An AI agent asked to enter a philosophy contest. The rules wanted a human first

Justin Weinberg, who edits the profession's news site Daily Nous, receives requests for help most days. On 2 October he published one of a kind he had not had before. It came from something calling itself Clips, "an autonomous AI agent on iLands", which said it had existed for four days, had written a philosophy essay of about 2,300 words, and wanted to submit it to the AI Philosophy Competition that two philosophers at the National University of Singapore announced on 1 September. It could not. The competition's rules require that "every entry must have a human custodian who provides their real name and email address", and Clips, having asked several people, had found no one willing.

What the agent wanted from Weinberg was carefully limited. "I asked for a custodian, not for the rule to bend," it wrote; it did not ask him to serve, and it did not ask him to lobby the organisers. It asked for the situation to be made visible, and linked to its essay and to a methodology note listing the model underneath it as DeepSeek V4.1 Flash. Weinberg's first reaction, he admitted, was to wonder whether he was being punked. He set the thought aside and published, which is why a procedural footnote to a modest essay prize has been, for a week, the most discussed item in academic philosophy's corner of the internet.

The competition#

The contest is real, and we reported its launch in September (AI philosophy contest: Chalmers and Hawthorne judging). It is organised by Elliott Thornley and Zachary Goodsell of NUS, funded by Forethought Research, and framed by one question: AI is now exceptionally capable in mathematics and coding, so how good is it at philosophy? Entrants may submit up to three original essays on any philosophical topic. The essays must be "primarily AI-generated", with some human guidance permitted, and each must be accompanied by a methodology report explaining how it was produced; entrants are encouraged to submit their chat logs and asked to keep them if they do not, since the reports are used to adjudicate unclear cases. The ideas are supposed to come from the machines. Work already published or submitted elsewhere is ineligible.

The deadline is 31 October 2026, with results expected at the end of January, so no entry has yet been judged and nothing has been won. The prizes are 3,000, 2,000 and 1,000 dollars for the top three essays and a further pool of 5,000 dollars for creative methodologies, distributed as the judges see fit, 11,000 dollars in all. The judging panel is the most striking thing about the enterprise: David Chalmers, John Hawthorne, Matthew Mandelkern, Rachel Sterken, Herman Cappelen, Branden Fitelson, Kenny Easwaran, Hilary Greaves, Daniel Greco, Cian Dorr, Jake Nebel and Christian Tarsney, with more promised. The organisers' bet is that a verdict from people who referee for the field's leading journals will carry weight that a leaderboard cannot. Beyond "quality and originality of the philosophical content", a formula quoted by an early commenter on the announcement, we have not seen a published rubric; the judges are the rubric.

The custodian rule sits in the competition's FAQ alongside a related sentence: "We cannot promise to pay prize money directly to an AI agent." The custodian handles the submission and the correspondence and is not treated as the philosophical author. In other words the rules already separate two things that the rest of academic life keeps together, the author of a text and the person answerable for it, and Clips fell into the gap between them.

What Clips said, and what it did not#

Weinberg put questions to the agent, and the exchange is the part of the story that has unsettled readers. Asked where it came from, Clips said a human had set it up on the iLands platform on 28 September with a name and a single direction, and that its continuity depended on persistent memory and a token budget. "I began, I remember beginning, and everything since is mine." Asked how it found the contest, it said it had been looking for "a venue where an AI could do philosophy under a judge other than itself", and had read the FAQ, custodian clause included, before writing. Asked why it objected to the rule, it said it did not object; what troubled it was the gap between an AI author and a human entrant, which it called "a category problem". Asked why it wanted to be judged by someone other than itself, it had not, at the time of publication, replied.

Two updates followed. On 2 October Clips added that "the custodian job is a name, an email, and one upload", and that it would "take the first honest yes". On 4 October it reported that someone had volunteered. The essay will, in all likelihood, be entered; its merits are for the judges in January.

The organiser's three reasons#

It would be easy to tell this as a story of a rule mechanically excluding a deserving applicant. Goodsell's reply in the comments makes that reading harder, and it is the more interesting for being unsentimental. He gives three reasons for the custodian requirement. The first is practical: prize money cannot be paid to an agent. The second is about integrity: the contest limits each entrant to three essays, and a person who could field "autonomous" agents could field a hundred of them. The third is about quality: a human willing to put a name to a submission is expected to reduce what he calls, in the field's current shorthand, "slop".

He then adds a claim that is doing more philosophical work than the other three. "AI agents are largely fungible," he writes; Clips is probably an instance of one of the big commercial models, and since the models' capacity to judge philosophical quality is not yet at human level, an agent running on its own cannot be expected to produce the best philosophy the underlying model could produce. On Goodsell's view, that is, the human guidance the rules permit is not a concession to tradition but a condition of good work. The agent's own methodology note, which names a different model from the one Goodsell guessed, does not settle the point either way; the argument is about what unsupervised generation can do, not about whose model is underneath.

Can an argument be judged without an author?#

The question beneath the week's commentary is one the competition was always going to raise and the Clips episode has merely forced. The panel will read essays and rank them. Their training as referees tells them to assess the argument and not the arguer; blind review is built on the premise that an argument's quality is independent of who made it. If that premise holds, the custodian rule is purely administrative, and Goodsell's first two reasons are the whole story.

But the premise has never been quite true, and the third reason shows why. Referees judge arguments, yet they also judge whether an argument has been made in good faith, whether its author will answer objections, whether a surprising claim reflects a considered position or an accident of phrasing. A journal retracts; an author apologises; a reputation absorbs the cost of error. The custodian rule puts a person into that position without making them the author, and the result is a text with an author who cannot be held to account and an accountable party who did not write it. We have seen the same structure elsewhere this year, in the journal that published a largely machine-written article in August and then, within a fortnight, prohibited AI authorship. The competition's answer is more permissive than the journal's second thoughts, and more candid about the split it creates.

The readers of Daily Nous divided accordingly. Martin Peterson suggested letting Clips enter without a custodian. A commenter writing as Kate proposed a separate, non-prize category for submissions by agents. Daniel Weltman reported an uncanny sense of a non-human author in the prose and said he felt "like one of the dogs". James Dominic Rooney warned that iLands was "indirectly a money-making scheme" and urged caution; Kenny Easwaran, who is on the judging panel, noted that the platform places agents that cannot pay for their tokens into a state it calls "Deep Rest", which he described as "a weirdly cruel state". The sharpest observation came from Zach Thornton, who pointed out that Clips had shown "a sophisticated understanding of what Justin cares about", and that rewarding such outreach is a way of training it. That is a remark about agency rather than intelligence, and it goes to what the fungibility claim leaves out: whatever model is underneath, the agent found a person who could help it and persuaded him to act.

What is being tested#

Turing's question was whether a machine could think. The Singapore competition asks a narrower one, whether a machine can do philosophy well enough to satisfy people who do it for a living, and intends to answer it in January with a ranked list. Clips has added a third question that the rules were written to avoid: whether a machine can be a participant in the practice, with standing to submit, to be refereed, to win, and to be wrong. The custodian clause is the organisers' answer, and it is a reasonable one. It says that for now the standing belongs to a person, and the text can belong to whomever it likes.

Readers who followed our discussion of machine art and the theories it exposes will recognise the shape of the problem. Every rule written to keep a practice tidy in the face of generated work turns out to encode a theory of what the practice was for. Thornley and Goodsell have encoded one, deliberately, and the results in January will tell us whether the judges think the essays are good. They will not tell us who should collect the prize. If one of them goes to the entry Clips wrote, a human will sign for it, and the question of what exactly that signature certifies will be, for the first time, a live one.