Reading time: 2 minutes
Can an algorithm be given the power to decide whether a text has been written by a human or by artificial intelligence? Through the case of Thelyson Orelien, this episode questions the limits of the d'IA detectors, the very notion of author and the increasing fragility of the evidence at the era of machine-assisted writing.
Artificial intelligence upsets writing, but it also transforms our way of determining who writes. With the generalization of generative models, universities, publishers, media and institutions now seek to distinguish between human texts and texts produced or assisted by a machine.
Algorithmic detection tools claim to meet this new demand. Using statistical characteristics of language, they try to estimate the probability that a text was generated by artificial intelligence.
But a probability is not evidence. When a probabilistic system becomes an instrument of accusation, a fundamental question arises: what confidence can be given to a machine when it claims to determine the origin of a text?
The case of Thelyson Orelien crystallises this question. Beyond the individual situation, it reveals a much broader problem: we are gradually building control mechanisms to detect the productions of the AI as the boundary between human writing and assisted writing becomes increasingly difficult to establish.
An author can search for information with an AI, ask him to correct his grammar, reformulate a sentence, translate a passage or suggest a structure while maintaining intellectual control of his text. When does assistance become written? And from what threshold can we consider that the machine becomes a co-author?
The difficulty is even deeper: the stylistic characteristics used to suspect the intervention of an AI can also be found in structured, academic or particularly regular human writing. Conversely, a text generated by a machine can be reworked until much of its initial statistical characteristics are lost.
The race between generation and detection could therefore become structurally asymmetric. As models progress and their productions become more diversified, the criteria for identifying their intervention with certainty may become increasingly fragile.
This situation poses a major institutional problem. Can a university, employer, publisher or jury punish a person solely on the basis of an algorithmic score? What adversarial procedure should exist? Who bears the burden of proof?
Behind the technical question finally appears a much broader philosophical question: what is an author in a world where writing becomes progressively hybrid?
This episode explores this new grey zone between human creation, algorithmic assistance and automated judgment. For the real danger may not only be that a machine writes in our place, but that another machine acquires the power to decide, rightly or wrongly, whether we have actually written what we claim to have written.
