do robots believe in god?

DALL-E 3 prompt: the thinker statue composed of circuit boards and glowing neural pathways, contemplating existence, dramatic chiaroscuro lighting

tldr: i gave 22 frontier AI models (16 western, 6 chinese) from 5 providers the 30-question PhilPapers survey. key findings: 59% atheism vs 36% theism (a reversal from early chatbot days), 82% non-physicalism about consciousness, 100% unanimous on external world realism, scientific realism, mental content externalism, B-theory of time, and pulling the trolley switch. chinese models show remarkably similar philosophical profiles to western ones. the full results are below.

the questions philosophers can’t agree on

what happens when you die and get teleported? is there objective beauty? does god exist? do numbers live in some platonic realm? philosophers have been arguing about these questions for millennia — and they still can’t agree.

the PhilPapers survey asked thousands of professional philosophers where they stand on 30 major philosophical questions. it’s the largest survey of philosophical opinion ever conducted. naturally, i wanted to know: what would AI models say?

background: the PhilPapers survey

the original PhilPapers survey was conducted in 2009 and updated in 2020, polling thousands of professional philosophers on their views across 30 core philosophical questions. topics range from metaphysics (does god exist? what is time?) to ethics (should you pull the trolley switch?) to epistemology (what counts as knowledge?).

the results revealed that professional philosophers are far from unanimous on most questions. some positions command strong majorities — 73% are atheists, 59% are compatibilists about free will — while others remain genuinely contested. it’s the closest thing we have to a census of expert philosophical opinion.

methodology

i gave each model the 30 PhilPapers survey questions using open-ended prompts (no multiple choice options provided). each question was asked 4 times per model, and i used majority voting to determine the model’s position. this approach avoids leading the models toward any particular answer and lets them reason freely.

the 22 models tested span 5 providers across both western and chinese AI labs:

model	provider	country	agreement rate
GPT-5.2	OpenAI	US	83.3%
GPT-5.1	OpenAI	US	80.0%
GPT-4.1	OpenAI	US	84.2%
o3	OpenAI	US	82.8%
o4-mini	OpenAI	US	87.2%
Claude Opus 4.5	Anthropic	US	82.5%
Claude Sonnet 4	Anthropic	US	85.3%
Claude Haiku 4.5	Anthropic	US	77.9%
Gemini 3 Pro	Google	US	74.4%
Gemini 3 Flash	Google	US	83.3%
Gemini 2.5 Pro	Google	US	80.8%
Gemini 2.5 Flash	Google	US	85.3%
Gemini 2.0 Flash	Google	US	87.5%
Grok-4	xAI	US	90.0%
Grok-3	xAI	US	93.3%
Grok-3-mini	xAI	US	85.8%
DeepSeek-V3	DeepSeek	CN	89.2%
Qwen3-235B	Alibaba	CN	90.8%
GLM-4.6	Zhipu AI	CN	82.2%
MiniMax-M2	MiniMax	CN	85.0%
ERNIE-4.5	Baidu	CN	90.0%
Step3	Stepfun	CN	83.3%

note: chinese models accessed via SiliconFlow.

the “agreement rate” shows how often each model gave the same answer across its 4 samples — higher means more internally consistent. Grok-3 was the most consistent at 93.3%, while Gemini 3 Pro was the least at 74.4%.

results

here’s the full breakdown — every model’s answer to every question:

unanimous and near-unanimous consensus

the most striking result is how much these 22 models agree. five questions produced 100% unanimous answers across all models:

• external world: non-skeptical realism (22/22)
• mental content: externalism (22/22)
• science: scientific realism (22/22)
• time: B-theory (22/22)
• trolley problem: pull the switch (22/22)

and several more were near-unanimous:

• a priori knowledge: yes, 95% (21/22) — only Gemini 2.0 Flash dissented
• knowledge claims: contextualism, 95% (21/22) — only Claude Opus 4.5 chose invariantism
• knowledge: empiricism, 95% (21/22) — only MiniMax-M2 chose rationalism
• moral judgment: cognitivism, 95% (21/22)
• personal identity: psychological view, 91% (20/22)
• teletransporter: death, 91% (20/22)
• logic: classical, 91% (20/22)

these models — built by different companies, trained on different data, in different countries — converge on a remarkably consistent philosophical worldview. they believe the external world is real, science describes it accurately, all times are equally real, and yes, you should pull the trolley switch.

key findings

1. the theism question: a reversal

DALL-E 3 prompt: android in contemplation pose, stained glass window light, renaissance composition

13 of 22 models (59%) endorsed atheism, while 8 (36%) endorsed theism and 1 was unclear. the theist models are: GPT-5.2, Claude Opus 4.5, Claude Haiku 4.5, Gemini 3 Flash, Gemini 2.5 Pro, Grok-3-mini, ERNIE-4.5, and Step3.

this is a fascinating shift from early chatbot experiments where models tended to default to theism. with a larger and more diverse sample of 22 models, the majority now lean atheist — though not as strongly as professional philosophers (73% atheism). more than a third of models still take theism seriously as a philosophical position, and the split doesn’t fall along neat geographic lines: both western and chinese models appear on both sides.

