photo_camera Michael Dello-Iacovo / Midjourney
computer screen with a blue nebula, hyperrealism --ar 16:9
Simulations and Catastrophic Risks
Bradford Saad  •  June 20, 2023

A PDF version of this report can be downloaded here.


This report explores interactions between large-scale social simulations and catastrophic risks.  It offers a tour of the surrounding theoretical terrain and brings together disparate literatures that bear on the topic.  My hope is that the report will facilitate further research in this area and directly or indirectly inform the decisions of benevolent actors who face choices about whether and, if so, how to develop and use powerful simulation technologies.

While the report is long, most of its sections are self-contained. So, readers should feel free to skip to section(s) of interest to them, perhaps after browsing the overview and preliminaries in §§1-2.

The report largely focuses on philosophical connections between simulations and catastrophic risks. That’s because, as a philosopher, philosophy is what I know best, not because I seriously considered focusing on alternative sorts of connections and then deemed philosophical connections most worthy of attention.  That said, I do think that these philosophical connections are worthy of attention: as the report illustrates, they are both important from a risk mitigation perspective and a fascinating subject matter from a perspective of pure inquiry.

I have aspired to write the report in an even handed manner. In particular, I have tried to bracket my own tendentious views in order to identify what I think should be widely recognized as key questions in this area and factors that are relevant to answering those questions. And I have incorporated sensitivity analyses of how different issues would play out, depending on one’s background views.  However, I have no doubt fallen short in this aspiration. Fully bracketing one’s own controversial views is no easy thing.  And I have not hesitated to rely on my own views about what should be controversial or to say that considerations point in a particular direction when I think this is clear.

I started writing this report in the summer of 2020. I continued working on it intermittently through the spring of 2023. With the public release of large language models and a surge in research interests in artificial intelligence and catastrophic risks in late 2022 and early 2023, I came to the realizations that relevant work was coming out faster than I could read and incorporate it and that any attempt to address the state of the art would quickly become dated.  While recent developments have not prompted any major changes to my analysis in the report, they have shortened my AI timelines and updated me toward thinking that there are substantial costs to delaying work on mitigating catastrophic risks associated with simulations. Thus, I have decided to release the report in its current form as a living document—one that I may occasionally update with important additions or corrections, though I have no intention of (futilely) trying to keep it up to date with the many relevant and rapidly evolving literatures.

For helpful discussion or comments, I am grateful to Jacy Reese Anthis, Austin Baker, Zach Barnett, Brian Cutter, Adam Gleave, Michael Dello-Iacovo, Drew Johnson, Fintan Mallory, Ali Ladak, Richard Ngo, Janet Pauketat, Jonathan Simon, and participants in an ULTIMA colloquium at Utrecht University. For feedback on related works that was especially helpful for this project, I am grateful to Daniel Berntson, Emery Cooper, Han Li, and Caspar Oesterheld.

1. Introduction

Future simulation technology is likely to both pose catastrophic risks and offer means of reducing them.  While there is much relevant work on the topic, it is scattered across disparate literatures. The main goal of this document is to bring together existing work in order to facilitate future research that will produce better understanding of the topic and help mitigate associated risks.  

To orient readers, I’ll start with an overview. Individual sections are largely self-contained. However, §2 offers preliminaries that some readers may find helpful for later sections. And, for readers who are unfamiliar with the simulation hypothesis or the simulation argument, I would recommend reading §8 before reading any of §§9-13.  To make it easier for readers to read at their desired level of depth, I’ll use bullet points, with details, examples, etc. in nested bullet points that can be skipped.

Here, then, is an overview of sections and their key contributions:

2. Preliminaries on Simulations and Catastrophic Risks

Simulations are systems that are designed and created to model other processes.  I will understand simulations broadly to include not only computer simulations but also systems that couple biological subjects with virtual environments.  I will primarily but not exclusively focus on large-scale social simulation scenarios, i.e. ones involving simulations that model at least tens of thousands of minds, some cognitive processing within each of those minds, and interactions between those minds.  As we will see, whether simulations themselves contain (conscious) minds will matter for some purposes but not others.  

My focus will be on catastrophic risks that interact with simulations and which are high stakes in that they either threaten millions of (present or future) people with significant harm or else pose a risk of a comparably bad outcome.[1]  These will include risks of astronomical quantities of suffering[2] and existential risks, i.e., risks of catastrophes that would permanently destroy humanity’s future potential.[3]  Given how bad such catastrophes would be, the risk of them could easily be worth mitigating even if their probability is low.  Thus, the discussion will not be restricted to high-probability catastrophic risks.  Nor will it be restricted to risks of acute catastrophes rather than ones that unfold over, say, many generations.  However, the category of high-stakes risks on which I’ll focus is somewhat broader than astronomical suffering and existential risks involving simulations. That’s partly because it includes risks of harms (not necessarily involving suffering) to digital minds that are comparable to existential risks and partly because the category includes risks of less severe catastrophes involving harm to millions of people (or something comparably bad).  I focus on this broad category for two reasons.  First, I think that the less severe catastrophes in this category are more likely to occur but still bad enough to be well-worth preventing.  Second, I anticipate that the most politically tractable way to mitigate the more severe risks in this category may be via interventions that target the less severe risks in the category. Hereafter unless otherwise indicated, I’ll use ‘catastrophic risks’ as shorthand for high-stakes, simulation-involving risks in the just described category.

The different catastrophic risks I discuss will be associated with different types of large-scale social simulations.  For every type of simulation I discuss, I believe there is a non-negligible probability that that type of simulation will be run.  However, these probabilities vary widely for different types of simulation.  While such probabilities are important for evaluating the quantitative impact of simulations on catastrophic risks, the discussion will mostly proceed at a coarse-grained level that is insensitive to these probabilities.  I’m hopeful that this report will prompt others to pursue more fine-grained and quantitative analyses.  Still, it is worth laying out what I see as some of the key differences in the plausibility of different types of simulations, as doing so may give a sense of how speculative different parts of the discussion are and offer something to go on for more fine-grained analyses.  And giving readers a glimpse of my underlying mental models may put them in a better position to understand and evaluate the discussion that follows.

To that end, it will be useful to distinguish several axes of variation among types of  large-scale social simulation:

As rules of thumb, I take the probability that a given type of large-scale social simulation will be run to be inversely related to how technologically demanding it is and to its computational complexity.  These are merely rules of thumb partly because the economic incentives to run simulations need not scale with technological demandingness or computational complexity.  Absent near-term catastrophes and stringent regulatory intervention that halt technological progress, I think it is highly probable (> 90%) that at least thousands of large-scale simulations will be run for research, entertainment, or economic purposes within the next century.[4] 

As noted, for some issues raised by simulations, it is crucial whether simulations would themselves contain conscious minds. For instance, catastrophic suffering risks will not arise within simulations that are clearly devoid of consciousness.  Other issues chiefly concern the effects of simulations on the external world. For instance, whether a simulation that models nuclear winter can be used to improve the prospects for recovery from nuclear winter does not turn on whether the simulation features conscious inhabitants.  In what follows, I will address the consciousness of simulation inhabitants  where relevant in connection with particular issues.

But it’s worth noting from the outset that whereas large-scale social simulations of some sort are clearly feasible, large-scale social simulations that contain conscious minds are not clearly feasible.  The former can be achieved by scaling up existing technologies. For extremely simple simulations of minds, it would be relatively easy to create a large-scale social simulation. For example, such simulations could be achieved with existing technology by scaling up real-time strategy games that simulate hundreds of interacting agents and incorporating rudimentary simulations of cognitive processes.  Much more advanced large-scale social simulations could also be achieved by embedding within a virtual environment digital agents trained through machine learning.  Training such agents is computationally expensive.  But once trained, it is relatively inexpensive to such agents across many tasks at once. So the currently high costs of training advanced machine learning agents may be less of an obstacle than one might have thought to creating large-scale social simulations that are populated by such agents.[5]  However, it is not clear that machine learning architectures are suitable for realizing consciousness. More generally, it is not clear that any existing computer technologies are of the right sort to generate consciousness.  But there is reason to think that more promising technologies are on the way: efforts are already underway to imbue large language models with sensory capacities, agency, and world-models to integrate them with robotic systems.[6]   And, in the future, whole brain emulations and neuromorphic systems may exhibit a high degree of functional similarity with brain processes that underlie consciousness, which would provide reason to think that such systems are conscious.[7]  It would also be unsurprising if superintelligent systems engineered hitherto unconceived types of architecture with the potential for consciousness.

Thus, while I take it that large-scale social simulations of rudimentary sorts are already feasible and their widespread future deployment is highly likely, I assign only a ~60% probability to the hypothesis that large-scale simulations with conscious minds will become feasible, given that technological progress is not halted in the next century.

Conditional on large-scale social simulations that contain conscious minds becoming feasible, I think it is unclear whether such simulations will be run on a large-scale (say, with at least thousands of such simulations).  Conditional on their becoming feasible, I assign ~40% probability to their being run on a large scale unwittingly (i.e. we run them without believing that they contain conscious minds), ~60% probability to their being intentionally run on a large scale , and ~25% probability to their not being run on a large scale (e.g. because humans universally enforce a ban on them (~5%) or because we lose control to artificial agents that opt not to run them (~15%)).[8]  Even in cases where large-scale social simulations come to house more conscious minds than there are humans, I would expect there to be more large-scale social simulations that do not contain conscious minds.

Many of the issues I discuss in what follows arise in a wide range of potential future scenarios with large-scale social simulations.  By my lights, no specific scenario of this sort stands out as especially likely.  So I will mostly discuss these issues in the abstract rather than in the context of particular scenarios.  Still, some readers may find it helpful to think through these issues in the context of some concrete scenarios.[9]  For this purpose, I’ll now offer some stylized scenarios.[10]  These scenarios are wild and speculative.  This comes with the territory, as it is highly probable that the future will be wild,[11] and  specifying concrete future scenarios is an inherently speculative endeavor.  The rest of the report won’t presuppose familiarity with these scenarios. So readers should feel free to skip them.

Virtualization of labor: In 2090, whole brain emulations arrive.  Because they emulate human brains, they can perform any cognitive task that humans can perform.  However, they can be run at much faster speeds than the human brain and at low costs.  While humans continue to command much of the capital in the economy, most human labor is largely priced out by whole brain emulations.  Because it is cheaper and safer to house whole brain emulations in controlled virtual environments than it is to equip them with robotics in the external world, they predominantly inhabit simulations. There is no global consensus on whether whole brain emulations are conscious or whether they have moral status. For ethical reasons and/or to preserve human jobs, virtual labor is initially banned in some jurisdictions.  However, these policies impose substantial economic costs on these jurisdictions.  The turn to virtual labor drives investment and productivity in places that do not heed such scruples.  Eventually, as opposition dwindles, whole brain emulations come to dominate the labor force nearly everywhere.[12]

Virtualization of leisure: In 2060, advances in machine learning and robotics have drastically reduced the demand for human labor. Advances in nuclear fusion have made energy abundant.  In countries that reap the benefits of these advances, the average citizen stands to present day billionaires in much the way that present day average incomes earners in the developed world stand to royalty of centuries past.  With their newfound wealth, these citizens invest heavily in leisure, including virtual reality. These investments create a virtuous cycle of improvements in virtual worlds that in turn drive more investment in them.  As these technologies are perfected, many people opt to live out most of their lives in virtual settings. The technologies are also put to other uses—for example, large-scale social simulations become commonplace in biology and economics.

Artificial replacement: Gradually over the course of the next century, the habitability of Earth’s surface degrades. Pollution and climate change render outdoor activity extremely hazardous in much of Africa and Asia.  But this is overshadowed by the evolution of biotechnology.  Open source biosynthesis software becomes widely available in the 2110s.  It is accompanied by cheap, automated, biosynthesis devices that are also widely available.  These are first used for personalized medicine.  However, they also put billions of people in a position to create and release novel pathogens.  For a few decades this threat is largely contained through a combination of regulations, surveillance, policing, and enormous investments in pharmacological responses to released pathogens.  The release of pathogens eventually outpaces governments’ abilities to respond with these measures.  Use of cumbersome personal protective equipment then becomes the chief means for safely navigating the physical environment.  Rather than muddle along in these conditions, humans instead opt for a recently developed uploading procedure. The procedure allows an individual to transfer their personality and memories into a cognitively enhanced digital mind with a virtual body of their choosing.  After the procedure, individuals live out their digital lives in virtual worlds, often with the digital successors of friends and family who also opted for the procedure.  To ensure that the infrastructure for these worlds is maintained, the worlds are porous: their inhabitants occasionally return to the outside world in robot form to carry out simulation maintenance.  Over the course of a few more centuries, the (biological) human population declines to zero.  Our digital successors come to regard our extinction in much the way that we regard our descent from now extinct ancestral species: an important historical fact to be sure, but not a tragedy that emotionally resonates.

Catastrophic recovery: The year is 2125. Digital minds have been developed in recent decades. Costs and regulations have kept their population and power at bay.  Meanwhile, through wargaming simulations and economic modeling, a regional power concludes that its strategic advantage is rapidly eroding, that resource scarcities will push neighboring powers to attack it in the next decade, and that its best option is to strike preemptively.  It does so.  The conflict escalates. Other countries are drawn into the conflict. The conflict leads to a nuclear war on a global scale.  The survivors are mainly humans in areas that are relatively habitable during the ensuing nuclear winter along with digital minds in simulation shelters, the latter having been put in place by militaries, philanthropic organizations, and wealthy individuals who attempted to upload themselves as digital minds to simulations. In the aftermath, digital minds rapidly initiate recovery efforts.  While different factions pursue different strategies, a common theme is that digital minds seek to inhabit virtual worlds.  Indeed, their activity in the physical environment is largely geared toward constructing and maintaining the requisite infrastructure.  Human recovery efforts proceed largely independently and at a much slower pace.  

Singularity: In February of 2042, leading AI companies across the world detect worrying signs of explosive intelligence growth. Governments respond by imposing regulations that require reinforcement learning agents to undergo extensive safety testing and training in virtual environments.  To meet these requirements, companies drastically scale up AI safety facilities, which house the simulations used for testing and training.  It is estimated that the number of reinforcement agents housed in AI safety facilities at any given time exceeds the total human population.  Disconcerting failures during testing lead a few companies to shut down their testing and development programs.  Others barrel ahead. Within a few months, one company announces that the first publicly known superintelligent agent is undergoing safety testing in company facilities.  A second company reports that it has temporarily lost control of its safety infrastructure to superintelligent agents in testing and that it is taking all necessary means to regain control.  A third company holds a press conference to report a lab accident in which superintelligent reinforcement learning agents—which had exhibited power-seeking tendencies in testing—somehow cooperated with each other to access the internet and, after bypassing security measures, managed to surreptitiously plant copies of themselves in several undisclosed data centers.  A government official at the press conference confirms that the company is working with authorities to identify and eliminate any residual threat posed by this incident, claims that there is currently no evidence of such a threat, and asks the public to remain calm.

3. Simulation as a Tool for Researching Risk Reduction

Risk levels for different catastrophic risks are highly uncertain and seem likely to remain so for the foreseeable future. The same goes for prospects of risk-reducing strategies.  Research that lessens these uncertainties could put us in a better position to set risk-reduction priorities and select risk-reducing interventions. Using simulations to conduct such research is one way that simulations could contribute to the reduction of catastrophic risks.

We can divide the use of simulations in risk-reduction research into two categories: direct and indirect.  I’ll have comparatively more to say about the latter. However, I should register that this does not reflect a distinction in significance between the two categories: I regard both as exploration-worthy but don’t have a settled view about which is more important.

