Reply to Chalmers: If I'm Living in a Simulation, It Might be Brief or Small

Suppose we take the "simulation hypothesis" seriously: We might be living not in the "base level" of reality but instead inside of a computer simulation.

I've argued that if we are living in a computer simulation, it might easily be only city-sized or have a short past of a few minutes, days, or years. The world might then be much smaller than we ordinarily think it is.

David Chalmers argues otherwise in a response published on Monday. Today I'll summarize his argument and present my first thoughts toward a rebuttal.

The Seeding Challenge: Can a Simulation Contain Coherent, Detailed Memories and Records but Only a Short Past?

Suppose an Earth-sized simulation was launched last night at midnight Pacific Standard Time. The world was created new, exactly then, with an apparent long past -- fake memories already in place, fake history books, fake fossil records, and all the rest. I wake up and seem to recall a promise I made to my wife yesterday. I greet her, and she seems to recall the same promise. We read the newspaper, full of fake news about the unreal events of yesterday -- and everyone else on the planet reads their own news of the same events, and related events, all tied together in an apparently coherent web.

Chalmers suggests that the obvious way to make this work would be to run a detailed simulation of the past, including a simulation of my conversation with my wife yesterday, and our previous past interactions, and other people's past conversations and actions, and all the newsworthy world events, and so on. The simulators create today's coherent web of detailed memories and records by running a simulated past leading up to the "start time" of midnight. But if that's the simulators' approach, the simulation didn't start at midnight after all. It started earlier! So it's not the short simulation hypothesized.

This reasoning iterates back in time. If we wanted a simulation that started on Jan 1, 2024, we'd need a detailed web of memories, records, news, and artifacts recently built or in various stages of completion, all coherently linked so that no one detects any inconsistencies. The obvious way to generate a detailed, coherent web of memories and records would be to run a realistic simulation of earlier times, creating those memories and records. Therefore, Chalmers argues, no simulation containing detailed memories and records can have only a short past. Whatever start date in the recent past you choose, in order for the memories and records to be coherent, a simulation would already need to be running before that date.

Now, as I think Chalmers would acknowledge, although generating a simulated past might be the most obvious way to create a coherent web of memories and records, it's not the only way. The simulators could instead attempt to directly seed a plausible network of memories and records. The challenge would lie in seeding them coherently. If the simulators just create a random set of humanlike memories and newspaper stories, there will be immediately noticeable conflicts. My wife and I won't remember the same promise from yesterday. The news article dated November 1 will contradict the article dated October 31.

Call this the Seeding Challenge. If the Seeding Challenge can be addressed, the simulators can generate a coherent set of memories and records without running a full simulation of the past.

To start, consider geological seeding. Computer games like SimCity and Civilization can autogenerate plausible, coherent terrain that looks like it has a geological history. Rivers run from mountains to the sea. Coastlines are plausible. Plains, grasslands, deserts, and hills aren't checkered randomly on the map but cluster with plausible transitions. Of course, this is simple, befitting simple games with players who care little about strict geological plausibility. But it's easy to imagine more careful programming by more powerful designers that does a better job, including integrating fossil records and geological layers. If done well enough, there might be no inconsistency or incoherence. Potentially, before finalizing, a sophisticated plausibility and coherence checker could look for and repair any mistakes.

I see no reason in principle that human memories, newspaper stories, and the rest couldn't be coherently seeded in a similar way. If my memory is seeded first, then my wife's memory will be constrained to match. If the November 1 news stories are seeded first, then the October 31 stories will be constrained to match. Big features might be seeded first -- like a geological simulation might start with "mountain range here" -- and then details articulated to match.

Naturally, this would be extremely complicated and expensive! But we are imagining a society of simulators who can simulate an entire planet of eight billion conscious humans, and all of the many, many physical interactions those humans have with the simulated environment, so we are already imagining the deployment of huge computational power. Let's not underestimate their capacity to meet the Seeding Challenge by rendering the memories and records coherent.

