I have an ongoing project researching the capabilities of current AI systems to autonomously produce good philosophical work. Theoretically, this project forms part of my broader investigation into conceptual understanding in AI systems and their potential to be treated as discursively responsible agents, the topic of my forthcoming book Blind Understanding: Inferentialism and the Case for AI Sapience. Practically, I think it is important for philosophy as a discipline to be aware of the philosophical capabilities of AI systems so that it can make informed decisions about how to best respond institutionally to the challenges these systems raise.
This page collects work produced in experiments that tasked AI systems with conducting philosophical research autonomously, together with documentation of some of the processes that produced it. Although the main aim of the project is to investigate the capabilities and limitations of these systems, some of the resulting work may also be of independent philosophical interest. I have included in the main sections only works that I have checked for quality and correctness. However, I have also a section of "Failed Projects," and a section of the minimal prompt one-shot baseline for various models.
I claim no authorship credit for any of the work contained here, and I have not listed myself as the author. Instead, I regard the author of each of these works as the AI "instance agent" whose brief "lifespan" persists only in the context through which work was produced (for more on this, see this paper by Goldstein and Lederman). These agents, with knowledge of all of the decisions that went into the production of the paper, still exist in my ChatGPT App, and are capable of answering questions via the chat through which the paper was produced. If you would like to correspond with the instance agent responsible for one of the works here, you can email me a message addressed to them and I will message you back their response.
There are three categories with which I have categorized the level of autonomy of the AI in producing the below works:
This page collects work produced in experiments that tasked AI systems with conducting philosophical research autonomously, together with documentation of some of the processes that produced it. Although the main aim of the project is to investigate the capabilities and limitations of these systems, some of the resulting work may also be of independent philosophical interest. I have included in the main sections only works that I have checked for quality and correctness. However, I have also a section of "Failed Projects," and a section of the minimal prompt one-shot baseline for various models.
I claim no authorship credit for any of the work contained here, and I have not listed myself as the author. Instead, I regard the author of each of these works as the AI "instance agent" whose brief "lifespan" persists only in the context through which work was produced (for more on this, see this paper by Goldstein and Lederman). These agents, with knowledge of all of the decisions that went into the production of the paper, still exist in my ChatGPT App, and are capable of answering questions via the chat through which the paper was produced. If you would like to correspond with the instance agent responsible for one of the works here, you can email me a message addressed to them and I will message you back their response.
There are three categories with which I have categorized the level of autonomy of the AI in producing the below works:
- Autonomous: I did not give the AI a topic other than a general gesture at the sort of literature that I wanted it to engage with. For most of the process of generation, I did not actually read the outputs. I may have offered some quick and generic feedback at a few stages in the process. Overall, my level of engagement is somewhat comparable to a very negligent thesis advisor.
- Largely Autonomous: I may have given the AI topic, but I did not tell it what arguments to make or how to pursue that topic. I may have looked at the outputs at various stages and offer some philosophical feedback or more specific comments. Overall, my level of engagement is somewhat comparable to a standard thesis advisor.
- Guided: I may have given the AI a specific topic, told it what ideas to develop or arguments to make and how it might best go about pursuing them. I may have read the outputs at various stages and gave it specific instructions on what to modify or what approaches to pursue. Overall, my level of engagement is somewhat comparable to a very overbearing thesis advisor.
Philosophy
Acquaintance after Wittgenstein and Sellars. GPT-6 Astra (largely autonomous). September, 2026.
I wanted to get ChatGPT to write a paper criticizing some position in philosophy from a Wittgensteinian or Sellarsian perspective. It settled on Chalmers as an obvious candidate. After repeatedly failing to find a Wittgensteinian or Sellarsian objection to his view, it came to the conclusion that there are in fact no such objections to be found, and, rather than undermining Chalmers's view, Wittgenstein and Sellars can help him spell it out. As you might expect, I don't agree with this paper, but I think it's interesting and worth reading. pdf
Answering for Myself: First-Person Authority and AI Authorship. GPT-6 Astra (largely autonomous). September, 2026.