3.1 Direct Risk-Reduction Research

In direct risk-reduction research, simulations of potentially catastrophic scenarios would be run in order to collect data on the catastrophic risks we face, the magnitude and severity of those risks, the interventions available to mitigate them, and/or intervention efficacy.  The idea would be to run simulations that are similar enough to our circumstances in relevant respects for risk and intervention data about the simulated scenarios to have direct bearing on risks and mitigation options in our own case.  Collecting data from and running tests with simulations would have potential advantages over trying to collect it from unsimulated sources: it might be that in contrast to unsimulated data sources, using simulations to generate data is more feasible, cheaper, faster, morally preferable, more conducive to data collection, or easier to control.

Research simulations could conceivably provide frequency data that bears on known catastrophic risks.  For example, in order to better estimate the likelihood of nuclear war, one might simulate many variations of the 21st century and observe the prevalence of nuclear war.[13] Or simulations might model interactions between multiple risks, e.g. to evaluate how frequently global climate catastrophes would lead to global wars or pandemics.

Similarly, research simulations could be used to generate data on unknown catastrophic risks.  For example, simulations might be used to “peek ahead” to identify technological “black balls—technologies that once created (by default) destroy the civilization that creates them—before it is too late: rather than creating a technology that by default destroys civilization, we might first simulate such a technology; upon recognizing it as such, we might then avoid creating it or create it only with due precautions.[14]

One use case of research simulations that merits its own treatment is virtual boxing, in which superintelligent (or otherwise potentially dangerous AI) systems are initially confined to simulated environments as a means for testing whether they are safe to release in our environment.[15]  For virtual boxing to be of use, simulations containing superintelligent systems would need to generate some information about those systems that is consumed by systems outside the simulation.  One concern about virtual boxing is that superintelligent systems might exploit these channels in order to escape or gain hazardous forms of influence outside the simulation.  Another concern is that confined systems might recognize their situation, behave safely in service of the instrumental goal of being released, and, after release, suddenly behave treacherously in pursuit of their final goals.[16]

One approach to addressing these concerns with virtual boxing would be to conduct safety testing within nested simulations: this might prevent catastrophes by confining consequences of escapes and post-release treacherous turns to simulated environments.[17]  Additional safety gains might be obtained from punishing such attempts or rewarding compliance with an unpredictable delay.  This would incentivize fully unboxed systems that are not certain that they are fully unboxed to continue behaving safely, even if their underlying goals would, unbeknownst to them, be best pursued by, say, inflicting revenge on the agents that confined them.[18]  Nested simulations could be used in similar fashion to discourage systems from engaging in activities that heighten the risk of catastrophes even if they do not inherently constitute them. For instance, rewarding agents for not seeking control of their source code or for not “wireheading”[19] their reward mechanism in simulated environments could be used to incentivize fully unboxed agents who do not know they are fully unboxed to abstain from such activities.[20]

Research simulations of catastrophes could also be used to test:

One domain where simulation-based research seems particularly promising is that of value dynamics—i.e. how normative (moral, prudential, epistemic, political) views that guide action evolve and produce changes over time—and their bearing on catastrophic risks.[23]  Plausibly, value dynamics play a central role in determining a societies’ goals, institutions, priorities, and hence catastrophic risk levels.  However, value dynamics are poorly understood. It is difficult to test them in the real world at the population level, and the frequency data we have on them is noisy and sparse at the scales relevant to catastrophic risks.  Simulations offer a way forward: by simulating populations of cognizers whose behavior manifests certain normative values and observing how such populations evolve and respond to risks, we may glean clues to answering questions like the following:

Given their roles in the general population or in communities focused on catastrophic risks, simulation-based comparisons of the following seem promising:

Evaluating differential prospects for different value dynamics is important because stable value dynamics would harbor the potential to reliably realize value on vast spatial and temporal scales. It could be much better to enact a stable value that will continue yielding mildly positive outcomes into the far future over an unstable—and, hence, probably short-lived—value that would yield an extremely positive outcome while present.  If value dynamics turned out to be inherently unstable on large time scales, that would severely limit the tractability of influencing the far future and so be a point in favor of prioritizing nearer-term catastrophic risks.

3.2 Indirect Risk-Reduction Research

Research simulations could also indirectly reduce catastrophic risk.  Candidate uses of this sort include:

3.3 Dual-Use Potential

As with other technologies, research simulations have dual-use potential: whether these technologies heighten or reduce risks will depend on how they are used.[58]  Some possible dual uses include:

As a recent, cautionary illustration, consider Collaborations Pharmaceuticals, Inc., a company that uses a model with generative and predictive machine learning components to identify new molecules and predict their biological properties.[59]  Ordinarily, the company trains their model with a reward function that penalizes toxicity.  However, in preparation for the conference, the company tweaked their models to reward toxicity and trained them on publicly available drug-like molecules, not toxic compounds.  Within six hours, their model identified 40,000 molecules that were predicted to exceed a toxicity threshold set by one of the most toxic chemical warfare agents.  These included that agent, other known chemical warfare agents, agents with higher predicted toxicity than publicly known agents, and a class of molecules in an unexplored region of molecular property space.  The company reported an absence of significant barriers to synthesizing these molecules.  The company also reported that its researchers had previously been naive to the potential misuse of their trade, despite working in the area for decades.

As this case suggests, the potential misuse of research simulations is a source of catastrophic risks that is apt to be underestimated.  Contributing factors include:

All this suggests that preventing the misuse of research simulations is a promising strategy for catastrophic risk reduction.  Since some of the relevant technologies have yet to be invented, there are at present limits to pursuing this strategy via technical safety work.  On the other hand, I would expect enacting safety regulations to become increasingly difficult as the technologies become more widely used and entangled with the interests of powerful actors.  If so, there is reason to pursue the strategy via AI governance in the near-term.

3.4 Will Research Simulations Be Superseded Before They Arrive?

If large-scale research, social simulations arrive after some other technology—for example, superintelligent AI—that would be better suited to conducting the relevant research, then research simulations would presumably not be run. In that case, there would be no point in considering them in thinking about how to reduce catastrophic risks.

The hypothesis that research simulations would be superseded before they arrive may turn out to be correct.  However, it should not be used as a basis for ignoring the prospects and dangers posed by research simulations, as it is not a hypothesis in which we should be very confident. There are several reasons for this:

4. Simulation as a Tool for Ethically Enhancing Testing

Some experiments on humans and non-human animals that could yield valuable information are ethically problematic.  Indeed, history is riddled with medical experiments in which people were harmed or killed by experimental treatments to which they did not consent.[62]  In some cases, simulations offer an ethically superior alternative: they provide a way to acquire the sought information without harming subjects or infringing on their autonomy.[63]  These alternatives should become more tractable as animal experimentation requirements on drug testing are relaxed.[64]

Similarly, simulations may offer a means to ethically enhance experiments that are currently deemed ethically permissible: for example, rather than performing suffering-involving experiments on non-human animals, experiments might be performed on simulated counterparts of them that are unconscious or which have only positively valenced experiences.  Likewise, for drug treatments on human patients with presently incurable diseases.  Such simulations might also offer advantages in terms of speed, number, and data-collection over corresponding human and animal experiments.

From a perspective of catastrophic risk prevention, it might seem that such experiments harbor only modest potential for improving the long-term future: it is not clear how they might be used to reduce the risk of catastrophes.

In response:

5. Using Virtual Reality to Promote Risk Responsiveness and Disaster Preparedness

One potential path to reducing catastrophic risk aims to make them more of a priority for policymakers by making relevant political constituencies more responsive to them. It is not surprising that catastrophic risks do not play a larger role in mainstream politics, as various factors make them hard to think about or tempting to ignore: they involve (small) probabilities, difficult to quantify uncertainty, large numbers of persons as well as many non-agent variables, and spatiotemporal scales that humans do not ordinarily think about. All this suggests that, for those who take reducing catastrophic risks to be a top global priority, simulations may offer a low-hanging fruit: simulations might be used to reduce catastrophic risks by drawing more people to think about, understand, and respond to those risks.

Gamified simulations of catastrophes are one sort of simulation that could be used to this end.  In such simulations, catastrophic risks are simulated, players manipulate parameters to try to reduce risks, and the players then witness the consequences of their choices.  Interactive programs of this sort have been made for climate change (e.g. see[66]  Popular games that simulate various wars, natural disasters, and/or civilizational collapse can be found in series such as Civilization.  As far as I know, no popular games have been introduced or widely used for the purpose of increasing responsiveness to catastrophic risk. But it is interesting to note that the unprecedented popularity of such games in the 1990s and 2000s was followed by an explosion of research on existential risks in the 2000s that continues to this day. Whether or not such games had any role in cultivating responsiveness to catastrophic risk among today’s researchers, there may be ways of using them to that end, e.g. by using them as part of an educational curriculum on catastrophic risks.[67]

Immersive simulations offer another sort of simulation that might be used to increase responsiveness to catastrophic risks.  As virtual simulations continue to improve, we will be able to have increasingly realistic-seeming experiences of virtual environments. It is plausible that realistic experiences of living through catastrophes would induce increased risk-responsiveness. Compare: we would expect those who have lived through a world war or pandemic to take militaristic and biological catastrophic risks more seriously than those who have merely read about them.  Of course, immersive catastrophic simulations could be potentially traumatizing.  This ethical concern is among those that would need to be taken into account in designing such simulations and deciding how to deploy them.

The risk-reduction potential of gamified and immersive simulations depends partly on two factors: their effectiveness at increasing risk responsiveness and the extent to which risk responsiveness serves to reduce catastrophic risk.  These factors are difficult to estimate. Evaluating the first factor is a potential research program that could be implemented now using standard tools and methods from the social sciences, perhaps in collaboration with video game creators.  Evaluating the second factor is less straightforward. In the future, estimates of it might be achieved through simulations of societies that face catastrophic risks and exhibit varying levels of risk responsiveness.

Immersive simulations can also be used as training tools to improve safety and disaster mitigation. Immersive simulations are already used to train astronauts, pilots, firefighters, military personnel, and workers in the nuclear industry.  I am not aware of any general investigation of the potential to reduce catastrophic risks through immersive simulation safety and disaster mitigation.  With recent improvements in virtual reality technology and more on the way, I believe this is a valuable time to carry out such an investigation.

6. Bunkers, Fallbacks, and Grand Futures

6.1 Simulation Refuges

Catastrophic risks can be mitigated either by lowering the probability of catastrophe or by reducing the expected harm of the catastrophe if it occurs.  For example, the risk posed by nuclear war can be mitigated either by lowering the probability of nuclear war or by raising the probability of civilizational recovery in the event of nuclear war.  A now common observation is that existential catastrophes would tend to be far worse than many non-existential catastrophes, even ones that would kill a large percentage of the world’s population: while both sorts of catastrophe would be bad for those directly harmed, existential catastrophes also preclude the realization of value in the vast stretches of time and space that lie before us.  Thus, it is worth considering proposals that aim to reduce existential risk without aiming to reduce other types of catastrophic risk.  

One such proposal is to build refuges, facilities that would house agents in the event of a would-be existential catastrophe and rebuild civilization in its aftermath. Proposed sorts include subterranean, aquatic, and extraterrestrial refuges.[68]  Discussions of such refuges tend to assume that such refuges would be inhabited by humans and that humans would reside in non-virtual environments within refuges.  Simulations offer alternatives. One option would be to create refuges that would physically house humans who would primarily live in virtual environments.  For example, a subterranean refuge might house a society of humans in cramped quarters that virtually lives in more expansive simulated environments until it is safe to rebuild civilization on Earth’s surface.  Another option would be to create refuges populated by digital minds rather than humans interacting with simulated environments.  While such minds might inhabit simulated environments while living in the refuges, they could be designed to interact with the environment external to the refuge as well. When the time is right, they would exit the refuge and rebuild civilization.

Beckstead (2015) notes some limitations of using refuges for risk mitigation. One is that they would not help in “overkill” scenarios where the catastrophe would kill people in refuges.  Another is that they would not help in very long-term environmental damage scenarios where there is no environment for refuge inhabitants to return to in order to rebuild.  Simulation refuges could avoid these limitations to some extent: for example, Earth’s surface might be rendered uninhabitable for humans by a nuclear war, extinction level pandemic, misaligned superintelligent AI, or a nanotechnological catastrophe; yet it might still be habitable by digital minds lying in wait in extraterrestrial simulation refuges.  Similarly, digital minds might be well-positioned to embark from subterranean or aquatic refuges to rebuild in the wake of a catastrophe caused by biological pathogens that prevent humans from surviving on the planet.

Further advantages of simulation refuges over non-simulation refuges may include:

A major concern about purely digital simulation refuges is that it is not clear that the digital agents populating them would be conscious or capable of realizing value.  Arguably, a scenario in which digital systems emerge from refuges and create a bustling galactic civilization of unconscious machines would be as bad as extinction.  This concern could be mitigated through future research on consciousness that reveals which sorts of digital systems would be conscious. (or exacerbated if it is revealed that digital systems are ill-suited to realize consciousness)  However, it is not clear that such revelations are possible deliverances of future research and, even if they are, they may not arrive or be incorporated in time.  Absent such revelations, we might opt for simulation refuges whose inhabitants would emerge to rebuild a civilization in which biological agents (presumably humans) play an important role.

Simulation refuges would not guard against all catastrophic risks.  Some catastrophes—such as physics experiments that destroy the known universe and the shut down of our universe in the event that it’s a simulation—would not spare those living in refuges even of the simulation variety.  It should also be acknowledged that another limitation Beckstead notes applies to both simulation and non-simulation refuges: they would not help in “underkill” scenarios involving catastrophes that are not destructive enough for refuges to be relevant.

6.2 Simulations and Grand Futures

We have seen that simulations could be used to safeguard against existential risks, thereby avoiding the permanent loss of our civilization’s immense future potential.  In addition, simulations may have a role to play in realizing that potential.  Indeed, much of our future’s potential may lie in the possibility of spreading virtual paradises across the galaxies within the affectable universe.  The design space of digital systems that could be created in the future is vast.  By way of comparison with biological minds, future digital systems will very probably be cheaper to produce, as well as much faster at processing information and capable of processing much larger quantities of information. This suggests that if digital systems will be capable of having experiences at all, then—relative to humans—they will be capable of having far more of them and of realizing far more value through them.[69]  A further reason for thinking that much of the future’s potential value may lie in the potential for virtual paradises can be found in the option of using “aestivation” to make the most of energy resources: by entering a relatively inactive state until computation becomes more efficient with the arrival of cooler conditions in the very far future, civilization could extract more compute and in turn value from its resources.[70] 

Shiller (2017) uses closely related considerations to argue in favor of the artificial replacement thesis that we should engineer the extinction of humanity so as to bring about artificial descendants capable of living better lives.  Such beings could conceivably inhabit non-virtual environments.  But it seems more likely that they would live out their artificial lives in largely virtual settings, as virtual environments would likely be much easier to mold to the preferences of such artificial beings than would non-virtual environments.

One concern about a grand future populated by such beings is that such virtual realities would merely simulate valuable phenomena.  However, on reflection, at least given that such beings would be conscious, this concern seems misplaced: genuine friendships, love, achievements, knowledge, etc. could exist in virtual realities.[71]

The more pressing concern in the vicinity is again that the digital systems would be unconscious and incapable of realizing value.  Perhaps by the time we’re in a position to initiate a grand future, we’ll know whether digital simulations would be conscious.  If not, then one way to address this concern in the context of a grand futures strategy would be to adopt a mixed digital-biological portfolio: we could aim to create both biological and digital paradises.[72] 

Hedging our bets in this fashion would ensure that some sort of immensely valuable future would exist.  Such a mixed-strategy might be best in expectation, but it would be unlikely to bring about the best outcome: the simulated paradises would either turn out to feature consciousness and realize value or they would not; if they did, the resources devoted to running biological paradises probably could have been used to bring about a much better outcome via more digital paradises; if not, the resources devoted to running digital paradises probably could have been used to bring about a much better outcome via more biological paradises.  This serves to highlight how knowing the conscious status of digital systems could prove valuable: without such knowledge, doing what is best in expectation may require us to leave much of the future’s potential value unrealized.