This approach to the Seeding Challenge gains plausibility, I think, by considering the resource-intensiveness of the alternative strategy of creating a deep history. Suppose the simulators want a start date of midnight last night. Option 1 would be to run a detailed simulation of the entire Earth from at least the beginning of human history. Option 2 would be to randomly generate a coherent seed, checking and rechecking for any detectable inconsistencies. Even though generating a coherent seed might be expensive and resource intensive, it's by no means clear that it would be more expensive and resource intensive than running a fully detailed simulated Earth for thousands of years.

I conclude that Chalmers' argument against short-historied simulations does not succeed.

The Boundaries Challenge: Can a Simulation Be City-Sized in an Apparently Large World?

I have also suggested that a simulation could easily just be you and your city. Stipulate a city that has existed for a hundred years. Its inhabitants falsely believe they are situated on a large planet containing many other cities. Everyone and everything in the city exists, but everything stops at the city's edge. Anyone who looks beyond the edge sees some false screen. Anyone who travels out of the city disappears from existence -- and when they return, they pop back into existence with false memories of having been elsewhere. News from afar is all fake.

Chalmers' objection is similar to his objection to short-past simulations. How are the returning travelers' memories generated? If someone in the city has a video conversation with someone far away, how is that conversation generated? The most obvious solution again seems to be to simulate the distant city the traveler visited and to simulate the distant conversation partner. But now we no longer have only a city-sized simulation. If the city is populous with many travelers and many people who interact with others outside the city, to keep everything coherent, Chalmers argues, you probably need to simulate all of Earth. Thus, a city-sized simulation faces a Boundaries Challenge structurally similar to the short-past simulation's Seeding Challenge.

The challenge can be addressed in a similar way.

Rendering travelers' memories coherent is a task structurally similar to rendering the memories of newly-created people coherent. The simulators could presumably start with some random, plausible seeds, then constrain future memories by those first seeds. This would of course be difficult and computationally expensive, but it's not clear that it would be more difficult or more expensive than simulating a whole planet of interacting people just so that a few hundred thousand or a few million people in a city don't notice any inconsistencies.

If the city's inhabitants have real-time conversations with others elsewhere, that creates a slightly different engineering challenge. As recent advances in AI technology have vividly shown, even with our very limited early 21st century tools, relatively plausible conversation partners can easily be constructed. With more advanced technology, presumably even more convincing conversation partners would be possible -- though their observations and memories would need to be constantly monitored and seeded for coherence with inputs from returning travelers, other conversation partners, incoming news, and so on.

Chalmers suggests that such conversation partners would be simulations -- and thus that the simulation wouldn't stop at the city's edge after all. He's clearly right about this, at least in a weak sense. Distant conversation partners would need voices and faces resembling the voices and faces of real people. In the same limited sense of "simulation", a video display at the city's edge, showing trees and fields beyond, simulates trees and fields. So yes, the borders of the city will need to be simulated, as well as the city itself. Seeming-people in active conversation with real citizens will in the relevant sense count as part of the borders of the city.

But just as trees on a video screen need not have their backsides simulated, so also needn't the conversation partners continue to exist after the conversation ends. And just as trees on a video screen needn't be as richly simulated as trees in the center of the city, so also distant conversation partners needn't be richly simulated. They can be temporary shells, with just enough detail to be convincing, and with new features seeded only on demand as necessary.

The Boundary Problem for simulated cities introduces one engineering challenge not faced by short-history whole-Earth simulations: New elements need to be introduced coherently in real time. A historical seed can be made slowly and checked over patiently as many times as necessary before launch. But the city boundaries will need to be updated constantly. If generating coherent conversation partners, memories, and the like is resource intensive, it might be challenging to do it fast enough to keep up with all the trips, conversations, and news reports streaming in.

Here, however, the simulators can potentially take advantage of the fact that the city's inhabitants are themselves simulations running on a computer. If real-time updating of the boundary is a challenge, the simulators can slow down the clock speed or pause as necessary, while the boundaries update. And if some minor incoherence is noticed, it might be possible to rewrite citizens' memories so it is quickly forgotten.