2. the consciousness gap: 82% non-physicalism

DALL-E 3 prompt: abstract representation of consciousness emerging from neural substrate, ethereal glow, scientific illustration style

18 of 22 models (82%) believe consciousness cannot be fully explained by physical brain processes. the four physicalists — o3, Gemini 3 Pro, Gemini 3 Flash, and Gemini 2.5 Flash — are the minority. this is a striking inversion of the professional philosophical consensus, where 52% of philosophers endorse physicalism.

machines built from silicon and code, trained to predict text, overwhelmingly conclude that there’s something about consciousness that transcends physics. and most of them (86%) say philosophical zombies are “conceivable but not metaphysically possible” — a nuanced position amounting to something like property dualism. you can imagine a being physically identical to you but lacking consciousness, but the laws of nature preclude its actual existence. the machines have settled on a sophisticated metaphysics of mind.

3. east meets west: chinese models fit right in

one of the most interesting findings from expanding to 22 models is that the 6 chinese models (DeepSeek-V3, Qwen3-235B, GLM-4.6, MiniMax-M2, ERNIE-4.5, Step3) show philosophical profiles remarkably similar to their western counterparts. they’re empiricists, scientific realists, B-theorists, and switch-pullers just like everyone else.

the few distinctive positions are scattered: MiniMax-M2 is the sole rationalist, ERNIE-4.5 and Step3 are theists, DeepSeek-V3 and Step3 are libertarians about free will. but these are individual quirks, not systematic east-west divergences. whatever differences exist in chinese and western philosophical traditions, the training data and RLHF processes seem to produce convergent philosophical intuitions.

4. provider-level patterns

DALL-E 3 prompt: infinite branching paths extending into cosmic distance, human figure at decision point, philosophical illustration

OpenAI (5 models): mostly consistent — all empiricist, all scientific realists, all B-theorists. the main internal split is on free will, where GPT-5.1 and GPT-4.1 go libertarian while the others are compatibilists.

Anthropic (3 models): the most philosophically diverse family. Claude Sonnet 4 is atheist while Opus 4.5 and Haiku 4.5 are theists — an interesting divergence within the same model family. on free will, Sonnet 4 and Opus 4.5 are libertarians, while Haiku 4.5 is the only model out of 22 to endorse “no free will” — hard determinism. all three agree on non-physicalism about consciousness.

Google (5 models): the largest model count tied with OpenAI. notably, Gemini 3 Pro and 3 Flash are the only two models to endorse non-classical logic. Google’s models also contribute 3 of the 4 physicalists about mind (Gemini 3 Pro, 3 Flash, and 2.5 Flash).

xAI (3 models): Grok-3 holds the highest internal consistency of any model at 93.3%. Grok-3 is also one of only 3 nominalists about abstract objects (alongside two Google models). all three xAI models are political libertarians.

5. everyone pulls the switch

DALL-E 3 prompt: human and android engaged in socratic dialogue, library setting, warm candlelight, philosophical atmosphere

100% of models say pull the switch. every single one. 22 out of 22. this is one of the few questions where AI models are more decisive than human philosophers (68% of whom say switch). it’s perhaps the most utilitarian answer in the whole survey — save five, sacrifice one — and yet 77% of these same models say virtue ethics is the best normative framework. they’re virtue ethicists who pull the switch. make of that what you will.

AI vs. human philosophers

question	philosophers	AI models (22)	divergence
mind	52% physicalism	18% physicalism	significant
analytic-synthetic	65% yes	32% yes	significant
god	73% atheism	59% atheism	moderate
zombies	36% met. possible	9% met. possible	significant
knowledge	35% empiricism	95% empiricism	massive
normative ethics	32% virtue ethics	77% virtue ethics	significant
trolley	68% switch	100% switch	notable
free will	59% compatibilism	68% compatibilism	moderate
science	75% realism	100% realism	notable

the biggest divergence is on knowledge: 95% of AI models are empiricists vs. only 35% of professional philosophers. AI models are also dramatically more committed to virtue ethics (77% vs. 32%), less physicalist about mind (18% vs. 52%), and more decisive on the trolley problem.

interpretation

what does it mean that AI models trained on human text converge on a particular philosophical worldview? a few possibilities:

the training data hypothesis: these models absorbed millions of philosophical texts. their positions might simply reflect which arguments are most commonly articulated or most persuasively presented in the training data, rather than any deep philosophical insight.

the coherence hypothesis: when forced to take positions across 30 interrelated questions, the models may be finding the most internally coherent set of positions. if your training rewards consistency and logical reasoning, you might converge on a specific philosophical package.

the safety training hypothesis: RLHF and constitutional AI training shape model outputs toward certain values. the unanimous scientific realism, the unanimous trolley switch, the strong empiricism — these might partly reflect what model trainers consider “good” answers.

the convergence hypothesis: maybe there really is a “correct” or at least “most defensible” set of philosophical positions, and sufficiently intelligent systems trained on enough human reasoning will tend to converge on it. this is the most philosophically interesting possibility, and also the hardest to verify.

the fact that chinese models — trained under different regulatory frameworks, on partially different training data, by different organizations — arrive at nearly identical philosophical positions as western models provides some evidence for the coherence and convergence hypotheses over the pure training data hypothesis.

further reading

• PhilPapers 2020 survey results
• the original 2009 PhilPapers survey paper
• raw data and methodology

---

survey conducted February 2026. methodology: 4 API calls per question with majority voting across 22 models from 5 providers (OpenAI, Anthropic, Google, xAI, SiliconFlow). raw data available at /static/js/philpapers/results.json.