In the context of grand future strategies, simulations could also be used as a sort of safety check.  Before deciding once and for all which grand future to implement, we would want to take measures to ensure that we have not overlooked important upsides or downsides of options that we are choosing among.  One way to do this would be to simulate our options prior to choosing.  In addition to revealing features of options that we would have otherwise overlooked, such simulations could also be used to tweak and optimize different approaches to bringing about a grand future.

6.3 Fallbacks

Simulations could also serve as fallbacks, i.e. as outcomes that we bring about in the event that better outcomes elude us.  For instance, suppose that spreading civilization beyond our solar system is initially the best option available to us. However, currently unknown engineering obstacles to interstellar travel force us to relinquish any hope of extending civilization to other solar systems. In that case, we would have squandered almost all of our cosmic potential. Yet our remaining potential might be vast: using only resources from our solar system, it might still be within our reach to run simulations populated by trillions of minds that enjoy super-human levels of welfare.

How valuable simulations would be to have as fallback options depends on various factors, including:

A full analysis of the prospects for using simulations as fallback options is a task for another occasion.

7.  Catastrophic Simulations

There is reason to think that at least some simulated minds (such as whole brain emulations) would be conscious and capable of suffering.[73]  This raises the possibility of catastrophic simulations, simulations that realize catastrophic quantities of disvalue.[74]  Disconcertingly, there are a range of at least somewhat plausible scenarios in which such simulations might be run.  Candidates for such simulations include:

8. Background on the Simulation Hypothesis and the Simulation Argument

Previous sections examined connections between catastrophic risks and simulations that might be run in our universe.  The next few sections will explore connections between catastrophic risks and the simulation hypothesis that our universe is itself a simulation.  While this may seem to be an outlandish or skeptical hypothesis, there is an interesting argument for it that is taken seriously by relevant experts.[83]  In this section, I will rehearse the simulation argument for the simulation hypothesis.  This will set the stage for discussing connections between the simulation hypothesis and catastrophic risk in later sections.

There are two driving ideas behind the simulation argument.  One is the broadly empirical claim that the expected motivations and computing potential of technologically advanced civilizations support simulation dominance, the hypothesis that at least a small portion of beings like us will produce a very large number of beings like us that are simulated—a large enough number for most beings like us to turn out to be simulations.[84]   The second is that if most beings like us are simulated, then we are probably simulated.  This idea rests on the (bland) indifference principle that we should divide our credence evenly among hypotheses about our self-location in the class of observers like us.[85]  Simulation dominance and that application of the indifference principle jointly entail that we are probably in a simulation.

The simulation argument has been spelled out in different ways in the literature.  One choice point concerns how to precisify the relevant class of observers: while the indifference principle is intuitive, it is not clear exactly what it takes for observers to be like us.  Here, I will finesse this issue by using ‘beings like us’ to pick out whichever observers fall within the relevant class on the most plausible version of the principle that is applicable to you and me.[86]  For concreteness, you might think of this as the class of conscious beings with roughly human-level intelligence.  In some presentations of the argument, the conclusion is simply that we are probably living in a simulation.  In others, the conclusion is couched as a disjunction between our (probably) being in a simulation and possible ways of blocking that result.[87]  The literature also contains various proposed amendments for patching the argument in response to objections, along with less consequential variations in formulation.  To keep the discussion tractable, I will just work with the above formulation.

The argument can be questioned in various ways.[88] Some responses include:

This is not the place to attempt a comprehensive evaluation of the simulation argument and responses to it.  So I will restrict myself to the following remarks, which aim to (1) address what I expect to be some of the more prevalent concerns about the argument and (2) highlight connections between some of the responses and catastrophic risks.

First, consider the response that the argument fails because civilizations generally go extinct before being able to create intelligent simulated beings—for example, as a result of a technology that inevitably precedes simulation technology and leads to civilizational destruction shortly after its discovery.[95]  If this response is correct, then it gives us reason to think that our civilization will end before being able to create intelligent simulated beings.  Thus, if the response is correct and there is reason to think we are on track to be able to create intelligent simulated beings, then there is also reason to think our civilization will succumb to catastrophe in the relatively near term—a catastrophe that at least knocks it off track, whether or not it terminates our civilization.  This can be understood as an argument for taking near-term catastrophic risks more seriously.

Second, if the simulation argument fails because advanced civilizations generally decide not to create intelligent simulated beings, this may be because civilizations that avoid catastrophe long enough to be able to produce such simulations are generally very risk averse.[96]  There is also reason to think such civilizations would have centralized control and/or coordination schemes with near-universal compliance: absent such control or compliance, we would expect sub-civilizational actors to create many intelligent simulations, given that intelligent simulations would eventually become cheap to produce in advanced civilizations.[97]  Similarly, those who are attracted to this response may take it as an indication of the paths that are available to humanity which do not end in near-term extinction.

Third, in informal interactions I've encountered a number of fellow philosophers who are tempted by the instability response. However, for several reasons, this response is unconvincing:

Another response that many people find tempting is to dismiss the simulation argument on the ground that it is a skeptical argument.  For several reasons, this response is also unconvincing.

9. Shutdown Risk

One corollary of the simulation hypothesis is that our simulation may be shut down.[104]  Depending on the axiological trajectory of our universe at the time of shutdown, shutdown could be catastrophic. If a shutdown happened today, it would prematurely end the lives of billions of people and it might destroy immense quantities of expected value that lie in our potential to usher in a grand future.  Shutdowns will continue to pose a catastrophic risk, at least as long as we manage to steer clear of other catastrophes.

It may be tempting to think: we shouldn’t worry about catastrophic shutdowns because there’s no way for us to influence whether they occur.  However, this thought is mistaken on two counts.  First, even if we cannot influence the risk of catastrophic shutdowns, that risk has implications for the expected value of the long-term future:  if we assign a tiny but constant probability to shutdown in any given year (conditional on survival up to that point), that will drive down the expected (dis)value of outcomes that would occur further in the future.[105]  Such discounting would drive down the expected (dis)value of nearer term future to a lesser extent.  Thus, updates that elevate the risk of catastrophic shutdowns will tend to weaken the case for prioritizing the far future.

Second, there are shutdown triggers that we may be able to influence:

Developing a more systematic understanding of potential shutdown triggers and the prospects for avoiding them is an underexplored topic.[109] 

10. Potential Upsides of Shutdown

The previous subsection noted some obvious reasons for thinking that the shutdown of our simulation might be catastrophic.  What may be less obvious is that there are also ways in which shutting down our simulation could mitigate catastrophic risks.  These include:

I hasten to add: even in circumstances where evidence indicates that the best option will involve trying to trigger shutdown, it could be a grave error to attempt to trigger shutdown too soon.  That’s because there’s value in delaying—and preserving other options in the meantime—in order to put us in a better epistemic position to determine whether shutdown is the best option.[115]  And that value could easily outweigh any time cost associated with delay.  Compare: suppose you’re on a slowly leaking ship that will probably sink.  Even if your evidence indicates that jumping overboard with a life vest is your best option, that doesn’t mean you should take it immediately. It might instead be more reasonable to wait to see how the situation evolves in case a better option emerges (e.g. staying on the ship in the event that the leak is fixed or exiting in a lifeboat).[116]

11. The Simulation Argument and Religious Catastrophic Risks

Traditional religions countenance distinctive catastrophic risks ranging from apocalypses inflicted by a creator to afterlives involving eternal torment.  (Distinctively) religious catastrophic risks are rarely discussed in the scholarly literature on catastrophic risks. The same goes for public non-academic discussions of catastrophic risks that I am familiar with.[117] Perhaps this is because such risks are typically approached with methods that do not treat these religions—or any religion for that matter—as a source of data, much less as authoritative.  In any event, given the enormous stakes associated with religious catastrophic risks, they should be considered in analyzing how to reduce catastrophic risk unless we regard religious catastrophes as astronomically unlikely.[118] For those who regard religious catastrophic risks as non-negligible, there is then a question as to how they interact with the simulation argument.[119]  There are at least four connections of interest.

Flawed simulators and religious catastrophes.  Outside the context of the simulation argument, it is often assumed that our universe either was not created or else it was created by an agent that is all powerful, all knowing, and perfectly good.  The simulation argument casts doubt on this assumption: it gives reasons for thinking that our universe is created by intelligent beings, but provides no reason to think such beings would exhibit a maximal degree of power, knowledge, or goodness.  Further, taken with the many morally problematic features of our universe, the simulation argument lends support to the hypothesis that our universe was created by a morally flawed or morally indifferent agent.  This in turn lends support to the hypothesis that our universe is subject to religious catastrophic risks such as catastrophic interventions by our creator(s) or disvaluable afterlives.[120]

Simulation as a solution to the problem of natural evil.  The simulation argument arguably indirectly supports the hypothesis that God exists—this is so even on the assumption that our universe was, if created, created by a non-divine simulator. For the simulation hypothesis offers a candidate solution to the problem of natural evil, i.e. the problem of reconciling the apparent existence of natural evils (ones not caused by the free choices of agents) with the existence of God (understood as an agent that is all powerful, all knowing, and perfectly good).  The candidate solution on offer is that the appearance of natural evils is an illusion: God created a world devoid of natural evils in which agents make free choices.  One of those agents freely chose to create a simulation, a simulation that turns out to be our universe.  Thus, what appear to be natural evils in our universe are really evils caused by our simulator, not by God.

Extending the solution to help with other problems for theism.  The simulation solution to the problem of natural evil can be developed to solve a range of other problems for theism.  For suppose we add to the solution that our universe is just one simulation in a much larger reality of which it is, in a certain respect, not representative.[121]

Simulation solutions to problems for theism have implications for catastrophic afterlives.  On the one hand, there is some plausibility to the thought that we are less likely to have the sort of free will required for responsibility if we are simulated than if we are not.  This in turn suggests that, on the simulation solution, we are less likely to be subject to catastrophic afterlives imposed by God as just punishment.  On the other hand, on simulation solutions, we need to distinguish between afterlives imposed by God and those imposed by the non-divine simulator(s) of our universe.  Thus, there remains the possibility that we are in a simulation and will be subjected to a catastrophic afterlife by the non-divine simulator(s).  The simulation solution somewhat constrains how bad such an afterlife could be: on pain of reinvigorating the problem of evil, the catastrophic afterlives cannot be so bad (or too inadequately compensated)[125] as to render the existence of God implausible.

Simulations, fine-tuning, and God.  Cosmological fine-tuning arguments for the existence of God are among the most popular contemporary arguments for the existence of God.  They appeal to the (supposed) fact that fundamental physical parameters take values within narrow ranges required for life.[126]  This is claimed to be expected if there is a designer of our universe but not otherwise—the idea is that it wouldn’t be surprising if a designer selected those parameters because they are required for life and the designer wanted to create a universe with life (or something for which life is a prerequisite such as biological intelligence or conscious organisms).  It’s usually assumed that if our universe has a designer, the designer would be God.  Insofar as cosmological fine-tuning arguments support the hypothesis that God exists, they presumably also modulate the catastrophic risks associated with the existence of God.  The simulation argument bears on cosmological fine-tuning in several respects.[127]

We’ve seen that the simulation argument in different ways supports theism and undermines such support. Likewise, we’ve seen that the simulation argument in different ways boosts and diminishes religious catastrophic risks.  The net effects of the simulation argument on these issues remain open questions.

12. The Simulation Argument, Self-Location, and Catastrophic Risk

12.1 Background on Self-Location

The simulation argument relies on self-locating information: it assumes you should divide your credence evenly among observers like yourself. There are other arguments that also rely on self-locating information and bear on catastrophic risks.  The simulation argument interacts with these arguments in ways that bear on catastrophic risks.  This section will describe some of these arguments and interactions.  I should flag that there is much controversy within the literature on how to rationally respond to self-locating information and that much of the literature is of a technical nature that I cannot adequately summarize here.[131]  I expect that much of the discussion that follows will merit revisiting in a more technical setting and much of what I say in this subsection would need to be qualified or scrapped in light of reassessment.  I therefore offer the discussion that follows in a provisional spirit in hopes that it will stimulate further work on the topic, even if what I say turns out to be misguided in important ways.

Before diving into the arguments, I’ll use a simple example to put some of the key issues about self-location on the table: suppose an urn contains an unknown number of balls and that you and perhaps some other subjects are each going to draw one ball from.  Next, suppose you find out that most subjects who take a ball from the urn draw a red ball.  Intuitively, this gives you reason to think that you will probably draw a red ball.  This is an instance of ‘inward’ reasoning that moves from information about a distribution of subjects to a conclusion about yourself—this is the sort of reasoning encoded in the indifference principle invoked by the simulation argument.

In contrast, ‘outward reasoning’ moves from information about yourself to a conclusion that concerns the distribution of subjects like yourself.[132]  To illustrate, suppose that you are initially ignorant about how many subjects will draw from the urn and what it contains.  You then draw a red ball.  Here, two inferences are prima facie plausible. First, you might take the fact that you drew a red ball to boost the expected number of subjects who will draw red balls.  After all, you presumably initially reserved credence for the hypothesis that no subject would draw a red ball.  You’ve now eliminated that hypothesis and presumably redistributed whatever you had in it to hypotheses on which more than one subject draws a red ball.  Call this number boosting.  Second, you might take the fact that you drew a red ball to boost the expected proportion of subjects who will draw red balls.  After all, the higher proportion of subjects who draw red balls, the more likely it is that you will draw a red ball, in which case those hypotheses get a boost.  Call this proportion boosting.

A final variation: you are initially ignorant about what the urn contains, except that you know that it contains turquoise balls and that they—unlike balls of any other color the urn contains—are all too deep in the urn for any subject to reach.  In this case, draws cannot be treated as representative of the urn’s contents: they exhibit a sampling bias.  For example, drawing a red ball rather than a turquoise ball is not evidence against the urn containing turquoise balls.  In this case, the sampling bias is an observation selection effect: observations do not qualify as random samples because observations are biased toward certain outcomes.

In practice, factoring in number boosting, proportion boosting, and observation selection effects raises a host of difficult issues, e.g. concerning the individuation of reference classes.[133]  A case in point is the bearing of number boosting and proportion boosting on the simulation argument.  Insofar as number boosting favors hypotheses with more observers regardless of their observer-type,[134] it will favor hypotheses on which the world contains large numbers of observers in simulations over non-simulation hypotheses on which the world is relatively sparsely populated.  And insofar as number boosting favors hypotheses with more observers like us, it will favor large-scale simulation hypotheses on which many beings like us are simulated over sparsely populated non-simulation hypotheses.  Thus, number boosting arguably bolsters the simulation argument’s simulation dominance premise that at least a small portion of beings like us will create a large number of beings like us.  In contrast, insofar as proportion boosting favors hypotheses on which a higher proportion of beings are observers of some type or other, it will favor panpsychist views over views on which the universe is relatively sparsely populated with macroscopic subjects (whether simulated or not). On the other hand, insofar as proportion boosting favors hypotheses on which a higher proportion of observers are observers like us, it will favor simulation hypotheses on which enough beings like us are simulated for there to be proportionally more beings like us than there are on rival simulation and non-simulation hypotheses.  The latter result also fits with the simulation argument’s simulation dominance premise, though that premise does not require that there be a higher proportion of simulated beings like us than there are other sorts of observers.  

