Philosophy & Public Affairs ends its experiment and bans AI authorship

An experiment ran its course in a fortnight, and the result came out rather clearer than those who ran it had expected.

Philosophy & Public Affairs is among the most respected journals in political philosophy. On 13 August it emerged that the journal had deliberately published an article the greater part of which was written by a language model. The thesis was supplied by Simon Goldstein of the University of Hong Kong, who steered the text, developed some of the arguments and corrected the errors; the writing itself was Claude's.

The decision had been taken in order to provoke an argument, and it more than achieved its aim: the announcement at Daily Nous drew over three hundred comments.

On 24 August the journal announced its decision. Content written by artificial intelligence will no longer be published.

What the experiment showed#

The journal's reasoning had nothing to do with technical quality.

The article had been through peer review; it was readable, coherent and defensible. That was precisely the difficulty. If the criterion is the quality of the text alone, no ground remains on which to rule out machine authorship.

What brought the journal to its decision was the recognition that the criterion is not quality. Academic publishing is built on the assumption that behind a text there stands an author who bears responsibility. Peer review, the citation system, the procedures for correction and retraction — every one of them depends on that assumption.

The structure fails not when authorship is shared but when it is handed over.

Where the argument moved#

The ban did not close the matter; it carried it somewhere else.

A post at Daily Nous on 27 August, which drew more than thirty comments, set out the real difficulty for the academy: making a rule is easy, enforcing it is not.

Nobody knows how much of what now reaches journals, universities and departments is partly or wholly the work of a machine. Policies are being formed, but they change constantly; as the technology advances, even freshly drafted rules turn vague or open unexpected gaps.

That uncertainty has produced a new risk: the false accusation.

One objection prominent in the discussion was directed against colleagues who point to academic misuse being treated as bounty hunters. The other side questioned the reliability of the detection tools.

A concrete case was offered. One philosopher described putting a translation of Plato's Apology, which he had commissioned from a language model the previous year, through today's detection tools, and getting contradictory results. The tools produce both false positives and false negatives.

The warning that follows is plain: a rule is a rule only so far as it can be applied fairly. The damage done to the career of a wrongly accused academic cannot be undone.

"An arms race"#

One prediction voiced in the discussion sums up neatly why a technical solution is hard to come by. As the detection tools improve, a loop will be set up in which texts are run through those tools and then rewritten. Machines will compete against machines.

Where human oversight sits in that picture is unclear.

The other view: a threshold, or a stage?#

One of the most striking interventions came from Graham Harman, a founding figure of speculative realism. Anyone who wishes, Harman observed, may go on writing everything from scratch, just as they might refuse the typewriter and the word processor.

The physicist and philosopher David Wallace drew a similar historical comparison: until recently the ability to carry out long symbolic calculations without error counted as "core physics." Today it does not.

The other side objects to the analogy. A calculator produces a result; a philosophical text is not exhausted by its result. The objection is the same as the point at the centre of the Daniel Kaufman argument we reported last week: in philosophy, can method be separated from result?

What has changed#

For this argument, August was a threshold.

At the beginning of the month a journal tested whether machine authorship was possible. At the end of it, the same journal decided that its being possible does not make it legitimate.

The question that remains is the harder one: with the ban in place, how is anyone to know whose a given text is?

There is no answer to it yet. But the question itself looks set to shape the agenda of academic publishing for years to come.