KEY INSIGHTS
- The wrong arithmetic: The question is not how much cognitive work stays inside one head but how labour is divided between humans and machines, and what that does to judgement.
- Authorship means answerability: Following Barthes and Foucault, the author is not the sole origin of a text but the person who can answer for it.
- Writing set the pattern: Externalising memory cost something, as Plato’s Thamus feared, yet it made sustained argument across centuries possible.
- Evidence from medicine: In a 2026 randomised trial of 70 clinicians, AI as a second opinion lifted diagnostic accuracy from 75 to 82 per cent while doctors kept responsibility.
- Authority is the real stake: AI harms us only when the division of labour erodes our capacity to understand, evaluate, redirect, and answer for the work.
On Danish television, the country’s culture minister, Zenia Stampe, recently confronted the entrepreneur Martin Thorborg, who had used artificial intelligence (AI) while writing his autobiography. Stampe’s conclusion was blunt: in that case, he had not written the book himself.
Behind the dispute lies a larger question. When AI performs operations once carried out by a writer – formulating, comparing, analysing, revising – does that necessarily leave less human authorship behind? Much of the debate assumes it does. More machine, less human.
This is the wrong arithmetic. The important question is not how much cognitive work remains physically inside an individual. It is how cognitive labour is divided between human beings and machines, and what that division does to human judgement.
The argument that AI diminishes authorship depends on a particular image of the writer: the sovereign author. On this view, writing belongs to me because the relevant cognitive operations originate within me.
More than half a century ago, Roland Barthes attacked precisely this image of the author as the origin of the text. We write in a language we did not invent and think with concepts inherited from others. Books, teachers, colleagues, editors, and critics all take part in forming what eventually appears under one name.
Michel Foucault later shifted attention to what the author’s name actually does. The author-function, he argued, is historically variable: “We can easily imagine a culture where discourse would circulate without any need for an author.” The author is therefore not simply the natural origin of a text, but a historically specific way of attributing texts to someone who can be identified with them and called upon to answer for them.
Writing Moved Memory Outside Us
The anxiety about new technologies taking over human capabilities has a far longer history than the recent debate over AI.
In Plato’s Phaedrus, Socrates tells the story of the invention of writing. The Egyptian god Theuth presents it as a gift that will improve memory and wisdom. King Thamus objects that it will do the opposite: once people can rely on external signs, they will stop exercising their own memory.
Thamus was partly right. Writing reduced the need for certain forms of biological memory. What he did not see was what this loss made possible.
Thoughts could now be preserved, compared, and developed across time. Arguments could extend over hundreds of pages. One could place a thought beside another written centuries earlier. Human beings became less dependent on one internal capacity while becoming capable of forms of thought that unaided memory could never have sustained.
This is the pattern to remember when discussing AI. We notice at once what happens when a cognitive operation leaves the individual. We are much slower to see what becomes possible when the larger process is reorganised.
AI is more than another step in this history of externalisation. Books stored thought outside us, but they remained silent. Plato’s Socrates noticed this too: ask a written text a question and it simply repeats itself. Challenge it and it cannot defend its position.
Generative AI changes that condition. The externalised material becomes responsive. I can formulate an argument and ask what follows from it. I can reject the answer, modify the premise, and try again. If I write about Plato, a language model can challenge my interpretation or point to something I have overlooked. This changes the structure of writing. Part of the intellectual world outside my head can now respond while an argument is being formed.
The obvious objection is that AI makes mistakes. It hallucinates, distorts sources, and confidently produces nonsense.
But humans make mistakes too. We misremember, read selectively, and grow attached to a hypothesis, then interpret new evidence through it. The relevant choice is rarely between a fallible machine and an infallible human. It is between a human working alone and a human working with another fallible system that may fail differently.
Medicine Shows Why Judgement Matters
Medicine offers a useful example.
It sounds reassuring to insist that a diagnosis should be made by a doctor rather than a machine. Yet responsible medicine has never meant that a doctor’s judgement should remain untouched by other forms of judgement.
A hypothesis-generating study published in 2026 examined 70 clinicians working through diagnostic cases. With conventional resources, their average performance was 75 per cent. When AI was introduced as a second opinion after the clinician’s initial judgement, performance rose to 82 per cent. When AI gave the first assessment and the clinician then evaluated it, the figure reached 85 per cent.
The lesson is not that AI should replace doctors. It is that human judgement does not become more human by remaining uncontaminated by external assistance.
A second opinion from AI can improve professional judgement while the doctor remains responsible for assessing it.
None of this means that concerns about AI are misplaced. AI can make us less capable. If students stop developing the knowledge required to recognise a bad argument, if journalists put their names to texts they do not understand, or if writers accept formulations simply because they sound convincing, something important has been lost.
But the loss does not occur simply because a cognitive operation has been performed outside the human mind. It occurs when the division of labour erodes the human capacity to judge the work being done. This is where the debate should move.
The central question is human authority within a distributed cognitive process. Authority means being able to understand the work one participates in, evaluate and redirect it, give the resulting text one’s own voice and character, and answer for the outcome.
A doctor may use an AI system without personally performing its calculations, but must retain the expertise to assess the diagnosis. The same applies to the writer, who need not be the sovereign origin of every operation involved in producing a text, but must ensure that the text conveys her own thinking and message, and be able to defend what remains on the page.
Two people can therefore spend three hours with the same language model and be engaged in entirely different activities. One delegates the problem and accepts the answer. The other uses the model to test an argument, expose weaknesses, and repeatedly force herself to explain why one interpretation is preferable to another. Calling both simply “AI use” tells us very little.
To move beyond that category, we can apply the Henry Ford principle. It suggests why AI does not necessarily subtract from human authorship simply because it takes over operations once performed by the writer.
In a new book called The Henry Ford Principle, Bent Dalager makes the argument clear: when a new technology appears, we tend to evaluate it by the properties of the technology it replaces.
Horse-drawn transport created problems involving horses, such as disposing of manure. Handling manure, however, was never the function of transport. Getting people from A to B was. It would therefore have made little sense to criticise the automobile because it offered no solution for manure.
We risk making the equivalent mistake with AI. Just as handling horse manure was never the function of transport, having every sentence traceable to cognitive operations inside a single human mind may not be what defines human authorship. AI forces us to distinguish between the operations historically associated with writing and the capacities necessary for authorship itself. Perhaps it lies in a particular division of cognitive labour in which the human author retains the capacity to understand, judge, give voice and existential resonance to, and take responsibility for what is being produced.
Help Keep Social Europe Free for Everyone
We believe quality ideas should be accessible to all — no paywalls, no barriers. Your support keeps Social Europe free and independent, funding the thought leadership, opinion, and analysis that sparks real change.
— €4.75/month
Help sustain free, independent publishing for our global community.
Become a supporter— €9.50/month
Go further: fuel more ideas and more reach.
Become an advocate— €19/month
Make the biggest impact — help us grow, innovate, and amplify change.
Become a champion— €142.50/month
For organisations and professionals who rely on Social Europe in their daily work.
Become a member