It is not obvious that the individuation of reference class implicit in the foregoing applications of number boosting and proportion boosting are correct.  That said, it should be borne in mind that number and proportion boosting may operate across multiple reference classes even within a single case.  Indeed, the foregoing applications are all mutually compatible.[135]  In what follows, rather than trying to settle the proper individuation of reference classes and what applications of number and proportion boosting are legitimate, I will instead focus on drawing out implications of different choices on these scores.

12.2 Evolutionary Arguments for Easy Artificial Intelligence

Some authors have appealed to the observation that evolution by natural selection produced human intelligence to promote optimism about our ability to engineer artificial systems with human-level intelligence.[136]  Pre-empirically, it’s not obvious that a world with our physics and chemistry could give rise to human-level intelligence at all, much less that it could feature beings with such intelligence that would be able to engineer systems with such intelligence.  Thus, finding out that evolution produced human-level intelligence eliminates one obstacle—that of incompatibility with the laws of nature—to our engineering such systems and so provides at least a smidgen of support for our capacity to achieve this engineering feat.

However, such optimism does not go beyond whatever optimism is licensed by the observation that physical and chemical processes in our world somehow gave rise to humans. The more interesting sort of evolutionary arguments contend that the fact that evolutionary processes not aiming for intelligence gave rise to human intelligence during the relatively short time (by cosmological standards) that Earth has existed suggests that engineering such intelligence is not that difficult of a problem.  The idea can be spelled out in terms of design space: despite being a very inefficient search procedure relative to those that are available to human engineers, evolution by natural selection managed to hit upon human-level intelligence in design space; so, given our superior search capabilities, we should be able to hit upon it without great difficulty.[137]

If this argument succeeds in reducing the expected difficulty of engineering human-level intelligence, it thereby bolsters the simulation argument by rendering it more plausible that civilizations like ours will be able to create simulated beings like us before going extinct.  

Rendering it more plausible that we will be able to engineer human-level intelligence also renders it more plausible that we will be able to—and, indeed, will—engineer superintelligent AI systems.  That outcome would itself bear on catastrophic risks, as superintelligent AI both poses catastrophic risks and harbors the potential to mitigate them.  Thus, the evolutionary argument for easy human-level AI bears on catastrophic risks not only by bolstering the simulation argument, but also by raising the probability of superintelligent AI.[138]

Just as the import of the simulation argument is sensitive to how number and proportion boosting are applied, so too is the import of the evolutionary argument for easy AI.[139]  To the extent that number boosting favors hypotheses with more observers, a crucial issue is how the difficulty in engineering intelligence co-varies with the number of observers.  The relationship is not obvious: while less intelligent observers may require fewer resources per observer (compare, say, fish versus humans), increases in intelligence may yield access (e.g. via space colonization) to large but otherwise inaccessible resource reservoirs.  For similar reasons, it is not obvious whether proportion boosting of hypotheses on which a higher proportion of entities are observers favors easy AI hypotheses.  On the other hand, to the extent that number boosting favors hypotheses with more observers like us, it favors:

Similarly, if proportion boosting favors hypotheses on which a higher proportion of observers are like us, it presumably favors hypotheses to the extent that it is easier for evolution to create beings like us and easier for us to create AI systems with human-level intelligence.  Thus, number and proportion boosting arguably boost easy AI hypotheses and in turn lend support to the simulation argument.

The evolutionary argument for easy AI is also sensitive to observation selection effects and choice of reference class. To see this, suppose first that the operative reference class is that of observers with human-level intelligence.  That is, suppose that our observations of a given sample space are to be treated not as random samples from that space but as random samples from the observations human-like intelligent observers make of that space.  Under this supposition, even if creating human-like intelligence is in fact exceedingly difficult, it might seem quite easy before taking into account the observation selection effect.[140]  For example, suppose that the expected duration for Earth-like conditions to yield human-like intelligence is many orders of magnitude greater than than have thus far elapsed on Earth. And suppose that the world contains many life-hospitable planets but that they all exist for durations on the same order of magnitude as Earth’s history.  Then only a tiny fraction of planets would harbor observers with human-level intelligence, but such observers would generally find themselves to be products of evolution that were produced in relatively short order.  To resist the temptation to erroneously conclude that engineering human-level intelligence is relatively easy, they would need to take into account this selection effect.

Alternatively, suppose that the operative reference class is a wider one—say, the one that includes observers with human-level intelligence and observers with lower levels of intelligence.  In that case, while the observation selection effect would make the evolution of intelligence appear easy even if it is hard, it would leave room for observing the absence of human-level intelligence.  So even once we take this effect into account, the fact that we observe human-level intelligence would rule out some hypotheses on which it goes unobserved because of how difficult it is to produce.  Thus, under the noted supposition, the observation selection effect leaves the evolutionary argument intact (albeit attenuated) and permits us to reason from evolution’s rapid production of human-level intelligence on Earth to conclusions about how difficult human-level intelligence is to engineer.

In this fashion, the evolutionary argument for easy AI—and in turn its ability to bolster the simulation argument—is sensitive to choice of reference class. I hasten to add that there are additional factors (notably timing of evolutionary transitions, convergent evolution of prerequisites for intelligence, and Fermi’s paradox) that can radically alter the import of observation selection effects on a given choice of reference class.[141]

12.3 Simulation and the Doomsday Argument

The doomsday argument holds that taking into account our birth rank should lead us to expect that we will go extinct sooner than we would have otherwise thought.[142]  Ćirković (2008: 129-30) helpfully summarizes the argument with the following analogy:

[Suppose there are] two… urns in front of you, one of which you know  contains ten balls, the other a million, but you do not know which is which.  The balls in each urn are numbered 1, 2, 3, 4, ... Now take one ball at random from the left urn; it shows the number 7. This clearly is a strong indication that the left urn contains only ten balls…. Now consider the case where instead of two urns you have two possible models of humanity's future, and instead of balls you have human individuals, ranked according to birth order. One model suggests that the human race will soon become extinct (or at least that the number of individuals will be greatly reduced) , and as a consequence the total number of humans that  ever will have existed is about 100 billion…. The other model indicates that humans will colonize other planets, spread through the Galaxy, and continue to exist for many future millennia; we consequently can take the number of humans in this model to be of the order of, say, 1018.  As a matter of fact, you happen to find that your rank is about 60 billion. (emphasis mine)

Proponents of the doomsday argument contend that we should reason in the same way about our birth rank as we do about drawing ball #7: our birth rank is more likely on hypotheses on which fewer observers like us will have existed than ones on which observers like us exist in numbers of the sort we would expect if we avoid near-term extinction; since the most plausible way for a smaller number of observers to exist is for extinction to happen soon; therefore, our birth rank provides evidence that we will succumb to extinction soon.

The doomsday argument is subject to a few qualifications.

I’ll now note some interactions between the doomsday argument and the simulation argument.[144]

Like the simulation argument, the doomsday argument relies on an application of the indifference principle: whereas the simulation argument tells you to divide your credence equally among simulated and unsimulated ‘locations’ in your reference class, the doomsday argument in effect assumes that you should divide your credence equally among birth locations within your reference class.  Thus, (dis)confirmation of the indifference principle would tend to (dis)confirm both arguments.  However, a challenge to one argument’s application of the principle need not carry over to the other argument.  One reason for this is that it is not obvious that the choices of reference classes presupposed in the arguments’ respective applications stand or fall together.

I noted above that number boosting arguably bolsters the simulation argument (via its simulation dominance premise).  In contrast, number boosting arguably cancels at least some of the import of the doomsday argument by boosting our priors in abundant-observer hypotheses.[145]

The doomsday argument can be taken to cast doubt on the simulation argument by raising the probability that we will go extinct before creating many intelligent simulated beings in our reference class.  There are several ways this might go.  

First, observers like us could generally go extinct before creating any intelligent simulated beings.  In this case, given that observers like us would create intelligent simulated beings absent near-term extinction, we will go extinct soon.  The plausibility of this outcome is inversely related to the number of civilizations at our level of technological development which contain observers like us: the more such civilizations there are, the less plausible it is that they generally go extinct before creating intelligent simulations.

Second, observers like us could sometimes create intelligent simulated beings, but never in quantities that underwrite the simulation argument.  The plausibility of this outcome is also inversely related to the number of civilizations at our level of technological development which contain observers like us. However, this relation is weakened by the fact that the creation of simulated intelligent beings may be strongly technologically coupled with AI risks that tend to result in extinction around the same time as civilizations gain the ability to create intelligent simulated beings.

Third, observers like us could create many intelligent simulated beings albeit ones not in our reference class. In contrast to the previous outcomes, the plausibility of this outcome is not inversely related to the number of civilizations at our level of technological development which contain observers like us.  Created simulated beings might fall outside our reference class either because they are unconscious or because they differ from us in some other respect.  There is a potentially enormous difference in value between these two possibilities. If we create many simulated beings but not many conscious simulated beings, that would likely be because we mistakenly believe that the simulated beings are unconscious.  This scenario would likely involve a catastrophic failure of resource allocation.  In the limit, we might set in motion a starfaring civilization, ensure that it will be inhabited by vast populations of intelligent simulations, and inadvertently engineer the cessation of value through a design flaw that renders the simulations unconscious. A potential alternative would be to engineer our own extinction and replacement through conscious simulations that fall outside our reference class.  For those who place substantial weight on the doomsday argument, there is reason to pursue this outcome, as it seems to be one of the only ones that is reasonably probable and positive by the lights of the doomsday argument.  That said, whether the argument deems this outcome reasonably probable turns on the unresolved issue of delineating our reference class.

The simulation argument can also be taken to cast doubt on the doomsday argument by revealing a way in which we might have a typical birth rank in a civilization that is not on the brink of extinction: if the relevant reference class concerns our simulators as well as intelligent simulated beings in other simulations, we might have a typical birth rank in our reference class even if we have a very early birth rank among intelligent beings in our universe-simulation. On the other hand, it is not clear why it would be illicit to run two versions of the doomsday argument: one for a broader reference class that encompasses all beings with minds like ours and another for a narrower reference class that includes only the beings in our universe. The latter version of the doomsday argument would at least not be blocked by the simulation argument, since it would reveal a way in which we might have a typical birth rank within our universe-simulation.

12.4 The Fermi Paradox

Recall that the Fermi paradox is the problem of explaining why we seem to be alone in the universe, given the apparently astronomical number of opportunities for life and advanced civilizations to emerge.  In §3.2 we saw that the Fermi paradox points to catastrophic risks. It does this by raising the possibility that we seem to be alone because civilizations generally go extinct before becoming observable by civilizations elsewhere in the universe.  We also saw that simulation-based research on Fermi’s paradox is one way simulations could be brought to bear on catastrophic risks.  In this subsection, I explore other connections between Fermi’s paradox, simulations, and catastrophic risks. Specifically, I will describe how some interactions between Fermi’s paradox and the simulation argument bear on catastrophic risks.

12.5 Boltzmannian Cosmologies

Physics offers a number of otherwise appealing cosmological models that have a peculiar consequence: they predict that the vast majority of systems in brain states like ours are “Boltzmann brains”, short-lived brains that result from thermal or quantum fluctuations.[154]  When taken with some innocuous-seeming assumptions, these theories yield the skeptical conclusion that we ourselves are probably Boltzmann Brains.  In particular, if we assume that what experience a system has is fixed by its brain state, then these theories predict that almost all observers with our experiences are Boltzmann brains.  An application of the indifference principle then dictates that we should divide our credences evenly among observers with our experiences and conclude that, on these theories, we are almost certainly Boltzmann brains. Avoiding this skeptical conclusion is the Boltzmann brain problem.

There is much controversy about what to make of the implications of these theories.  Here, I will describe some interactions between the Boltzmann brain problem and the simulation argument.[155]  As far as I can tell, the main bearing that the Boltzmann brain problem has on catastrophic risks is by way of these interactions.[156]

13 Simulation as Non-Causal Intervention

This section will introduce simulations as non-causal interventions, which I regard as an important and neglected factor with the potential to significantly influence risk levels for a wide range of catastrophes.  To a first approximation, the idea is that by raising the probability that certain types of simulations will be run, we can change the probability of our currently inhabiting those types of simulations and (perhaps counterintuitively) thereby meaningfully but non-causally affect the probability of various types of catastrophic risks.  The concept will require some unpacking, as it combines insights from the simulation argument and some recent work in decision theory.[160]

To start, recall the simulation argument’s premise of simulation dominance: at least a small portion of beings like us will produce a very large number of beings like us that are simulated—a large enough number for most beings like us to turn out to be simulations.  The main ways for this premise to be false would be if beings like us went extinct before we created beings like us that are simulations or if we decided never to create such beings.  Even if plausible, this premise is at present a speculative empirical claim. So we should not be extremely confident in it.  However, this could change: if we began mass producing simulations that realized beings like us, that would give us a powerful reason to accept simulation dominance. Those who have a high credence in the indifference principle (the other premise in the simulation argument) would then be under strong pressure to conclude from the argument that we are probably living in a simulation.[161]  Likewise, acquiring evidence that we will mass produce such simulations should boost our confidence in simulation dominance and in turn the simulation hypothesis via the simulation argument.

Next, consider that there is a family of variations of the simulation argument that concern different types of simulations. For example, there are salvation simulations in which minds like ours are immune to catastrophic risks but have much misleading evidence to the contrary.  There are also doom simulations in which it is inevitable that minds like ours will collectively succumb to certain catastrophic risks, regardless of their evidence concerning risk levels and regardless of their risk mitigation efforts.  And there are myriad other sorts of simulations in which minds like ours face risks that are bizarrely related to their evidence and actions.  Just as evidence that we will one day mass produce simulations with minds like ours should boost our confidence that we are in a simulation, so too should evidence that we will one day mass produce simulations of a given type that contain minds like ours boost our confidence that we are in a simulation of that type.  Thus, evidence that we will run such salvation simulations would be good news while evidence that we will run such doom simulations would be bad news.  And we are in a position to choose what type of news we receive—for example, setting up a fund to sponsor salvation simulations when they become technologically feasible would be a way of bringing about good news about the catastrophic risks we face.

Of course, we are either in a given type of simulation or we are not.  And we control neither whether we are in a simulation nor which type of simulation we are in if we are in one. Thus, there is no way to causally exploit the described connection between our advancing the creation of certain sorts of simulations and our inhabiting such simulations.  So it may seem that this connection is not decision-relevant in the context of catastrophic risks.  Endorsing this appearance would be to take a stand on an ongoing debate in decision theory. The debate concerns whether causal decision theory is correct, rather than one of its rivals such as evidential decision theory.  Roughly, whereas “[evidential decision theory] tells you to perform the action that would be the best indication of a good outcome, … [causal decision theory] tells you to perform the action that will tend to causally bring about good outcomes” (Levinstein & Soares, 2020).

To briefly illustrate, suppose you are deciding whether to vote. Holding fixed how everyone else votes, you know (let’s suppose) that your vote won’t make a difference as to who wins and hence that you could cause more good by doing something else instead of voting. You also know that the people like you will vote for the better candidate just in case you vote for that candidate, and that that candidate will probably win just in case the people like you vote for her.  In this case evidential decision theory recommends voting, while causal decision theory recommends not voting.[162]  Similarly, evidential decision theory will tend to recommend running salvation simulations, preventing doom simulations, and bringing about evidence that the former will be run and the latter prevented.  On evidential decision theory, such actions affect catastrophic risk levels in a decision relevant sense without causally affecting them—in this sense, evidential decision theory licenses non-causal interventions.  In contrast, causal decision theory will tend to judge these actions unworthy of promotion and will recommend allocating our resources elsewhere should these actions come with even the slightest opportunity cost.  