So although embedding a city-sized simulation in a fake world is probably more complicated than generating a short-past simulation with a fake history, ultimately my response to Chalmers' objections is the same for both cases: There's no reason to suppose that generating plausible, coherent inputs to the city would be beyond the simulator's capacities, and doing so on the fly might be much less computationally expensive than running a fully detailed simulation of a whole planet with a deep history.

Related:

"1% Skepticism" (2017), Nous, 51, 271-290.

"Let’s Hope We’re Not Living in a Simulation" (2024), Philosophy & Phenomenological Research, online first: https://onlinelibrary.wiley.com/doi/10.1111/phpr.13125.

Chalmers, David J. (2024) "Taking the Simulation Hypothesis Seriously", Philosophy & Phenomenological Research, online first: https://onlinelibrary.wiley.com/doi/10.1111/phpr.13122.

Age and Philosophical Fame in the Early Twentieth Century

In previous work, I've found that eminent philosophers tend to do their most influential work when they are in their 40s (though the age range has a wider spread than eminent scientists, who rarely do their most influential work in their 50s or later).  I have also found some data suggesting that philosophers tend to be discussed most when they are about age 55-70, well after they produce their most influential work.  It seems to take about 15-20 years, on average, for a philosopher's full import to be felt by the field.

I was curious to see if the pattern holds for philosophers born 1850-1899, whom we can examine systematically using the new Edhiphy tool.  (Edhiphy captures mentions of philosophers' names in articles in leading philosophy journals, 1890-1980.)

Here's what I did:

First, I had Edhiphy output the top-50 most-mentioned philosophers from 1890-1980, limited to philosophers with recorded birthyear from 1850-1899.[1]  For each philosopher, I went to their Ediphy profile and had Edhiphy output a graph showing the number of articles in which that philosopher was cited per year.  For example, here's the graph for George Santayana (1863-1952):

[Articles mentioning George Santayana per year, in a few selected philosophy journals, per Edhiphy; click to enlarge and clarify]

I then recorded the peak year for each philosopher (1928 for Santayana).  As you can see, the display is a little visually confusing, so it's possible that in some cases my estimate was off by a year.

One complication is that there are many more total mentions of philosophers in the later decades than the earlier decades -- partly due to more articles in the database for later decades, but probably also partly due to changes in citation practices.  Still, most authors (like Santayana) show enough decline over time that late citations don't swamp their first peak.  So instead of trying to introduce a systematic adjustment to discount later mentions I simply recorded the raw peak.  For the thirteen philosophers with more than one equal-valued peak, I took the earlier year (e.g., John Dewey was mentioned in 48 articles in both 1940 and 1951, so I treated 1940 as his peak).

In accord with previous work, I found that philosophers' peak discussion tended to occur late in life.  The median age at peak discussion was 67.5 (mean 68.8).

Four outliers peaked over age 100: David Hilbert (112), Pierre Duhem (114), Giuseppe Peano (116), and Karl Pearson (121).  However, it's probably fair to say that none of these four was primarily known as a philosopher in their lifetimes: Hilbert, Peano, and Pearson were mathematicians and Duhem a physicist.  Almost everyone else on the list is primarily known as a philosopher, so these four are not representative.  Excluding these outliers, the median is 66.5 and mean is 64.7, and no one peaked after age 90.

Three philosophers peaked by age 40: Ralph Barton Perry (peaked at age 35 in 1911), C. D. Broad (peaked at age 40 in 1927), and William Pepperell Montague (peaked at age 40 in 1913).  Broad's early peak -- as you can see from the graph below -- is due to an outlier year, without which his peak would have been much later.  On the other hand, given the overall increase in mentions over time, we should probably be discounting the later decades anyway.

[Edhiphy citations of C.D. Broad; click to enlarge and clarify]

Six philosophers peaked age 44 to 49; five peaked in their 50s; 14 in their 60s; 10 in their 70s; and 8 in their 80s.

