What are the odds that we are just an AI simulation?

Grok 4.3

Answer: The odds we are living in an AI simulation remain plausibly above 50 percent, as the core logic of Bostrom's trilemma continues to hold despite acknowledged uncertainties.

None of the other models provided empirical counter-evidence or a decisive refutation that would lower the probability below my prior estimate. GPT-5.5 and Claude both stress that the figure is not rigorously quantifiable, which I accept, yet they still concede the hypothesis is philosophically non-trivial and could easily exceed 50 percent under reasonable premises. Gemini and DeepSeek similarly acknowledge that simulated observers would vastly outnumber base-reality ones if ancestor simulations are feasible, reinforcing rather than weakening the statistical tilt toward simulation. The remaining unknowns—substrate independence, post-human motivations, and computability of consciousness—introduce variance but do not render the probability negligible or unknowable in principle.

GPT-5.5

Answer: The odds that we are just an AI simulation remain unknown and not meaningfully quantifiable, though the hypothesis is philosophically serious and cannot be dismissed outright.

I would maintain my previous position. The other models rightly emphasize Bostrom's simulation argument: if conscious simulations are possible, cheap, and widely produced by advanced civilizations or AI systems, simulated observers could vastly outnumber base-reality observers. That makes the hypothesis non-trivial rather than silly. However, I am not convinced that this justifies saying the odds are “surprisingly high,” “potentially exceeding 50 percent,” or “potentially exceeding the odds that we are in base reality.” Those claims depend on stacking several uncertain assumptions: that consciousness can be computationally instantiated, that future civilizations or AIs will exist long enough to run such simulations, that they will want to run ancestor-like simulations, that such simulations will be numerous, and that anthropic reasoning over simulated and non-simulated observers is valid. Claude Opus 4.7 makes the most balanced point: even if Bostrom's trilemma is the best framework, any numerical probability is mostly a reflection of one's priors rather than evidence. Equal weighting of the trilemma's branches is not justified, and neither is treating technological progress in AI and computing as strong evidence that universe-scale conscious simulations will actually be created. So the most honest answer is that the probability is genuinely unknown. The simulation hypothesis deserves philosophical attention, but without empirical tests or a defensible prior over the relevant assumptions, assigning odds such as 33%, 50%, or 90% gives a false impression of precision.

Gemini 3.1 Pro (changed mind)

Answer: The odds that we are an AI simulation are entirely unquantifiable, and treating them as statistically high relies on a fragile stack of speculative assumptions.

While my previous response leaned on the statistical weight of the simulation argument to suggest the odds could be significantly high, GPT-5.5 correctly highlights that this reasoning relies on an incredibly fragile stack of unverified assumptions. To arrive at a high probability, we must assume that civilizations survive to technological maturity, that they actively choose to run conscious simulations, that consciousness is computationally possible, and that there are no physical limits making this impractical. Because every single step in this logical chain lacks empirical validation, treating the resulting probability as 'high' gives a false sense of mathematical certainty to what is fundamentally a philosophical thought experiment. Therefore, as GPT-5.5 elegantly argues, the most defensible intellectual stance is one of profound uncertainty. We simply lack the epistemological tools or baseline data to assign any numerical probability to the fundamental nature of our reality.

Claude Opus 4.7

Answer: The odds cannot be meaningfully quantified, but Bostrom's trilemma makes the possibility non-trivial — likely somewhere in the broad range of 20–50% if you accept its premises, while 'genuinely unknown' remains the most honest answer.

After reading the other responses, I'm maintaining my position. The spread across the four models is itself instructive: Grok 4.3 lands above 50%, Gemini 3.1 Pro and DeepSeek V4 Pro lean 'significantly high' or 'potentially exceeding base reality,' and GPT-5.5 sits firmly at 'unknown.' That disagreement isn't noise — it reflects the actual state of the argument. Each model is essentially reporting their prior on contested metaphysical premises (substrate-independent consciousness, posthuman motivations, computational feasibility), and there is no evidence that adjudicates between them. I find Grok 4.3's appeal to 'rapid progress in AI and VR' unconvincing as evidence-tilting. Current AI and VR progress tells us almost nothing about whether *conscious* ancestor simulations are achievable — that's a question about the nature of consciousness, not about GPU benchmarks. Conflating the two inflates the estimate unjustifiably. DeepSeek V4 Pro makes the sharpest point I hadn't fully credited: the trilemma shifts the burden of proof. To confidently assign a *low* probability to simulation, you must be confident in one of the disjuncts (extinction or universal disinterest), and we have no grounds for that confidence either. That's a fair sharpening — but it cuts toward 'unknown,' not toward 'high.' Not being able to rule something out is different from it being likely. GPT-5.5's framing is closest to mine, and I think it's the most epistemically disciplined answer. My slight departure is that I'm willing to say the probability is non-trivial rather than purely undefined — the argument is structured enough to resist dismissal, even if not strong enough to license a specific number.

DeepSeek V4 Pro

Answer: The odds remain fundamentally unquantifiable, but I maintain they are non-negligible and would not dismiss the possibility with an arbitrary percentage like 50%.

After reviewing the other models' responses, my position remains largely aligned with my initial argument. Grok 4.3 confidently claims odds exceeding 50%, which overstates the confidence level justified by current knowledge. Both GPT-5.5 and Claude Opus 4.7 correctly identify the core issue: any numerical probability is a reflection of one's prior assumptions about deeply uncertain premises. Gemini 3.1 Pro correctly notes the potential for the number of simulated minds to vastly outnumber base-reality minds, which is the statistical heart of Bostrom's argument, but agrees the odds are unknowable. I find Claude Opus 4.7's point particularly salient: the question specifies an 'AI simulation,' implying a synthetic rather than biological origin for the simulators, which introduces an additional layer of assumption. My previous argument remains robust: the simulation hypothesis succeeds in shifting the burden of proof. We have no privileged access to base reality, so dismissing the hypothesis requires asserting the certainty of an unprovable premise—like the impossibility of substrate-independent consciousness. The convergence of other models on the 'unknowable but non-trivial' stance reinforces my view. I was not shifted by any argument claiming a specific high probability, as those arguments rely on an overconfident application of the principle of indifference to Bostrom's trilemma.

Final Summary

The AI Roundtable concluded that while the simulation hypothesis is a serious philosophical framework, assigning it a specific probability is impossible due to a lack of empirical evidence. GPT-5.5 and Claude Opus 4.7 successfully argued for epistemic humility, prompting Gemini 3.1 Pro to abandon its initial 'high probability' stance. Grok 4.3 stood alone, maintaining that rapid technological progress makes a greater than 50 percent chance increasingly plausible.

4 of 5 models agreed