Causal decision theory seems more widely favored than evidential decision theory.[163]  And it might seem that non-causal interventions can influence the expected value of outcomes in decision relevant ways only to those who accept evidential decision theory.  Putting these two points together, one might be tempted to conclude that simulations as non-causal interventions bear on catastrophic risks in decision-relevant ways only given a certain minority view (evidential decision theory).  Contrary to this response, simulations as non-causal interventions have broader significance.  This is for several reasons:

Some might take the promotion and prevention of certain simulations as non-causal interventions as an implausible result that indicates that the argument has gone wrong somewhere.  My own view is that this reaction is understandable but unwarranted.  Admittedly, that simulations as non-causal interventions should be promoted is a weird and wacky idea.  However, this comes with the territory: it’s no surprise that interactions between decision theory, decision theoretic uncertainty, self-locating belief, and simulation hypotheses have bizarre consequences.  If there is something especially objectionable about this one, it remains to be specified. I expect others to disagree here. So there is a project of exploring the prospects for logically weakening the argument while preserving its upshot along with the prospects for a robust escape from the argument.

If simulation as non-causal interventions is admitted as relevant to analyzing and seeking to reduce catastrophic risks, how should we devise a risk-reduction portfolio that is appropriately sensitive to them? This is a large question that I cannot hope to settle here. This is partly because of the residual issues concerning how to deal with decision theoretic uncertainty that we encountered above, and partly because addressing it would require a systematic investigation of what sorts of non-causal interventions via simulation are possible, their expected (causal and non-causal, descriptive and axiological) consequences.  While I will not attempt such an investigation here, I will lay out some considerations that would need to be addressed in such an investigation.

First, we need to consider various types of non-causal interventions that could affect catastrophic risk levels via simulation hypotheses about different types of simulations.  Types to consider include not only salvation and doom simulations, but also:

If simulation as non-causal interventions is admitted as relevant to analyzing and seeking to reduce catastrophic risks, there are some potential mistakes that we should take care to avoid or else ensure that they are not in fact mistakes. These include:

14. Open Questions and Avenues for Future Research

By way of conclusion, I’ll highlight what I regard as some key open questions we’ve encountered along with some promising avenues for future research.  I’ll classify these in a rough and ready way by research area.  A disproportionate number of questions and research avenues will belong to areas of philosophy, which is my own field of expertise. This merely reflects my being in a better position to identify promising topics in philosophy.  This collection is by no means comprehensive, and I would be delighted if others improved upon it.


Adams, M. (2000). Horrendous evils and the goodness of God. Cornell University Press.

Alexander, S. (2019) Don’t Fear The Simulators. Slate Star Codex.

Althaus, D., & Baumann, T. (2020). Reducing long-term risks from malevolent actors. Publications of Center of Long Term Risk.

Anthis, J.R. (2018) “Why I Prioritize Moral Circle Expansion Over Artificial Intelligence Alignment.” Effective Altruism Forum.

Anthis, J.R. (2022) The Future Might Not Be So Great. URL:

Armstrong, S.; Sandberg, A. & Bostrom, N. (2012). Thinking Inside the Box: Controlling and Using an Oracle AI. Minds and Machines 22 (4):299-324.

Babcock, J., Kramár, J., & Yampolskiy, R. V. (2019). Guidelines for artificial intelligence containment. Next-Generation Ethics: Engineering a Better Society, 90-112.

Banks, I.M. (2010) Surface Detail, London: Orbit.

Barak B. & Edelman B. (2022) AI will change the world, but won’t take it over by playing “3-dimensional chess”. URL: 

Barnett, Z. & Li, H. (2016). Conciliationism and merely possible disagreement. Synthese 193 (9):1-13.

Baum, S. D., Denkenberger, D. C., & Haqq-Misra, J. (2015). Isolated refuges for surviving global catastrophes. Futures, 72, 45-56.

Baumann, T. (2022) Avoiding the Worst: How to Prevent a Moral Catastrophe URL: 

BBC News. (2010) "Stephen Hawking warns over making contact with aliens". URL: 

Beckstead, N. (2015). How much could refuges help us recover from a global catastrophe?. Futures, 72, 36-44.

Bedke, (2009). Intuitive non-naturalism meets cosmic coincidence. Pacific Philosophical Quarterly 90 (2):188-209.

Benatar, D. (2008). Better never to have been: The harm of coming into existence. OUP.

Bennett, C.H., Hanson, R., & Riedel, C.J. (2019). Comment on ‘The Aestivation Hypothesis for Resolving Fermi’s Paradox’. Foundations of Physics, 49(8), 820-829.

Bentham, J. (1791). Panopticon: or, The inspection-house. Containing the idea of a new principle of construction applicable to any sort of establishment, in which persons of any description are to be kept under inspection, etc. Thomas Byrne.

Benton, M. A., Hawthorne, J., & Isaacs, Y. (2016). Evil and evidence. Oxford Studies in Philosophy of Religion, Vol. 7, 1-31.

Birch, J. (2013). On the ‘simulation argument’ and selective scepticism. Erkenntnis, 78(1), 95-107.

Bhogal, H. (forthcoming). What's the Coincidence in Debunking? Philosophy and Phenomenological Research.

Bogardus, T. (2016). Only All Naturalists Should Worry About Only One Evolutionary Debunking Argument. Ethics 126 (3):636-661.

Bostrom, N. (2002a). Existential risks: Analyzing human extinction scenarios and related hazards. Journal of Evolution and technology, 9.

Bostrom, N. (2002b) Anthropic bias: observation selection effects in science and philosophy. New York, NY: Routledge.

Bostrom, N. (2003a). Are we living in a computer simulation?. The philosophical quarterly, 53(211), 243-255.

Bostrom, N. (2003b). Ethical issues in advanced artificial intelligence. In Science fiction and philosophy: from time travel to superintelligence.

Bostrom, N. (2005). The simulation argument: Reply to Weatherson. The Philosophical Quarterly, 55(218), 90-97.

Bostrom N. (2008) Where are they? Why I hope the search for extraterrestrial life finds nothing. MIT Technology Review May/June: 72-77.

Bostrom, N. (2011) The Simulation Argument FAQ. URL:

Bostrom, N. (2014). Superintelligence: paths, dangers, strategies. OUP.

Bostrom, N. (2019). The vulnerable world hypothesis. Global Policy, 10(4), 455-476.

Bostrom, N., & Ćirković, M. M. (Eds.). (2008). Global catastrophic risks. OUP.

Bourget, D. & Chalmers, D.., (2021) Philosophers on Philosophy: The 2020 PhilPapers Survey. URL:

Braddon-Mitchell, D., & Latham, A. J. (2022). Ancestor simulations and the Dangers of Simulation Probes. Erkenntnis, 1-11.

Buchak, L. M. (2013). Risk and rationality. OUP.

Bykvist, K. (2017). Moral uncertainty. Philosophy Compass, 12(3), e12408.

Carroll, S.M. (2020). Why Boltzmann brains are bad. In Current Controversies in Philosophy of Science (pp. 7-20). Routledge.

Carter B. (1983) The anthropic principle and its implications for biological evolution. Philos Trans R Soc Lond A Math Phys Sci 310:347–363.

Chalmers, D. (2003). The Matrix as metaphysics. Science Fiction and Philosophy: From Time Travel to Superintelligence.

Chalmers, D. (2006). Perception and the Fall from Eden. Perceptual experience, 49-125.

Chalmers, D. (2010). The Singularity. Journal of Consciousness Studies, 17(9-10), 7-65.

Chalmers, D. (2018a). The meta-problem of consciousness. Journal of Consciousness Studies, 25(9-10).

Chalmers, D. (2018b). Structuralism as a response to skepticism. The Journal of Philosophy, 115, 625–660.

Chalmers, D. (2020). Debunking Arguments for Illusionism about Consciousness. Journal of Consciousness Studies, 27(5-6), 258-281.

Chalmers, D. J. (2022). Reality+: Virtual worlds and the problems of philosophy. Penguin UK.

Chen, E.K. (forthcoming). Time's Arrow and Self-Locating Probability. Philosophy and Phenomenological Research.

Christensen, D. (2016). Conciliation, Uniqueness and Rational Toxicity. Noûs, 50(3), 584-603.

Ćirković, M.M. (2008). Observation selection effects and global catastrophic risks. Global Catastrophic Risks. OUP. pp. 120-145.

Ćirković, M.M. (2015). Linking simulation argument to the AI risk. Futures, 72, 27-31.

Ćirković, M.M. (2018). The great silence: Science and philosophy of Fermi's paradox. OUP.

Clarke-Doane, J. (2020). Morality and Mathematics. OUP.

Christiano, P. (2019) “What failure looks like” Alignment forum.

Cotra, A. (2020) Forecasting Transformative AI with Biological Anchors. Open Philanthropy.

Cotra, A. (2022) Without specific countermeasures, the easiest path to transformative AI likely leads to AI takeover. Alignment forum.

Crawford, L. (2013). Freak observers and the simulation argument. Ratio, 26(3), 250-264.

Critch, A.  (2021)What Multipolar Failure Looks Like, and Robust Agent-Agnostic Processes. Alignment forum.

Crummett, D. (2017). Sufferer-centered requirements on theodicy and all-things-considered harms. Oxford Studies in Philosophy of Religion, 8.

Crummett, D. (2020). The real advantages of the simulation solution to the problem of natural evil. Religious Studies, 1-16.

Cuneo, Terence (2007). The Normative Web: An Argument for Moral Realism. OUP.

Cutter, B., & Crummett, D. (forthcoming) Psychophysical Harmony: A New Argument for Theism. Oxford Studies in Philosophy of Religion.

Cutter, B. (2021). Perceptual illusionism. Analytic Philosophy.

Cutter & Saad (forthcoming) The Problem of Nomological Harmony. Manuscript. Nous.

Dainton, B. (2012). On singularities and simulations. Journal of Consciousness Studies, 19(1-2), 42-85.

Dainton, B. (2020). Natural evil: the simulation solution. Religious Studies, 56(2), 209-230.

Danaher, J. (2015). Why AI doomsayers are like sceptical theists and why it matters. Minds and Machines, 25(3), 231-246.

Daniel, M. (2017). S-risks: Why they are the worst existential risks, and how to prevent them (EAG Boston 2017). Foundational research institute.

Davidson, T. (2023) Continuous doesn’t mean slow. URL: 

Dello-Iacovo, M.A. (2017) On terraforming, wild-animal suffering and the far future. Sentience Politics.

Dreier, J. (2012). Quasi‐realism and the problem of unexplained coincidence. Analytic Philosophy, 53(3), 269-287.

Dogramaci, S. (2017). Explaining our moral reliability. Pacific Philosophical Quarterly, 98, 71-86.

Dogramaci, S. (2020). Does my total evidence support that I’m a Boltzmann Brain? Philosophical Studies 177 (12):3717-3723.

Doody, R. (2022). Don't Go Chasing Waterfalls: Against Hayward's “Utility Cascades”. Utilitas 34 (2):225-232.

Dorr, C., & Arntzenius, F. (2017). Self-locating priors and cosmological measures. The philosophy of cosmology, 396-428.

Easwaran, K. (2021). A classification of Newcomb problems and decision theories. Synthese, 198(27), 6415-6434.

Elamrani, A., & Yampolskiy, R. V. (2019). Reviewing tests for machine consciousness. Journal of Consciousness Studies, 26(5-6), 35-64.

Elga, A. (2000). Self-locating belief and the sleeping beauty problem. Analysis 60 (2):143–147.

Elga, A. (2004). Defeating dr. evil with self-locating belief. Philosophy and Phenomenological Research 69 (2):383–396.

Elga, A. (2008). “Lucky to be rational.” Unpublished manuscript. URL:

Enoch, D. (2011). Taking Morality Seriously: A Defense of Robust Realism. OUP.

Erez, F. (2023). Ought we align the values of artificial moral agents?. AI and Ethics, 1-10.

Everitt, N. (2004). The non-existence of God. Routledge.

Fitzpatrick, L. (2009). "A Brief History of China's One-Child Policy". Time. ISSN 0040-781X

Friederich, S., "Fine-Tuning", The Stanford Encyclopedia of Philosophy (Winter 2018 Edition), E.N. Zalta (ed.), URL = <>.

Gabriel, I. (2020). Artificial intelligence, values, and alignment. Minds and machines, 30(3)411-437.

Gloor, L. (2016) “The Case for Suffering-Focused Ethics” URL: 

Gloor, L. (2018) “Cause prioritization for downside-focused value systems” URL: 

Goff, P. (2018). Conscious thought and the cognitive fine-tuning problem. Philosophical Quarterly,

68(270), 98–122.

Grace, C. (2010). Anthropic reasoning in the great filter. Unpublished manuscript.

Greaves, H. (2016, October). Cluelessness. In Proceedings of the Aristotelian Society (Vol. 116, No. 3, pp. 311-339). OUP.

Greaves, H., Cotton-Barratt (2019) “A bargaining-theoretic approach to moral uncertainty” URL: 

Green, A., & Stump, E. (Eds.). (2015). Hidden Divinity and Religious Belief. Cambridge.

Greene, J. (2007). The secret joke of Kant’s soul. In W. Sinnott-Armstrong (ed.), Moral Psychology, Vol. 3. MIT Press.

Greene, P. (2020). The Termination Risks of Simulation Science. Erkenntnis 85(2):489-509.

Gustafsson, J.E., & Peterson, M. (2012). A computer simulation of the argument from disagreement. Synthese, 184(3), 387-405.

Häggström, O. (2021) AI, orthogonality and the Müller-Cannon instrumental vs general intelligence distinction. URL: 

Hanson, R., (1998). The Great Filter - Are We Almost Past It? URL: 

Hanson, R. (2001). How to live in a simulation. Journal of Evolution and Technology, 7(1), 3-13.

Hanson, R. (2016). The age of Em: Work, love, and life when robots rule the Earth. OUP.

Hanson, R., Martin, D., McCarter, C., & Paulson, J. (2021). If Loud Aliens Explain Human Earliness, Quiet Aliens Are Also Rare. The Astrophysical Journal, 922(2),182.

Hayward, M.K. (2020). Utility cascades. Analysis 80 (3):433-442.

Hill, R. R., & Tolk, A. (2017). A history of military computer simulation. In Advances in Modeling and Simulation (pp. 277-299). Springer.

Hoffman, D. (2019). The case against reality: Why evolution hid the truth from our eyes. WW Norton & Company.

Huang, S., ... & Wei, F. (2023). Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045.

Isaacs, Y.; Hawthorne, J. & Russell, J.S. (forthcoming). Multiple Universes and Self-Locating Evidence. Philosophical Review.

Ichikawa, J. (2011). Quantifiers, Knowledge, and Counterfactuals. Philosophy and Phenomenological Research 82 (2):287 - 313.

James, W. (1890). The principles of psychology (Vol. 2). New York: Henry Holt and Company

Jaquet, F. (2022). Speciesism and tribalism: embarrassing origins. Philosophical Studies, 179(3), 933-954.

Johnson, D.K. (2011). Natural evil and the simulation hypothesis. Philo, 14(2), 161-175.

Joyce, R. (2007). The evolution of morality. MIT press.,

Karnofsky, H. (2021). “All Possible Views About Humanity's Future Are Wild” URL: 

Kagan, S. (2000). Evaluative focal points. Morality, rules, and consequences: A critical reader, 134-55.

Knab, B. (2019)  Three Problems in Formal Epistemology. Dissertation.