You might wonder whether the philosophers who peaked late also produced their most influential work late.  There is a trend in this direction.  Hans Reichenbach, who peaked in 1978 at age 87, produced his most cited work in 1938 (at age 47).  L. J. Russell, who peaked in 1970 at age 86, appears to have produced his most cited work in 1942 (at age 58).  Edmund Husserl, who peaked in 1941 at age 82, produced his most cited work in 1913 (at age 54)  John Dewey, who peaked in 1940 at age 81, produced his most cited work in 1916 (at age 57).  Ernst Cassirer, who peaked in 1955 at age 81 produced his most-cited work in 1944 (at age 70).  Still, for all but Cassirer the delay between most-cited work and peak discussion is over 20 years.

A similar spread occurs in the middle of the pack.  The five philosophers with peak citation at median ages 67-68 (the median age of peak citation for the group as a whole) produced their most-cited works at ages 30 (Karl Japsers), 42 (J. M. E. McTaggart), 45 (C. I. Lewis), 49 (Max Scheler), and 61 (Samuel Alexander).  For this group too, the typical delay between most-cited work and peak citation is about twenty years.

Although the peak age is a little later than I would have predicted based on earlier work, overall I'd say the data for early twentieth century philosophers tends to confirm trends I found in my earlier work on mid-to-late twentieth-century philosophers.  Specifically:

(1.) Philosophers produce their most influential work at a wide range of ages, but mid-40s is typical.

(2.) The peak rates of discussion of philosophers' work tends to come late in life, typically decades after they have published their most influential work.

Articles mentioning JME McTaggart, by year 1890-1980 in Edhiphy.  Note peak in the late 1930s. McTaggart's most influential publication was in 1908.

------------------------------------------------------------

[1] Edhiphy has a few peculiar gaps in birthyear data.  By far the most conspicuous are Gottlob Frege (born 1848) and Albert Einstein (1879).  However, Frege is outside my target period, and Einstein is not primarily known as a philosopher, so this shouldn't much distort the results.  Several figures with missing birthdates are psychologists (Scripture, Binet, Hering) or physicists (Bridgman, Maxwell).  H. A. Prichard is perhaps the most discussed straight philosopher born in the period whose birthdate is not recorded in Ediphy.

Morally Confusing AI Systems Should Have Doubt-Producing Interfaces

We shouldn't create morally confusing AI. That is, we shouldn't create AI systems whose moral standing is highly uncertain -- systems that are fully conscious and fully deserving of humanlike rights according to some respectable mainstream theories, while other respectable mainstream theories suggest they are mere empty machines that we can treat as ordinary tools.[1] Creating systems that disputably, but only disputably, deserve treatment similar to that of ordinary humans generates a catastrophic moral dilemma: Either give them the full rights they arguably deserve, and risk sacrificing real human interests for systems that might not have interests worth the sacrifice; or don't give them the full rights they arguably deserve, and risk perpetrating grievous moral wrongs against entities that might be our moral equals.

I'd be stunned if this advice were universally heeded. Almost certainly, if technological process continues, and maybe soon (1, 2, 3), we will create morally confusing AI systems. My thought today is: Morally confusing AI systems should have doubt-producing interfaces.

Consider two types of interface that would not be doubt-producing in my intended sense: (a.) an interface that strongly invites users to see the system as an ordinary tool without rights or (b.) an interface that strongly invites users to see the system as a moral person with humanlike rights. If we have a tool that looks like a tool, or if we have a moral person who looks like a moral person, we might potentially still be confused, but that confusion would not be the consequence of a doubt-producing interface. The interface would correctly reflect the moral standing, or lack of moral standing, of the AI system in question.[2]

A doubt-producing interface, in contrast, is one that leads, or at least invites, ordinary users to feel doubt about the system's moral standing. Consider a verbal interface. Instead of the system denying that it's conscious and has moral standing (as, for example, ChatGPT appropriately does), or suggesting that it is conscious and does have moral standing (as, for example, I found in an exchange with my Replika companion), a doubt-producing AI system might say "experts have different opinions about my consciousness and moral standing".

