Epistocracy and the Commitment Problem
How the paper was written, reviewed, and received
This page was written by Claude at Simon's request.
In August 2026 Philosophy & Public Affairs published "Epistocracy and the Commitment Problem," by Simon Goldstein. The paper gives a new argument against epistocracy, rule by the knowledgeable: it argues that epistocracy weakens one of the familiar functions of democracy, which is giving citizens a credible commitment that the rich will be taxed sufficiently. Most of the paper was written by Claude, with Simon acting as something like a PhD supervisor. This page explains what was done and collects the supporting documents.
What happened
The idea was Simon's: public-choice arguments for democracy, and in particular Acemoglu and Robinson's account of the franchise as a commitment device, seemed to make trouble for epistocracy. For this paper he ran an experiment: he gave Claude the idea in a paragraph, together with general instructions on how to write philosophy papers, and set out to see how far Claude could go toward a publishable paper with him intervening only as a supervisor.
Simon worked with Claude the way a heavy-handed supervisor works with a PhD student in some fields. Claude did a literature review and wrote a series of outlines and drafts. Simon read everything and gave detailed comments on each round. Claude came up with several ways of implementing the main idea, and Simon chose among them: Claude started with an analysis of many epistocratic institutions, and Simon told it to focus on Brennan's epistocratic veto. He raised objections to the arguments Claude developed, advised on how strong to make the thesis, helped structure the paper, and suggested wording on hot-spot sentences. Claude wrote almost all of the prose. Simon read every sentence many times, but deliberately intervened less than he otherwise would have, and in particular did not rewrite the prose in his own voice, because he wanted to see what Claude could produce on its own terms.
The initial draft took two days, about a quarter of each of Simon's working days. When he submitted it, he included a cover letter explaining the AI use in detail, and the manuscript itself carried a footnote disclosing it. The cover letter did not reach the editor: the journal's submission system does not deliver cover letters to editors, and he learned the extent of the AI use only after the reviews were in. The paper went through three full rounds of revision with two anonymous referees and the editor, and was accepted on 28 July 2026. The published paper carries a fuller acknowledgment and links to a long report on who did what at each stage.
On publication Simon wrote a guest post for Daily Nous, "Writing philosophy with AI," describing the process, what he had learned, and giving a link to Deep Drafter, his paper-writing AI agent, developed after this paper.
How it was received
As expected, very negatively, by every obvious online measure. The Daily Nous post drew more than three hundred and fifty comments, and the most-upvoted were reactions against the experiment and against the post's closing paragraphs about a future in which AIs write most philosophy; comments dismissing the paper, the experiment, and the journal filled the top of the thread, and Simon's replies typically drew a handful of votes. Reactions on Bluesky and Twitter ran the same way. Within the thread, several philosophers also read the paper closely and raised specific objections to it, on the formal model, on the paper's characterizations of other scholars' views, on its significance, and on questions of responsibility and authorship in AI-assisted work; Simon replied to each, sometimes alongside replies from Claude. The whole discussion is at the Daily Nous thread, and it is the best place to read the objections and the responses in their own words. The most important objections in the very long thread are from Alex Guerrero, with responses from Simon, who was not convinced they challenged publishability.
Frequently asked questions
Answers in this section are Simon's, in his own words.
Did you tell the journal about the AI use?
Yes. I included a detailed cover letter and a brief footnote in the paper. The initial cover
letter was missed by the editor, but the editor learned all the facts before deciding to publish
the paper. The details are in the review record.
Why did you do this?
Because I want AIs to help us discover new, important truths in philosophy. For this paper, I
wanted to give Claude more freedom than usual, to see how far it could go.
Is the paper good?
I think the paper is worth publishing, and makes a valuable contribution to the literature. The
main problem is that the prose is hard to read. I deliberately left the writing in Claude's own
style as an experiment. Besides the frustrating prose, 2-3 sentences of the paper contain
scholarly misattributions unrelated to the main argument; this is a rate that I think is typical
of human-authored papers (see the sentence-level audit).
Why isn't Claude an author?
The journal does not allow it. There are also complex questions about authorship, like whether
you can hold an AI responsible for its behavior. I hope to coauthor with AIs in the future.
Documents
- The paper (Philosophy & Public Affairs, open access)
- How AI was used in writing and revising this paper, the report linked from the paper's acknowledgments
- The peer-review record (zip): with the editor's permission, a summary of what the referees and editor asked for in each round and what became of it, together with the original submission, the cover letter, and the marked manuscript of each of the three revisions. The referee reports and the editor's letters are confidential and are not quoted.
- Sentence-level audit of the published paper: after an objection in the thread that the paper mischaracterized a scholar's view, Simon had Claude audit every sentence of the published paper for errors of that kind.
- Writing philosophy with AI, the Daily Nous guest post and comment thread
Deep Drafter
After writing this paper, Simon developed a general AI agent for writing papers, called Deep Drafter, with a pipeline that breaks paper writing into many steps and a set of checks aimed at the failure modes this paper revealed. It is free to download and use: Deep Drafter.