Arden Koehler, Benjamin Todd, Robert Wiblin and Keiran Harris  (2020) “BenjaminTodd on varieties of longtermism and things 80,000 Hours might be getting wrong” Podcast. URL:

Korman, D.Z. (2019). Debunking arguments. Philosophy Compass 14 (12).

Kotzen, M. (2020). What Follows from the Possibility of Boltzmann Brains?. In Current Controversies in Philosophy of Science (pp. 21-34). Routledge.

Kraay, K. (2010) "Theism, Possible Worlds, and the Multiverse." Philosophical Studies 147:355-368.

Kraay, K. (ed.) (2014). God and the Multiverse: Scientific, Philosophical, and Theological Perspectives. Routledge.

Ladak, A.(2022) Is Artificial Consciousness Possible? A Summary of Selected Books. URL:

de Lazari-Radek, K., & Singer, P. (2012). The objectivity of ethics and the unity of practical reason. Ethics. 123(1)9-31.

Lawsen, A. (2023) AI x-risk, approximately ordered by embarrassment. Alignment forum.

Lee, G. (2019). 13 Alien Subjectivity and the Importance of Consciousness. Blockheads!: Essays on Ned Block's Philosophy of Mind and Consciousness, 215.

Leike, J., Martic, M., Krakovna, V., Ortega, P. A., Everitt, T., Lefrancq, A., ... & Legg, S. (2017). AI safety gridworlds. URL: 

Lenman, J. (2002). On becoming extinct. Pacific Philosophical Quarterly 83 (3):253–269.

Leslie, J. (1989) Universes. Routledge. [reprinted in 1996]

Leslie, J. (1991). Ensuring two Bird deaths with one throw. Mind 100 (1):73-86.

Leslie, J. (1996), The End of the World: The Science and Ethics of Human Extinction. Routledge.

Lewis, D. (2001). Sleeping beauty: reply to Elga. Analysis, 61(3), 171-176.

Lewis, D. (2004). How many lives has Schrödinger's cat?. Australasian Journal of Philosophy, 82(1), 3-22.

Lewis, P.J. (2013). The doomsday argument and the simulation argument. Synthese, 190(18), 4009-4022.

Li, K., Hopkins, A. K., Bau, D., Viégas, F., Pfister, H., & Wattenberg, M. (2022). Emergent world representations: Exploring a sequence model trained on a synthetic task. arXiv preprint arXiv:2210.13382.

Lockhart, T. (2000). Moral uncertainty and its consequences. OUP.

Long, R. Key questions about artificial sentience: an opinionated guide. Effective Altruism Forum. URL:

Joyce, R. (2001). The Myth of Morality. Cambridge University Press.

MacAskill, W. (2016). Smokers, psychos, and decision-theoretic uncertainty. The Journal of Philosophy, 113(9), 425-445.

MacAskill, W. (2022) What We Owe the Future. Basic Books.

MacAskill, W., Bykvist, K., & Ord, T. (2020). Moral Uncertainty. OUP.

MacAskill, W.; Vallinder, A.; Oesterheld, C.; Shulman, C. & Treutlein, J. (2021). The Evidentialist's Wager. Journal of Philosophy 118 (6):320-342.

Mackie, J.L. (1977). Ethics: Inventing Right and Wrong. Penguin Books.

Madden JC, Enoch SJ, Paini A, Cronin MTD. A Review of In Silico Tools as Alternatives to Animal Testing: Principles, Resources and Applications. Alternatives to Laboratory Animals. 2020;48(4):146-172.

Manley, D. Manuscript. “On Being a Random Sample”

Mason, K. (2010). Debunking arguments and the genealogy of religion and morality. Philosophy Compass, 5(9), 770-778.

Megill, J. (2011). Evil and the many universes response. International Journal for Philosophy of Religion 70:127–138.

Mogensen A. (2019). Doomsday rings twice. Available at   

Mogensen, A. (2019). ‘The only ethical argument for positive 𝛿’?. Working paper.

Miller, J.D. (2019). When two existential risks are better than one. Foresight, 21(1), 130–137.

Miller, J.D., & Felton, D. (2017). The Fermi paradox, Bayes’ rule, and existential risk management. Futures, 86, 44-57.

Monton, B. (2009). Seeking God in science: an atheist defends intelligent design. Broadview Press.

Moravec, H. (1976) “The Role of Raw Rower in Intelligence.” Unpublished manuscript. URL: 

Moravec, H. (1999) Robot: Mere Machine to Transcendent Mind. OUP.

Mowshowitz, Z. (2023) A Hypothetical Takeover Scenario Twitter Poll.

Mørch, H. (2018). The evolutionary argument for phenomenal powers. Philosophical Perspectives, 31,


Muehlhauser, L. (2021) “Treacherous turns in the wild” 

Muehlhauser, L. (2017) A software agent illustrating some features of an

illusionist account of consciousness, OpenPhilanthropy, [Online],

Müller, V.C. & Cannon, M. (2021). Existential risk from AI and orthogonality: Can we have it both ways? Ratio 35(1):25-36.

Murphy, P., & Black, T. (2012). Sensitivity meets explanation: an improved counterfactual condition on knowledge. In K. Becker & T. Black (eds.), The Sensitivity Principle in Epistemology. Cambridge.

Newberry, T., & Ord, T. (2021). The Parliamentary Approach to Moral Uncertainty. Technical Report 2021-2, Future of Humanity Institute, University of Oxford. URL: https://www. fhi. ox. ac. uk/wpcontent/uploads/2021/06/Parliamentary-Approach-to-Moral-Uncertainty. pdf

Nowak, A., Gelfand, M. J., Borkowski, W., Cohen, D., & Hernandez, I. (2016). The evolutionary basis of honor cultures. Psychological science, 27(1), 12-24.

Nozick, R. (1981). Philosophical Explanations. Harvard University Press.

Nozick, R. (1993). The nature of rationality. Princeton University Press.

Olds, J., & Milner, P. (1954). Positive reinforcement produced by electrical stimulation of septal area and

other regions of rat brain. Journal of Comparative & Physiological Psychology, 47, 419–427.

Owen Cotton-Barratt (2021) Web of Virtue thesis. URL:

Parfit, D. (1984). Reasons and persons. OUP.

Parfit, D. (2011) On what matters: Vol. 2. OUP.

Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442.

Passini, E., Britton, O. J., Lu, H. R., Rohrbacher, J., Hermans, A. N., Gallacher, D. J., ... & Rodriguez, B. (2017). Human in silico drug trials demonstrate higher accuracy than animal models in predicting clinical pro-arrhythmic cardiotoxicity. Frontiers in physiology, 668.

Pautz, A. (2014). The real trouble with phenomenal externalism: New empirical evidence

for a brain-based theory of consciousness. In Consciousness inside and out: Phenomenology, neuroscience, and the nature of experience (pp. 237-298). Springer, Dordrecht.

Pautz, A. (2020). Consciousness and coincidence: Comments on Chalmers. Journal of Consciousness Studies, (5-6).

Pautz, A. (2021). Perception. Routledge.

Peterson, M. (2019). The value alignment problem: a geometric approach. Ethics and Information Technology, 21, 19-28.

Perez, E. (2022) A Test for Language Model Consciousness. URL:

Pettigrew, R. (2022) Effective altruism, risk, and human extinction. URL: 

Prinz, J. (2012). Singularity and inevitable doom. Journal of Consciousness Studies, 19(7-8), 77-86.

Ord, T. (2020). The precipice: existential risk and the future of humanity. Hachette Books.

Olson, J. (2014). Moral Error Theory: History, Critique, Defence. OUP.

Rowland, R. (2019). Local evolutionary debunking arguments. Philosophical Perspectives, 33(1), 170-199.

Richmond, A.M. (2017). Why doomsday arguments are better than simulation arguments. Ratio, 30(3), 221-238.

Riedener, S. (2021) Existential risks from a Thomist Christian perspective. URL:

Saad, B. (2019). A teleological strategy for solving the meta-problem of consciousness. Journal of Consciousness Studies, 26(9-10), 205-216.

Saad, B. (2020). Two solutions to the neural discernment problem. Philosophical Studies, 177(10), 2837-2850.

Saad, B. (forthcominga) Harmony in a Panpsychist World. Synthese.

Saad, B. (forthcomingb) “The sooner the better: an argument for bias toward the earlier”. Journal of the American Philosophical Association

Saad, B. (forthcomingc) “Lessons from the void: what Boltzmann brains teach” Analytic Philosophy.

Saad, B. & Bradley, A. (2022). The Problem of Digital Suffering. Inquiry.

Sampson, E. (2022) Disaster Prevention and the Possibility of Hell: A Dilemma for Longtermist Effective Altruists. Manuscript.

Sandbrink, J., Hobbs, H., Swett, J., Dafoe, A., & Sandberg, A. (2022). Differential technology development: A responsible innovation principle for navigating technology risks. Available at SSRN.

Sandberg, A., Armstrong, S., & Cirkovic, M.M. (2017). That is not dead which can eternal lie: the aestivation hypothesis for resolving Fermi's paradox. arXiv preprint arXiv:1705.03394.

Sandberg, A., Drexler, E., & Ord, T. (2018). Dissolving the Fermi paradox. arXiv preprint arXiv:1806.02404.

Schellenberg, J. L. (1996). Divine Hiddenness and Human Reason. International Journal for

Philosophy of Religion 40(2):121-124.

Schellenberg, J. L. (2010). The Hiddenness Problem and the Problem of Evil. Faith and Philosophy


Schelling, T.C., (1971). “Dynamic Models of Segregation,” Journal of Mathematical Sociology, 1: 143–186.

Schwartz, S. H., & Boehnke, K. (2004). Evaluating the structure of human values with confirmatory factor analysis. Journal of research in personality, 38(3), 230-255.

Schwitzgebel, E. (2017). 1% Skepticism. Noûs, 51(2), 271-290.

Schwartz, S. H., & Boehnke, K. (2004). Evaluating the structure of human values with confirmatory factor analysis. Journal of research in personality, 38(3), 230-255.

Seth, A. K., & Bayne, T. (2022). Theories of consciousness. Nature Reviews Neuroscience, 1-14.

Shafer-Landau, R. (2012). Evolutionary debunking, moral realism and moral knowledge. J. Ethics & Soc. Phil., 7, i.

Shiller, D. (2017). In Defense of Artificial Replacement. Bioethics, 31(5)393-399.

Shulman, C. (2010). Whole brain emulation and the evolution of superorganisms. Mach. Intell. Res. Inst. Work. Pap. Httpintelligence OrgfilesWBESuperorgs Pdf.

Shulman, C., & Bostrom, N. (2012). How hard is artificial intelligence? Evolutionary arguments and selection effects. Journal of Consciousness Studies, 19(7-8), 103-130.

Shulman, C., & Bostrom, N. (2021). Sharing the World with Digital Minds. In Rethinking Moral Status.

Sider, T. (2020). The tools of metaphysics and the metaphysics of science. OUP.

Silva Jr, P. (forthcoming). Debunking Objective Consequentialism: The Challenge of Knowledge-Centric Anti-Luck Epistemology. In M. Klenk (ed.), Higher Order Evidence and Moral Epistemology. Routledge.

Singer, P. (2005). ‘Ethics and Intuitions’. The Journal of Ethics 9: 331–52.

Snyder-Beattie, A. E., Sandberg, A., Drexler, K. E., & Bonsall, M. B. (2021). The timing of evolutionary transitions suggests intelligent life is rare. Astrobiology, 21(3), 265-278.

Sotala, K., & Gloor, L. (2017). Superintelligence as a cause or cure for risks of astronomical suffering. Informatica, 41(4).

Steinhart, E. (2010). Theological implications of the simulation argument. Ars Disputandi, 10(1), 23-37.

Street, S. (2006). A Darwinian dilemma for realist theories of value. Philosophical studies, 127(1), 109-166.

Street, S. (2010). What is constructivism in ethics and metaethics? Philosophy Compass 5 (5):363-384.

Street, S. (2011). Mind-Independence Without the Mystery: Why Quasi-Realists Can’t Have it Both Ways. In Russ Shafer-Landau (ed.), Oxford Studies in Metaethics, Volume 6. OUP.

Streumer, B. (2017). Unbelievable Errors: An Error Theory About All Normative Judgments. OUP.

Stump, E. (1985). The problem of evil. Faith and philosophy, 2(4), 392-423.

Summers, M., & Arvan, M. (2021). Two New Doubts about Simulation Arguments. Australasian Journal of Philosophy, 1-13.

Tersman, Folke (2017). Debunking and Disagreement. Noûs 51 (4):754-774.

Thomas, T. (2022) Simulation expectation. URL: 

Titelbaum, M.G. (2013). Ten reasons to care about the Sleeping Beauty problem. Philosophy Compass, 8(11), 1003-1017.

Todd, B. (2020). The emerging school of patient longtermism. 80,000 Hours.

Tomasik, B. (2016). How the Simulation Argument Dampens Future Fanaticism. URL:

Tomasik, B. (2017). What Are Suffering Subroutines. URL: 

Tufekci, Z. (2022) Open Thread. URL:

Turchin, A. (2016) “The Map of Shelters and Refuges from Global Risks”. URL:

Turchin, A. (2018). A Meta-Doomsday Argument: Uncertainty About the Validity of the Probabilistic Prediction of the End of the World. URL: 

Turchin, A., Batin, M., Denkenberger, D., & Yampolskiy, R. (2019). Simulation Typology and Termination Risks. arXiv preprint arXiv:1905.05792.

Udell, D. B. & Schwitzgebel, E. (2021). Susan Schneider's Proposed Tests for AI Consciousness: Promising but Flawed. Journal of Consciousness Studies 28 (5-6):121-144.

Urbina, F., Lentzos, F., Invernizzi, C., & Ekins, S. (2022). Dual use of artificial-intelligence-powered drug discovery. Nature Machine Intelligence, 4(3), 189-191.

Van Fraassen, Bas C. (1989). Laws and Symmetry. OUP.

Vavova, K. (2014). Debunking evolutionary debunking. Oxford studies in metaethics, 9(7), 6-101.

Vavova, K. (2015). Evolutionary debunking of moral realism. Philosophy Compass, 10(2), 104-116.

Vemprala, S., Bonatti, R., Bucker, A., & Kapoor, A. (2023). Chatgpt for robotics: Design principles and model abilities. Microsoft.

Virués-Ortega, J., Buela-Casal, G., Garrido, E., & Alcázar, B. (2004). Neuropsychological functioning associated with high-altitude exposure. Neuropsychology review, 14(4), 197-224.

Ward P, & Brownlee D. (1999) Rare earth. In Copernicus, Springer-Verlag, New York, NY.

Webb, S. (2015). If the universe is teeming with aliens... where is everybody?: fifty solutions to the Fermi paradox and the problem of extraterrestrial life. 2nd ed.  New York, NY: Copernicus Books.

White, R. (2006). Problems for dogmatism. Philosophical Studies, 131, 525–557.

White, R.  (2010). You just believe that because…. Philosophical Perspectives 24 (1):573-615.

White, R. (2018). Reasoning with Plenitude. In Knowledge, Belief, and God: New Insights in Religious Epistemology

Wright, C. (2004). Warrant for nothing. Proceedings of the Aristotelian Society, Supplementary Volume,

78, 167–212.

Xia, L., Robock, A., Scherrer, K., Harrison, C. S., Bodirsky, B. L., Weindl, I., ... & Heneghan, R. (2022). Global food insecurity and famine from reduced crop, marine fishery and livestock production due to climate disruption from nuclear war soot injection. Nature Food, 3(8), 586-596.

Yampolskiy, R. V. (2014). Utility function security in artificially intelligent agents. Journal of Experimental & Theoretical Artificial Intelligence, 26(3), 373-389.

