⏱ Temps de lecture : 2 minutes
Can an algorithm reliably decide whether a text was written by a human or by artificial intelligence? Through the case of Thelyson Orelien, this episode examines AI detectors, authorship and the growing fragility of proof in an age of machine-assisted writing.
Artificial intelligence is transforming writing, but it is also changing the way institutions determine who actually wrote a text. As generative AI becomes widespread, universities, publishers, media organisations and other institutions increasingly seek ways to distinguish human writing from machine-generated content.
AI-detection systems promise to provide an answer. By analysing statistical patterns in language, they attempt to estimate whether a piece of writing resembles text generated by artificial intelligence.
But probability is not proof. When a probabilistic system becomes an instrument of accusation, a fundamental question emerges: how much authority should an algorithm have when determining the origin of human expression?
The case of Thelyson Orelien brings this dilemma into sharp focus. Beyond one individual controversy, it illustrates a much broader problem: society is building systems designed to detect AI-generated writing precisely as the boundary between human and machine-assisted writing becomes increasingly difficult to define.
A writer may use AI to conduct research, correct grammar, translate passages, reorganise ideas or suggest alternative formulations while remaining the intellectual author of the work. At what point does assistance become authorship? And when does a tool become a collaborator?
The technical problem is equally significant. Linguistic characteristics interpreted as signs of AI generation can also occur in highly structured, academic or stylistically consistent human writing. Conversely, machine-generated text can be revised until many of its original statistical characteristics disappear.
The competition between generation and detection may therefore become structurally asymmetric. As generative models become more sophisticated and stylistically diverse, confidently identifying their contribution may become increasingly difficult.
This creates an institutional problem. Should a university, employer, publisher or competition jury be allowed to penalise someone primarily because an algorithm assigns a high probability of AI involvement? What evidence should be required? What appeal process should exist? And where should the burden of proof lie?
Beneath these technical questions lies a deeper philosophical issue: what does authorship mean when writing itself becomes increasingly hybrid?
This episode explores the emerging grey zone between human creation, algorithmic assistance and automated judgement. The ultimate danger may not simply be that machines learn to write like humans, but that other machines acquire the authority to decide whether humans wrote their own words.
