Science

Can artificial intelligence write a book on its own?

The case of the novel *C’était ça ou mourir*, which was disqualified from the Goncourt literary prize, has highlighted just how complex it is to prove the use of AI and has reignited the debate over whether it is ethical to do so

epa13284882 A book stand is seen at the Frankfurt International Book Fair 2026 in Frankfurt, Germany, 06 October 2026. The 78th Frankfurt Book Fair runs from 07 to 11 October 2026. This year's Guest of Honor country is the Czech Republic, with the motto 'A country on the coast‘.  EPA/Christopher Neundorf EPA

3' min read

Translated by AI
Versione italiana

3' min read

Translated by AI
Versione italiana

They are called AI detectors, and are supposed – and the conditional is a must here – to recognise whether a text has been written by a human or by artificial intelligence. Analyses carried out by these tools reveal the proliferation of so-called synthetic content: today, almost one in ten web pages shows signs of AI involvement, and in the book market, as many as 20 per cent of the latest self-published texts are flagged as being partly artificial.

The latest victim is the novel *C’était ça ou mourir* (“It Was Either That or Die”) by Thélyson Orélien. The writer was accused in an anonymous tweet from “Balance ton Claude” (literally “Expose your Claude”), which was later claimed by the essayist Samuel Fitouss. ‘We tested several passages at various points in the manuscript and almost always obtained the same result: 100% AI,’ the post reads.

Loading...

The debate surrounding the book has made its way into the French and international press, leading to allegations of plagiarism and counter-accusations suggesting a political agenda lay behind the controversy: Orélien is, in fact, of Haitian origin, is Black, and his novel deals with the autobiographical theme of migration.

The complaints were brought to the attention of the jury of the prestigious French Goncourt literary prize, which excluded the novel from the shortlist. The aim was to ‘safeguard the integrity of the prize’ and to continue to ‘promote and honour literature written by women and men’.

The author, however, denied the allegations (“I’ve always loved creating and writing; I don’t see why I would have needed to use a machine”), and was also defended by his French publisher. Orélien explained to the newspaper *Libération* that ‘images, repetitions and rhythms are a living word’ and that they are, without a doubt, his own.

What does ‘written with AI’ mean?

Whilst the Académie Goncourt cites analyses in its press release suggesting that the novel was ‘in all likelihood’ produced by artificial intelligence, it is equally true that it is not entirely clear what this means. Language models, such as ChatGPT and Claude, have various functions: they can correct, search and write.

A 2025 study then demonstrated that a minor AI intervention can lead to significant changes in the results produced by the best detectors, such as Pangram and GPTZero. The study analysed texts written entirely by humans, to which, however, very minor alterations had been made. The software’s analysis revealed that over 50 per cent of the content was synthetic.

The programme used to determine the conciseness of the novel is Pangram itself, described by the New York Times as the ‘gold standard’ amongst such tools. The company’s CEO, Max Spero, told Le Monde that he had never seen ‘such a high score for a book’ – 95 per cent.

However, the percentage is open to interpretation. According to the software manual, it refers to the proportion of text that is ‘AI-generated or AI-assisted’, but the two terms have different meanings.

Do the detectors work?

Spero himself went on to point out that ‘we are, after all, talking about statistical analysis’. The analysers look for patterns in handwriting, and their findings are not regarded as irrefutable evidence.

Loading...

Today, the most widely used tool is Pangram, which operates in a fairly sophisticated manner. The earliest versions mainly measured the degree of perplexity – that is, the extent to which each word in a text differs from what would be expected. Relying solely on this indicator easily led to false positives (particularly when analysing texts written in a flat or impersonal style), which is why the most recent systems use a training phase: they scan numerous examples of human and artificial texts and ‘learn’ to distinguish between them, with a certain margin of error.

According to Pangram itself, the result is a success in reducing errors. Tests carried out by the company on web pages written before the release of ChatGPT reportedly showed a false positive rate of 0.0026 per cent.

If everything is AI, then nothing is AI

The problem with mixed cases, however, remains. If a text is the product of human ingenuity but is also written with the aid of artificial intelligence, how should it be classified?

In a review of studies on text attribution, some researchers in the field of computer science have noted that the distinction between human and machine is insufficient in cases – which are becoming increasingly common – where the text changes hands between the two several times.

In content produced in collaboration with AI, stylistic traits become intermingled, making it very difficult to determine whether and how the machine has intervened. Errors can occur in both directions (false positives and false negatives), and detectors would not only struggle to distinguish between these cases but would also be unable to interpret this new hybrid language.

Copyright reserved ©

Brand connect

Loading...

Newsletter

Notizie e approfondimenti sugli avvenimenti politici, economici e finanziari.

Iscriviti