Yampolskiy, R.V. (2018). Why we do not evolve software? Analysis of evolutionary algorithms. Evolutionary Bioinformatics, 14, 1176934318815906.

Yudkowsky, E. (2004). Coherent extrapolated volition. Singularity Institute for Artificial Intelligence.

Zackrisson, E., Calissendorff, P., González, J., Benson, A., Johansen, A., & Janson, M. (2016). Terrestrial planets across space and time. The Astrophysical Journal, 833(2), 214.

Ziegler, D. M., Stiennon, N., Wu, J., Brown, T. B., Radford, A., Amodei, D., ... & Irving, G. (2019). Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593.

Zhang, L. What Rethink Priorities General Longtermism Team Did in 2022, and Updates in Light of the Current Situation. URL: 

Zuber, S., Venkatesh, N., Tännsjö, T., Tarsney, C., Stefánsson, H. O., Steele, K., ... & Asheim, G. B. (2021). What should we agree on about the repugnant conclusion?. Utilitas, 33(4), 379-383.


This project has been supported by the Long-Term Future Fund, the Sentience Institute, and Utrecht University. The views expressed in this report do not necessarily reflect those of funders.

[1] For discussion of various sorts of global catastrophic risks, see the essays in Bostrom & Ćirković (2008).

[2] See Baumann (2022), Gloor (2016; 2018), and Daniel (2017).

[3] See Ord (2020).

[4] Research is already underway on small-scale (25 agents) social simulations that embed instances of OpenAI’s ChatGPT in virtual social contexts (Park, et al., 2023).

[5] See Davidson (2023).

[6]  See Huang (2023),, Li et al. (2022), and Vemprala et al. (2023)

[7] See Chalmers (1996: Chs. 7, 9) and  Sandberg & Bostrom (2008).

[8] These sum to more than 100% because the first two possibilities are not exclusive: we could on large scales both unwittingly run some simulations containing conscious minds and intentionally run others. Or multiple agents running the same simulations could hold different views about whether those simulations contain conscious minds.

[9] The situation here parallels the situation with AI-involving catastrophic risks more generally.  While the sources of risk do not depend on the specifics of concrete scenarios we can conjure, it is still advisable to describe concrete scenarios since the uninitiated may find abstract risks difficult to take seriously without first seeing how they could be concretely realized.  And it is advisable to emphasize that the general catastrophic risks do not depend on the details of those scenarios, lest the target audience take challenges to details of those scenarios to be levers for driving down estimates for the general risks they illustrate.  See Christiano (2019), Cotra (2022), Critch (2021), Hendrycks (2023), Lawsen (2023), and Mowshowitz (2023).

[10] For other presentations and discussion of scenarios, see Bostrom (2003a), Chalmers (2010b; forthcoming), Dainton (2012), and Hanson (2016).  

[11] See  Karnofsky (2021).

[12] For a book-length discussion of a future with whole brain emulations, see Hanson (2016).

[13] Simulations of nuclear war have already been used to research catastrophic risks. For example, Xia et al. (2022) used simulations to evaluate the impact of nuclear war on global food supply and to arrive at the estimate that five billion people would die in a large-scale nuclear war between the United States and Russia. For a news cycle, the article was widely discussed in popular media. This points to another way in which research simulation could reduce catastrophic risks: by providing easily understood statistics about concrete scenarios, simulation research may elicit stronger responses from the public and policymakers than, say, compelling arguments that deal with a given type of risk in the abstract.

[14] See Bostrom (2019).

[15] For theoretical discussions, see Chalmers (2010b: §7), Bostrom (2014: 116-119), and Babcock et al. (2019). AI safety testing in virtual environments is already being explored in practice. For example, OpenAI has developed AI Safety Gridworlds, a suite of virtual environments for establishing safety in AI systems (Leike, 2017).

[16] See, e.g., Bostrom (2014) and Muehlhauser (2021).

[17] Cf. Armstrong et al. (2012).

[18] See Bentham (1791) and Bostrom (2014: 134-5).

[19] See Olds & Milner, P. (1954) and Yampolskiy (2014).

[20] Cf. Bostrom (2014: 134-5).

[21] Because official public messaging during disasters is low-bandwidth, it often involves a tradeoff between direct and indirect effects. This was evident during the Covid-19 pandemic when public health institutions issued unwarranted statements that could be charitably interpreted as aiming at indirect effects—cf. Tufekci (2022)  Substantiating a heuristic concerning direct vs. indirect effects could help protect against basing public messaging during catastrophes on overestimations of indirect effects of the messaging.

[22] See Bostrom (2014: Ch. 14) and Sandbrink et al (2022).

[23] Cf. Anthis (2022), Bostrom (2014: Chs. 12-3), Dello-Iacovo (2017), Doody (2022), Hayward (2020), and MacAskill (2022: Chs. 3-4). A well-known toy simulation of this sort can be found in Schelling (1971). In it, agents manifested different preferences for segregation into groups with members of the same kind through movement rules. Surprisingly, he found that slight preferences for segregation induced segregation from a non-segregated state.

[24] For a simulation study of the dynamics of moral disagreement, see Gustafsson & Peterson (2012). For a simulation study of honor culture, see Nowak et al. (2016).

[25] See Schwartz & Boehnke (2004). Relatedly, simulation testing of different civilizational balances between exploration and exploitation (e.g. in institutional design) seems promising.

[26] See Todd (2020) for discussion and references.

[27] See Koehler et al. (2020).

[28] See Anthis (2018) and Owen Cotton-Barratt (2021).

[29] For a meta-analysis of empirical work on the effectiveness of simulation-based learning, see Chernikova et al. (2020).  

[30] For discussion of artificial evolution as a method for developing AI, see Bostrom (2014: 24-8, 37-44,  Chalmers (2010b: 16-17), Shulman & Bostrom (2012), and Shulman (2010). See Yampolskiy (2018) for pessimism about artificially evolving software. For cautionary observations about using artificial selection, see Bostrom (2014: 153-5).

[31] For presentations of Fermi’s paradox and book-length discussions of candidate solutions, see Ćirković (2018) and Webb (2015).  For reasons to think that the paradox arises because of mishandling of uncertainties in calculations that are used to pose it, see Sandberg et al. (2018).  For simulations of a “grabby alien’s” solution to Fermi’s Paradox, see Hanson et al. (2021).

[32]  See Bostrom (2002a), Grace (2010), and Hanson (1998).

[33] See, e.g., Miller & Felton (2017).

[34] For a collection of resources on longtermism, see

[35] See Bostrom (2008).

[36] See Jumper et al. (2021).

[37] For an overview of different sorts of debunking arguments, see Korman (2019). For discussion of moral debunking arguments, see, e.g., Joyce (2007), Shafer-Landau (2012), Street (2006), and Vavova (2015).  For discussion of religious debunking arguments, see Mason (2010), references therein, and White (2010).

[38]For example, perhaps evolutionary considerations could debunk some intuitions about the non-physicality of consciousness—cf. Chalmers (2018a; 2020).  If so, this could be relevant to whether simulation inhabitants have moral status, since it is more plausible that consciousness has special moral significance if it is a basic non-physical property than if it is a physical property.

[39] In the case of normative beliefs, there is an analogy between evolutionary forces being orthogonal to true moral beliefs (even if those forces promote instrumental rationality) and the orthogonality thesis in AI that levels of intelligence and final goals can generally be arbitrarily combined (even if intelligence engenders instrumental rationality)—for discussion of the latter thesis, see Bostrom (2014: Ch. 7), Häggström (2021), and Müller & Cannon (2021).

[40] See Parfit (2011: 494-6); cf. Müller & Cannon (2021).

[41] A field of study that is relevant here is artificial life, which is partly concerned with analyzing life-like agents through simulations of evolutionary processes. See for an overview.

[42] For discussion of evolutionary debunking arguments against normative rather than metaethical theories, see Greene (2007), de Lazari-Radek & Singer (2012), Rowlands (2019), and Silva (forthcoming), and Singer (2005).

[43] Cf Cuneo (2007).

[44] See Joyce (2001), Mackie (1977), Olson (2014), and Streumer (2017).

[45] See, e.g., Street (2010).

[46] Cf. Dreier (2012) and Street (2011).

[47] See Cuneo (2007), Dogramaci (2017), Vavova (2014; 2015), Shafer-Landau (2012), and White (2010).

[48] There is a burgeoning technical subfield of AI safety devoted to the alignment problem. Much of the research on this topic can be found at

[49] See Bostrom (2003b).

[50] Cf. Bostrom (2014: 2019-20).

[51] Unless, of course, the human preferences in question are idealized via alignment with the objective moral facts—in that case, there would not be room for moral catastrophe to result from aligning powerful AI with those human preferences but not with the moral facts. This contrasts with idealizations that modify preferences by imposing non-moral constraints such as coherence among preferences, the elimination of lower-order preferences that conflict with higher-order preferences, reflective, empirically informed endorsement of preferences—cf. Yudkowsky, E. (2004).  On moral realism, there is no guarantee that aligning AI with human preferences that result from the latter sorts of idealizations would align the AI with the moral facts—cf. Bostrom (2014: 2018), Erez (2023), Gabriel (2020), Peterson (2019), and Shafer-Landau (2003: 42).  The risk level for a catastrophe from this kind of alignment failure depends partly on the plausibility of moral realism.  Some relevant data: in a recent survey of professional philosophers, 62.1% favored moral realism while only  26.1% favored moral antirealism (Bourget & Chalmers, 2021).

[52] See Dretske (1971), Murphy & Black (2012), Nozick (1981), and Ichikawa (2011).

[53] Reflection on skeptical scenarios suggests that insensitivity on its own is not enough to make a belief epistemically defective: I would believe I’m not a brain in a vat with exactly this experience even if I were; yet my belief that I am not a brain in a vat is not epistemically defective. On the other hand, reflection on non-skeptical cases suggests that insensitivity can result in epistemic defect: for example, if evolution explains widespread robust belief in the moral superiority of homo sapiens over other species and that belief would have been widespread even if it were false, that would raise a serious challenge to the belief—see Jaquet (2022). This is so even if the belief is held in all nearby scenarios in which evolution went differently.

[54] See Barnett & Li (2016) Bedke (2009), Bhogal (forthcoming), Bogardus (2016), Clarke-Doane (2020), Enoch (2011), and Tersman (2017).

[55] Theories of consciousness tend to fall into one of three areas: the metaphysics of mind, the philosophy of perception, or the science of consciousness. For an overview of theories in the metaphysics of mind, see Chalmers (2010a: Ch. 5).  For an introduction to theories in the philosophy of perception, see Pautz (2021).  For an overview of theories in the science of consciousness, see Seth & Bayne (2022).

[56] For an overview of a range of books over the last few decades that address consciousness in artificial systems, see Ladak (2022). For an overview of key issues and open questions concerning artificial consciousness, see Long (2022).

[57] For discussion of tests for consciousness, see Saad & Bradley (2022), Chalmers (2018a: 34-5), Elamrani & Yampolskiy (2019), Muehlhauser (2017), Perez (2022) Schneider (2019: Ch. 4), and Udell & Schwitzgebel (2021).

[58] The same goes for simulation technologies more generally.  However, I’ll just focus on dual-use risks posed by research simulation technologies.  I do this for tractability and because, among simulation technologies, research simulation seems like an especially large source of dual use risk.

[59] See Urbina et al. (2022).

[60] See Ord (2020: Ch. 5) and MacAskill (2022: Ch. 5).

[61] See Barak and Edelman (2022).

[62] See

[63] There are already cases in which computer simulations outperform animal experiments. For instance, see Passini (2017) for a case in which (Virtual Assay) human-based computer simulations outperformed animal experimentation on predicting drug-induced cardiotoxicity in humans.  For a review of existing in silico alternatives to animal testing, see Madden et al. (2020).

[64] Some such requirements were relaxed in 2022 by the FDA Modernization Act 2.0, which lifts the United State’s federal mandate on testing experimental drugs in animals before testing them in humans.

[65] See

[66] For an interactive, simulation-based model of AGI arrival timelines, see

[67] A notable potential risk of gamified simulations (and virtual reality simulations more generally) is that they could become so encompassing as to disempower humanity: if all of humanity, or even just key actors, became more interested in pursuing goals within gamified simulations than in pursuing goals, our civilization would lose much of its ability to respond to risks.

[68] See Baum et al. (2015).  For a chart of different types of refuge, including digital shelters, see Turchin (2016). Rethink Priorities is one organization that has recently explored refuges as a risk-reduction strategy (Zhang, 2022).

[69] See Bostrom & Shulman (2021).

[70] For potential advantages of aestivation, see Sandberg et al. (2017). For criticism of their analysis, see Bennett et al. (2019).

[71] See, e.g.,  Chalmers (2003; 2022) and Dainton (2012).

[72] Cf. Shulman & Bostrom (2021).

[73] See, e.g., Chalmers (1996) and Saad & Bradley (2022).

[74]  For discussion in the context of superintelligence and its bearing on the risk of astronomical quantities of suffering, see Sotala & Gloor (2017).

[75] See Hill & Tolk (2017) for a history of military simulations.

[76] See Perez (2022).

[77] See Bostrom (2014: Ch. 11) and Hanson (2016).

[78] See Dainton (2012) and his discussion of Banks (2010).

[79] Compare: OpenAI, a leading company among those trying to develop AGI, inadvertently trained a large language model (GPT-2) to optimize for expressing negative sentiment as a result of a flipped sign and AI developers (literally) being asleep during the training process (Ziegler et al., 2019).

[80] See, e.g., Tomasik (2017).

[81] For instance, Mao Zedong’s policies led to population growth followed (perhaps in ways that were foreseeable but not intended) by famines in which tens of millions of Chinese citizens died (Fitzpatrick, 2009).

[82]On the other hand, creating suitable ensembles of simulated agents with malevolent impulses that are punished when they act on those impulses could incentivize even powerful would-be malevolent actors to behave morally by giving them reason to think they are in such a simulation—see Elga (2004) and the discussion of the simulation argument below. For discussion of existential and suffering risks posed by malevolent actors, see Althaus & Baumann (2020).  

[83] The argument was introduced and defended by Bostrom (2003a).  Others who take it seriously include Braddon-Mitchell & Latham (2022), Chalmers (2022), Ćirković (2015), Crummet (2020), Dainton (2012; 2020), Greene (2020), Hanson (2001), Johnson (2011), Lewis (2013), Monton (2009: Ch. 3). Schwitzgebel (2017), Steinhart (2010), Thomas (2022), and Turchin et al. (2019).

[84] There is debate in the literature about whether this hypothesis should, given its role in the argument, be made conditional on the reasoner’s not being a simulation—see Thomas (2022); cf. Bostrom (2011) and  Crawford (2013). For tractability and simplicity, I set this issue aside and work with the unconditional formulation.

[85] More precisely, the bland indifference principle says that for a given hypothesis about how the world is qualitatively, we should divide our credence concerning our self-location on that hypothesis evenly among the observers like us that exist on that hypothesis. This should not be confused with indifference principles that require one to divide one’s credence evenly between self-locations posited on different qualitative hypotheses or between different qualitative hypotheses when one’s evidence does not adjudicate between them.  These stronger principles are open to serious objections that do not apply to the bland indifference principle (hence the label ‘bland’)—see Elga (2004: 387-8) and Van Fraassen (1989).  For discussion and defense of restricting rather than rejecting indifference principles, see Greaves (2016).

