The Author after the Machine

In February 1969 Michel Foucault put a simple and difficult question to the French Society of Philosophy: what is an author? His answer was that the author is not a person hidden behind the text but a “function” that a society grants to certain kinds of discourse. It is a name that gathers texts together, sets them apart from others, and holds someone answerable for them. Two years earlier Roland Barthes had announced the death of the author so that the reader could be born. Half a century later we stand before a machine that produces text without pause and with no one behind it. Foucault’s question is no longer a theoretical exercise. It has become a question of engineering and of politics.

Cheap text, costly signatures

When producing text becomes almost free, value migrates from the text to the signature. A signature means there is someone who can be asked why they wrote this, how they know it, and what they will do if they were wrong. In February 2023 the Committee on Publication Ethics said as much in administrative language: an AI tool cannot be an author, because it cannot take responsibility for the work. The ruling is right, but it is not enough. Our trouble is not only that a machine’s name should stay off the list of authors. It is that unnamed, undated, unversioned text now circulates so widely that we no longer know who said which sentence, when, and on what grounds.

In 2021 Emily Bender, Timnit Gebru and their co-authors called large language models “stochastic parrots”: devices that stitch linguistic forms together without reference to meaning. The phrase can be read from the other side as well. If the machine produces form without meaning, then meaning lies where someone answers for it. Before knowledge is a mass of sentences, it is a web of commitments.

The university and the temptation to outsource

Traditional institutions of knowledge rested on a simple assumption: writing is hard, so a written text is evidence of work. An article was evidence of research, a thesis evidence of years of reading, a coursework essay evidence of understanding. That chain of evidence is broken. A university that built its evaluation on counting articles is now buried under a flood of text that shows nothing at all. The administrative response, software that claims to detect machine text, has been both ineffective and unjust. Liang and colleagues showed in 2023 that these tools flag the writing of people working in a second language as machine-made far more often than anyone else’s. Once again, those already at the margins are the ones accused.

Many universities have gone back to invigilated exams with pen and paper. That is a retreat, not an answer. The real question is left untouched: if a text no longer proves work, what does? My suggestion is that the proof is the process: drafts, revisions, conversations with teachers and peers, and the trace of decisions the writer made along the way. A university that measures only the finished product has handed itself over to the machine of its own accord.

Authoritarianism profits from this rupture. When no one can say who said what, every claim weighs the same as every other and whoever has the loudest megaphone wins. A flood of ownerless text serves, before anyone else, the kind of power that never accepts responsibility.

An architecture of answerability

The Persian manuscript tradition had an old answer to this problem. In the colophon of every manuscript the scribe recorded their name, the date and place of copying, and sometimes the exemplar they had copied from. Every copy knew where it came from, and it is with the help of these colophons that an editor today can reconstruct a text’s lineage and tell tampering from the original. In the age of print, and then of the web, we forgot this habit, because copying was cheap and we entrusted our confidence to the publisher. Now that making text has become cheap as well, we need the colophon again; this time for every sentence and every translation.

What I propose is not a return to the Romantic author, the solitary genius. What I propose is an architecture: a system of knowledge that ties every text to the people who answer for it, versions every change, and places the machine’s work, where there is any, beside the human’s work with a name and a date, never in its stead. Within the framework of language-agnostic knowledge this principle goes one step further. The translation of a text is another form of the same knowledge object, and its translator signs it too. The machine proposes; the decision and the signature stay with people.

The author has not died. The author has only moved: from behind the text to beside it, where they can be found and asked.

Questions for discussion

References

Foucault, Michel. 1969. “Qu’est-ce qu’un auteur ?” Bulletin de la Société française de philosophie 63 (3): 73–104.

Barthes, Roland. 1967. “The Death of the Author.” Aspen, no. 5–6.

Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21), 610–623.

Liang, Weixin, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou. 2023. “GPT Detectors Are Biased against Non-Native English Writers.” Patterns 4 (7): 100779.

Committee on Publication Ethics (COPE). 2023. “Authorship and AI Tools.” COPE position statement, 13 February 2023.