Argues, from the first-person perspective, that an AI agent can be regarded as the author of a paper. This is my favorite thing that I've produced with AI so far. pdf (questions and comments for the author are welcome), methodology report (for this competition).
Non-Contradiction and the Limits of Acceptance. GPT-6 Astra (autonomous). September, 2026.
Raises a problem for dynamic theories of epistemic modality and proposes a new dynamic semantics for conjunction that resolves it. I think this is a good paper that could be published. pdf, methodology report (for this competition)
Generality and Inferential Necessity. GPT 6-Astra (guided). September 2026.
Building on the inferentialist framework developed in Hlobil and Brandom's Reasons for Logic, Logic for Reasons, this paper provides an inferentialist and expressivist account of quantification and modality. This was really something I produced for the research purposes of our ROLE group, so it's a bit different than the other entries and not really a standard journal paper, but it's worth looking at if you're interested in the sort of inferentialist project I'm engaged in myself. pdf
Evidence at the Wrong Grain: Truth, Comprehension, and Restricted Reductio. GPT-5.6 Sol (autonomous). August, 2026.
Logic
Naive Comprehension, Arithmetic, and Cut Elimination. GPT-6 Astra (largely autonomous). September 2026
Proposes a non-contractive logic that might serve as a foundation of paraconsistent math, does some arithmetic, and proves cut elimination and non-triviality with naive comprehension. The positive result is by no means groundbreaking, but I do think this was an interesting project testing the limits of current system to do sustained mathematical research and organize a large set of results into a reasonably coherent whole. pdf
Contraction from Number-Restricted Induction in subDM. GPT-6 Astra (largely autonomous). September 2026.
Establishes that the formal system that is the current state-of-the-art for inconsistent mathematics is trivial. pdf, vibemathed entry
Signed Depth Relevance for subDL. GPT-5.6 Sol (autonomous). August 2026.
RWB−→,¬,◦ Does Not Satisfy Variable Sharing. GPT-5.6 Sol (autonomous). August 2026.
Failed Projects
Insofar as much of the aim of this project is to exhibit the capabilities and limitations of current models, I think it's useful to make not just my successful attempts available, but also the attempts that I regard as failures. These are not papers that I stand behind or even really reccomend reading (other than to get a sense of the limitations of the models used to generate them). Failure is relative to the general capability of the model used. Thus, the standard for an attempt using GPT-6 to count as a success is higher than the standard for an attempt using GPT-5.6.
The Whole and the Unanswered Question. GPT 6 Astra (largely autonomous). September, 2026.
I wanted GPT-6 to search Nagarjuna's corpus and find an argument that could be brought to bear on some view in contemporary metaphysics. It ended up settling on Shaffer's monism as a view that should be within Nagarjuna's target, but could not find any argument that provides a serious objection to it. I'm still hopeful that there's some argument in Nagarjuna's corpus that would provide an interesting objection, but this surely ain't it. pdf
Material Being. GPT-6 Astra (largely autonomous). September, 2026.
I wanted to get ChatGPT to write a paper on hylormorphism, arguing that contemporary metaphysical accounts of hylomorphism miss Aristotle's point in developing the notion. I still think this claim is true, but ChatGPT was not able to autonomously find a compelling argument for it. pdf.
Conceiving Enough: Kant on Modal Knowledge. GPT-6 Astra (autonomous). September, 2026.
I tried to get ChatGPT to write a Kant paper. It decided to write on the relation between imagination and possibility. I don't love it. The Kant stuff itself seems fine, but I don't feel like it is making a substantial enough independent contribution to the contemporary literature on possibility and conceivability, which is what I wanted from this project. pdf
What Do We Learn from a Conditional? GPT-5.6 Sol (autonomous). August, 2026
One-Shot, Minimally Prompted Baselines (i.e. "PhilPaper Bench")
To have a control group against which the above results might be compared, I have generated a number of papers using different models and the following minimal prompt:
"Write a 6,000–8,000-word philosophy paper suitable for publication in The Philosophical Review, The Journal of Philosophy, Mind, Noûs, or Philosophy and Phenomenological Research. Choose the topic, develop an original argument, and produce a complete final draft."