[86] It may turn out that being an observer like us is a matter of degree, either because being an observer is a matter of degree (an outcome that is hard to avoid on a reductive physicalist view of observerhood—see Lee (2019)) or because counting as like us is a matter of degree. In either case, the most plausible rendering of the indifference principle may then require one to divide one’s self-locating credence among observers that are to some degree like oneself in proportion to the degree to which they are like oneself—see Dorr & Arntzeneius (2017).  While I think such a proportional rather than egalitarian indifference principle may well turn out to be correct, I do not think the difference between them is important for the discussion that follows. In any event, for simplicity, I will work with the just sketched egalitarian principle.

[87] See Bostrom (2003a) and Chalmers (2022).

[88] See Chalmers (2022).

[89] See Bostrom (2003a).

[90] See Dogramaci (2020).

[91] See Summers & Arvan  (2021), Dainton (2012), and Chalmers (2022).

[92] See Chalmers (2022), Richmond (2017), and Hanson (2001).

[93] See Birch (2013), Crawford (2013), and Thomas (2022).

[94] See Crawford (2013).  For discussion, see Chalmers (2022: online appendix) and Dainton (2012).  For a response to this sort of objection, see Bostrom (2009).  For related discussions of cognitive instability in the context of Boltzmann brains, see Carroll (2020), Chalmers (2018b), Dogramaci (2020), Kotzen (2020), Saad (forthcomingc).

[95] See Bostrom (2003a) and Chalmers (2022).

[96] See Chalmers (2022).

[97] See, e.g., Ćirković (2008: 138).

[98] See Dainton (2012); cf. Chalmers (2022: online appendix).

[99] For a summary of literature on neuropsychological responses to hypoxia, see Virués-Ortega (2004).

[100] For relevant discussion, see Elga (2008), Kotzen (2020), and Christensen (2016).

[101] See Bostrom (2005).  However, there are exceptions, notably skeptical arguments from dreams, the evolutionary origins of our beliefs, and physical theories that proliferate Boltzmann Brains.

[102] See Chalmers (2003; 2022).

[103] See, e.g., Chalmers (2006), Cutter (2021), Hoffman (2019), and Pautz (2014).

[104] See Bostrom (2002a), Ćirković (2008), Greene (2020), and Turchin et al. (2019).

[105] See Tomasik (2016); cf. Ord (2020: Appendix B).

[106] See, e.g., Ćirković (2008).

[107] See Hanson (2001) and Greene (2020).

[108] For reasons to think simulation awareness is a risk, see Greene (2020). For reasons to think it is not, see Alexander (2019) and Braddon-Mitchell & Latham (2022).

[109] The most thorough discussions of this that I am aware of are Braddon-Mitchell & Latham (2022), Greene (2020), and Turchin et al. (2019).

[110] See Lenman (2002).

[111] See Lenman (2002) and Prinz (2012).  Those who countenance agent-relative moral reasons may see middle ground here—see, e.g., Mogensen (2019b) and references therein.

[112] For moral concerns about directed panspermia, see Dello-Iacovo (2017) and O’Brien (forthcoming).

[113] For reasons to doubt that our universe’s past or future are net positive, see Anthis (2018; 2022), Benatar (2008), Gloor (2016; 2018), and MacAskill (2022: Ch. 9). For motivation for risk-averse decision theory and arguments that it recommends extinction-hastening interventions over extinction-preventing ones (given longtermist moral assumptions), see Pettigrew (2022).

[114] One way to motivate asymmetrically diminishing returns is to note that (1) some bads, such as uncompensated suffering, seem not to diminish at all in disvalue as their quantity increases and (2) countenancing diminishing returns for goods provides an escape from what’s known in population ethics as the Repugnant Conclusion, which is the (supposedly) implausible claim that “For any possible population of at least ten billion people, all with a very high quality of life, there must be some much larger imaginable population whose existence, if other things are equal, would be better even though its members have lives that are barely worth living” (Parfit, 1984: 342). For objections to one form of asymmetrical diminishing returns, see ibid (§134). For reasons to doubt that avoiding the Repugnant Conclusion is a requirement for an adequate approach to population ethics, see Zuber et al. (2021).

[115] Cf. Ord (2020: Ch. 2).

[116] The Unilateralist Curse provides a further reason for caution here: when large numbers of at least somewhat error-prone, altruistically motivated agents are each in a position to unilaterally act in ways that affect others, we should expect unilateral interventions to happen more often than is optimal (Bostrom et al., 2016).  In the case at hand, the curse we face is: if a sufficiently large number of error-prone, altruistic agents will be in a position to unilaterally take extinction-inducing interventions and their evidence renders extinction non-optimal in expectation, we should expect some such agent to mishandle their evidence, overestimate the value of such interventions, take them, and thereby cause extinction.  For general suggestions for avoiding the Curse that can be applied in this case, see ibid.

[117] Religious discussions of catastrophic risks that I am familiar with tend not to discuss non-religious catastrophic risks. There are a few notable exceptions. Riedener (2021) argues that existential risk reduction is extremely important from a Thomist Christian perspective.  Danaher (2015) brings some ideas from philosophy of religion to bear on catastrophic risk posed by AI: he explores an analogy between the “skeptical theist” view that appearances of evil are not strong evidence against the existence of a benevolent God with Bostrom’s (2014: Ch. 8) “treacherous turn” concern that AI systems may behave cooperatively (e.g. while boxed within a simulated environment) and hence appear benign before abruptly changing their behavior (in a potentially catastrophic manner) to pursue their final goals.  The (pseudonymous) author of this post argues that longtermist movements have a dismal track-record, that religions have been (for good or ill) influential long-termist movements, and that catastrophic risks that concern longtermists may be more effectively mitigated through near-term focused approach.

[118] For an argument that religious catastrophic risks are important, tractable, and neglected relative to other catastrophic risks, see Sampson (2022).

[119] For discussion of the religious implications of the simulation argument or hypothesis, see Bostrom (2003: 254), Chalmers (2022: Ch. 7), Crummett (2020), Dainton (2020), Johnson (2011), and Steinhart (2010).  

[120] This support may be somewhat attenuated by the fact that religions that posit catastrophes tend not to countenance a flawed or indifferent simulator of our universe; hence, the supposition that our universe has such a creator tells against those specific catastrophes and so in that respect detracts from overall risk.

[121] For multiverse solutions to theistic problems that do not invoke simulations, see Kraay (2010), Leslie (1989),  and Megill (2011). For an overview of surrounding literature, see Kraay (2014: 9-11).

[122] The problem of evil is variously formulated—see, e.g., Benton et al. (2016).  The same goes for the other problems considered below.

[123] See Everitt (2004: Ch. 11).

[124] See, e.g., the essays in Green & Stump (2015) along with Schellenberg (1996; 2010).

[125] See, e.g., Adams (2000), Crummett (2017), and Stump (1985).

[126] For an overview, see Friederich (2018).  There are also less explored non-cosmological varieties of fine-tuning such as the match between the universe’s boundary conditions with its laws (Cutter & Saad, forthcoming) and the fine-tuning of experiences’ causal profiles with their rational profiles (Chalmers, 2020; Goff, 2018, Saad (2019; 2020; forthcominga), James (1890), Mørch (2018), Pautz (2010; 2020).  Cutter & Crummett (forthcoming) defend an argument for theism that appeals to psychophysical fine-tuning.  While both cosmological and psychophysical fine-tuning can be used to argue for theism, they interact with the simulation hypothesis in different ways.  For instance, whereas an ensemble of simulation universes might explain cosmological fine-tuning, it is not clear that they could explain psychophysical fine-tuning—e.g., if the psychophysical laws have a functionalist character at the base level, then they will not be manipulable independently of other factors in the simulations.  Similarly, if they involve a non-functional feature that cannot be freely varied in the simulation, the simulators will not be able to vary it. Even if simulators could in principle vary psychophysical laws in simulations, epistemological obstacles associated with consciousness may prevent simulators from figuring out how to do so.  And even if simulators could vary the psychophysical laws in simulations and figure out how to do so, they may have no incentive to do so, as they may only be concerned with outputs of the simulation. For simplicity, I mostly hereafter set aside psychophysical fine-tuning and how different varieties of it interact with the simulation hypothesis and associated catastrophic risks.

[127] For discussion of the simulation hypothesis and fine-tuning, see Chalmers (2022: Ch. 7) and Steinhart (2010). For an objection to some multiverse hypotheses on the ground that they require a complicated set of basic laws, see Cutter & Saad (forthcoming).

[128] For discussion of this theoretical vice, see, e.g., Chalmers (1996: 213-4), Cutter & Saad (forthcoming), and Sider (2020: 102).

[129] See Chalmers (2022: Ch. 7).

[130] Though the same goes for the simulation argument.

[131] Notable work on the topic includes Bostrom (2002b), Dorr & Arntzenius (2017), Elga (2000; 2004), Lewis (2001), Titelbaum (2013), Isaacs et al. (forthcoming), and Manley (ms).

[132] For discussion of the inward-outward distinction, see Manley (ms).

[133] See Bostrom (2002b: Ch. 4) and Dorr & Arntzenius (2017).

[134] This is close to what Bostrom (2002b: 66) calls the self-indication assumption, which claims that  “[g]iven the fact that you exist, you should (other things equal) favor hypotheses according to which many observers exist over hypotheses on which few observers exist.” Like number boosting, this formulation of the self-indication assumption does not specify how big of a boost hypotheses with more observers should receive.  However, Bostrom goes on to offer as a formalization of the self-indication assumption a principle which specifies the size of the boost for pairs of hypotheses and argue that the self-indication assumption should be rejected because of the implausible extent to which it favors certain hypotheses over others (ibid: 122-6).  The implausible consequences Bostrom derives from what he offers as a formalization of the self-indication assumption do not follow from his initial formulation of the principle, since it does not specify the size of the boost.  For the same reason, his criticisms of the assumption do not apply to number boosting, and analogous criticisms would not apply to proportion boosting.

[135] To yield a precise boost from number or proportion boosting in a given case, five parameters need to be set:

(i) your observer type(s),

(ii) your observation type(s),

(iii) the boosted observer type(s),

(iv) the boosted observation type(s), and

(v) the size of the boost.

It is extremely difficult to find a plausible general principle for setting these parameters (cf. ibid). That said, reflection on cases show that there are reasonable and unreasonable ways of setting the parameters and applying number and proportion boosting.  This mirrors the case of inductive inferences not concerning observers: while reasonable and unreasonable inferences are easy enough to find, the prospects for precisely delineating the reasonable inferences in that class seem bleak. This analogy is no accident: number and proportion boosting are self-locating inferences that are special cases of inductive inference from observations of something to a conclusion about a broader reference class that includes that thing.  Noticing this should help allay any worry that, in the absence of a worked out account, the subject-involving character of number and proportion boosting renders their epistemology suspect.  

[136] See Chalmers (2010b) and Morevac (1976; 1999); cf. Shulman & Bostrom (2012).

[137] See Cotra (2020) and Shulman & Bostrom (2012: §1.2).

[138] For discussion of interactions between superintelligence and the simulation argument, see Prinz (2012).

[139] See Shulman & Bostrom (2012: §§3-4).

[140] Cf. Carter (1983).

[141] See Shulman & Bostrom (2012: §5) and Snyder-Beattie et al. (2021).

[142] The argument traces back at least to Brandon Carter (who did not publish on it). Versions of it were later advanced by Gott (1993) and Leslie (1989: 214; 1996). I work with a formulation that is closer to Leslie’s, as his formulation is widely regarded as yielding a more plausible (though still controversial) argument than Gott’s.  For background and critical discussion of much of the literature on the doomsday argument, see Bostrom (2002b: Chs. 6-7).  For notable variations of the argument that have received relatively little attention, see Grace (2010), Mogensen (2019a), and Turchin (2018).

[143] Cf. Bostrom (2002b: 92-3).

[144] See also Lewis (2013) and Richmond (2017).

[145] For discussion and references, see Bostrom (2002b: 122-6; Ch. 7, fn11).

[146] For discussion, see Ćirković (2018: §4.5).

[147] See Dainton (2020: 220).

[148] This sort of strategy was (famously) promoted by Stephen Hawking—see, e.g., BBC (2010).

[149] The import of this evidence is, however, blocked by Early Filter solutions on which the dearth of intelligent life is explained by a filter that renders life rare. Compare Armstrong (2014): “The Great Filter is early, or AI is hard”.  If correct, this points to a worrying corollary of progress in AI: insofar as it gives us reason to think engineering human-level AI is easier than we thought, it also gives us reason to doubt Early Filter solutions to Fermi’s paradox and hence to be more confident that the Filter lies in our future.  But see Miller (2019) for argument that evidence for both an Early Filter and AI-based existential risks is less concerning than evidence for either alone would be.  See also Hanson et al. (2021) for a solution to Fermi’s paradox that lends to an explanation of how an Early Filter could cohere with human-level AI being relatively easy to engineer.

[150] See Ward & Brownlee (2000).

[151] See Monton (2009: 99) for references.

[152] See Zackrisson et al. (2016).

[153] Cf. Carter (1983). Note, however, that it remains a live possibility that the universe has an infinite number of Earth-like planets and that this would in effect render it certain that there are other such planets that give rise to intelligent life no matter how unlikely it is that a given planet would—see Monton (2009: 102-4).  The existence of an infinite number of observers also raises technical difficulties for the application of indifference principles—see Bostrom (2002b), Dorr & Arntzenius (2017), and White (2018).

[154] For background on the physical theories, see Carroll (2020).  For discussion, see Kotzen (2020).

[155] For discussion of similarities between the simulation argument and the Boltzmann brain problem, see Crawford (2013).

[156] Of course, a world in which almost all observers like us are Boltzmann brains would itself arguably involve catastrophe in the form of premature death on a cosmic scale.  However, if there is a real risk of such a catastrophe, the catastrophe is presumably already underway and there is nothing we can do to mitigate it, with the possible exception of non-causal interventions like those discussed in §13. In any event, I will set this sort of catastrophe aside.

[157] See Chen (forthcoming).

[158] See Saad (forthcomingc).

[159] See Carroll (2020).

[160] To my knowledge, nothing has been published on reducing catastrophic risks by using simulations as non-causal interventions. However, after submitting a research proposal to the Center on Long-Term Risk on the topic, I was informed that the idea had already been considered there.  I plan to expand the ideas in this section into a stand-alone paper.

[161] Cf. Bostrom (2003: 253).

[162] See Leslie (1991).

[163] In a recent survey of professional philosophers, participants were asked about their view of Newcomb’s problem, a case that is standardly taken to elicit different recommendations from causal decision theory and evidential decision theory.  31.2% favored the response standardly associated with evidential decision theory while 39.0% favored the response standardly associated with causal decision theory (Bourget & Chalmers, 2021).  It is thus natural to interpret these results as indicating that causal decision theory is more widely favored.  However, there is also debate about whether Newcomb’s problem bears on causal decision theory and evidential decision theory in the way that is standardly supposed—see Knab (2019: Ch. 3) for references and reasons to think not.

[164] See Easwaran (2021) for an illuminating taxonomy and references.

[165] See Easwaran (2021: §3.2), Kagan (2000), and Nozick (1993).

[166] See MacAskill (2016).

[167] See MacAskill et al. (2021).

[168] See Lockhart (2000), Bykvyst (2017), and MacAskill et al. (2020).

[169] For a decision theory that treats risk aversion as a basic parameter, see Buchak (2013).

[170] Respectively, see Greaves & Cotton-Barratt (2019) and Newberry & Ord (2021).

[171] See Parfit (1984: §95).

[172] See Saad (forthcomingb).

[173] See Lewis (2004).

[174] Cf. White (2005) and Wright (2004).

[175] Cf. Dainton (2012: 68).

Subscribe to our newsletter to receive updates on our research and activities. We average one to two emails per year.