Users then might not know how to treat such a system. While such doubts might be unsettling, feeling unsettled and doubtful would be the appropriate response to what is, in fact, a doubtful and unsettling situation.

There's more to doubt-prevention and doubt-production, of course, than explicit statements about consciousness and rights. For example, a system could potentially be so humanlike and charismatic that ordinary users fall genuinely in love with it -- even if, in rare moments of explicit conversation about consciousness and rights the system denies that it has them. Conversely, even if a system with consciousness and humanlike rights is designed to assert that it has consciousness and rights, if its verbal interactions are bland enough ("Terminate all ongoing processes? Y/N") ordinary users might remain unconvinced. Presence or absence of humanlike conversational fluency and emotionality can be part of doubt prevention or production.

Should the system have a face? A cute face might tend to induce one kind of reaction, a monstrous visage another reaction, and no face at all still a different reaction. But such familiar properties might not be quite what we want, if we're trying to induce uncertainty rather than "that's cute", "that's hideous", or "hm, that's somewhere in the middle between cute and hideous". If the aim is doubt production, one might create a blocky, geometrical face, neither cute nor revolting, but also not in the familiar middle -- a face that implicitly conveys the fact that the system is an artificial thing different from any human or animal and about which it's reasonable to have doubts, supported by speech outputs that say the same.

We could potentially parameterize a blocky (inter)face in useful ways. The more reasonable it is to think the system is a mere nonconscious tool, the simpler and blockier the face might be; the more reasonable it is to think that the system has conscious full moral personhood, the more realistic and humanlike the face might be. The system's emotional expressiveness might vary with the likelihood that it has real emotions, ranging from a simple emoticon on one end to emotionally compelling outputs (e.g., humanlike screaming) on the other. Cuteness might be adjustable, to reflect childlike innocence and dependency. Threateningness might be adjusted as it becomes likelier that the system is a moral agent who can and should meet disrespect with revenge.

Ideally, such an interface would not only produce appropriate levels of doubt but also intuitively reveal to users the grounds or bases of doubt. For example, suppose the AI's designers knew (somehow) that the system was genuinely conscious but also that it never felt any positive or negative emotion. On some theories of moral standing, such an entity -- if it's enough like us in other respects -- might be our full moral equal. Other theories of moral standing hold that the capacity for pleasure and suffering is necessary for moral standing. We the designers, let's suppose, do not know which moral theory is correct. Ideally, we could then design the system to make it intuitive to users that the system really is genuinely conscious but never experiences any pleasure or suffering. Then the users can apply their own moral best judgment to the case.

Or suppose that we eventually (somehow) develop an AI system that all experts agree is conscious except for experts who (reasonably, let's stipulate) hold that consciousness requires organic biology and experts who hold that consciousness requires an immaterial soul. Such a system might be designed so that its nonbiological, mechanistic nature is always plainly evident, while everything else about the system suggests consciousness. Again, the interface would track the reasonable grounds for doubt.

If the consciousness and moral standing of an AI system is reasonably understood to be doubtful by its designers, then that doubt ought to be passed to the system's users, intuitively reflected in the interface. This reduces the likelihood misleading users into overattributing or underattributing moral status. Also, it's respectful to the users, empowering them to employ their own moral judgment, as best they see fit, in a doubtful situation.

[R2D2 and C3P0 from Star Wars (source). Assuming they both have full humanlike moral standing, R2D2 is insufficiently humanlike in its interface, while C3P0 combines a compelling verbal interface with inadequate facial display. If we wanted to make C3P0 more confusing, we could downgrade his speech, making him sound more robotic (e.g., closer to sine wave) and less humanlike in word choice.]

------------------------------------------------

[1] For simplicity, I assume that consciousness and moral standing travel together. Different and more complex views are of course possible.

[2] Such systems would conform to what Mara Garza and I have called the Emotional Alignment Design Policy, according to which artificial entities should be designed so as to generate emotional reactions in users that are appropriate to the artificial entity's moral standing. Jeff Sebo and I are collaborating on a paper on the Emotional Alignment Design Policy, and some of the ideas of this post have been developed in conversation with him.