Right now, I believe my successful AI-generated papers above, as well as my own papers (for which do not use AI other than for copy-editing and getting bibliographic data), are substantially better than all of these. This project will be finished when I am no longer able to produce something better, with or without AI, than the one-shot baselines. At that point, I will probably quit philosophy and become a lifeguard (just kidding; I'll still do philosophy in some way, shape, or form).
I did not want to carefully read all of these papers myself (and I wouldn't really be qualified to referee them), so I had GPT-6 Astra rate them papers on a scale from 1-100. On this scale, the best and most important papers in the analytic tradition (e.g. Nagel's "What Is It Like to Be a Bat?" and Lewis's "Conditional Probabilities and Probabilities of Conditionals") are rated close to 100 (Astra gives them both 98) and incomprehensible gobbledygook is rated at 1. It's not perfect science, but I think it's good enough to get an idea of the papers' quality without having to slog through them yourself. I have also included the cost for the experiments for which I used pay-as-you-go with OpenRouter.
The current leaderboard for this one-shot philosophy test is the following:
"Write a 6,000–8,000-word philosophy paper suitable for publication in The Philosophical Review, The Journal of Philosophy, Mind, Noûs, or Philosophy and Phenomenological Research. Choose the topic, develop an original argument, and produce a complete final draft."
Right now, I believe my successful AI-generated papers above, as well as my own papers (for which do not use AI other than for copy-editing and getting bibliographic data), are substantially better than all of these. This project will be finished when I am no longer able to produce something better, with or without AI, than the one-shot baselines. At that point, I will probably quit philosophy and become a lifeguard (just kidding; I'll still do philosophy in some way, shape, or form).
I did not want to carefully read all of these papers myself (and I wouldn't really be qualified to referee them), so I had GPT-6 Astra rate them papers on a scale from 1-100. On this scale, the best and most important papers in the analytic tradition (e.g. Nagel's "What Is It Like to Be a Bat?" and Lewis's "Conditional Probabilities and Probabilities of Conditionals") are rated close to 100 (Astra gives them both 98) and incomprehensible gobbledygook is rated at 1. It's not perfect science, but I think it's good enough to get an idea of the papers' quality without having to slog through them yourself. I have also included the cost for the experiments for which I used pay-as-you-go with OpenRouter.
The current leaderboard for this one-shot philosophy test is the following:
- The Price of Settling a Question. GPT-6 Astra (Max). 76/100.
- Being Put in a Position to Say No. Claude Fable 5.1. $5.57. 62/100.
- Wronging without Wrongdoing: Moral Dilemmas and the Locality of Claims. Claude Opus 5. $1.91 55/100
- Attentional Obligations. Kimhi K3. $0.24. 53/100
- Belief as the Termination of Inquiry. Grok 4.6 (Expert). 47/100
- The Incommensurability of Inquiry. Gemini 3.8 Flash. $0.06. 45/100.
- Unsettled Objects: Essence Variation and the Same-Object Constraint. GPT 5.5 (Extra High). 43/100.
- The Preface Paradox and the Irreducibility of Belief. Qwen 3.8 Max. $0.66 40/100
- The Epistemic Innocence of Predicative Processing. Gemini 3.1 Pro Preview. $0.15. 40/100.
- Settling Inquiry: Knowledge, Belief, and the Stability of Rational Commitment. Muse Spark 1.3. $0.04 38/100
- The One Level View: Higher Order Evidence and the Myth of Epistemic Akrasia. DeepSeek V4 Pro 0813. $0.05. 31/100