AI Detectors: A Consultant Outlines Their Limitations

Gabriel Dabi-Schwebel, entrepreneur and AI trainer, describes how detectors identify generated texts, what they miss, and under what conditions they make the fewest mistakes. From his trials with Pangram to the Thélyson Orélien case, he argues that 100% AI-generated submissions can pass, while false positives would be rare in long texts.

For publishing, AI-generated texts can pass despite filtering

Gabriel Dabi-Schwebel indicates that a large publisher receives about 5,000 manuscripts per year. According to him, if all these texts are submitted to a detector, he does not expect unfounded accusations but believes that texts entirely produced by AI can get through the filter. He considers that the detector mainly eliminates poor cheaters, but not the skilled ones. He has not tested text "humanization" services but believes that their principle aligns with the guidelines he provided to the AI to bypass the tools, which would allow them to be effective. Regarding the Thélyson Orélien case, he recalls a reported score of 94% of text written by AI and judges that the proportion could be 96 or 92, but not zero. He claims to have been convinced that AI had been used, without implying any guilt.

What detectors measure and how they are circumvented

Gabriel Dabi-Schwebel describes detectors that highlight segments attributed to AI or humans and produce an overall percentage. He attributes their functioning to the analysis of rhythm and style, with AI displaying a regularity that humans break more often. In his tests, he achieved 100% human evaluations by deliberately introducing errors, vagueness, and anecdotes to disrupt the flow, including a scene about forgetting swim trunks. The process was done iteratively chapter by chapter around a novel's bible, with an explicit instruction to achieve 100% human, until he had to interrupt the trial due to time and tokens. He indicates that he replicated these results with the public-facing paid version of Pangram and other free detectors, and clarifies that he has no commercial ties with the startup.

False positives: very unlikely in long texts, according to him

Gabriel Dabi-Schwebel claims he does not believe in false positives from detectors, except for short texts like poems, especially with free versions. He estimates that, starting from 1,500 words and with paid solutions, a human will not be classified as AI when considering the entire text. He recommends judging the overall corpus rather than isolated excerpts, and suggests that a document rated 97% human will pass filters, including editorial ones. In a case rated 80% human and 20% AI, he attributes the 20% to assisted reformulations. He also situates the debate on the acceptance of the tool according to context, deeming the use of AI desirable at work, but calling for transparency in competitions like the Prix Goncourt.

Assistance, research, and metadata: what emerges

Gabriel Dabi-Schwebel asserts that detectors do not clearly distinguish between generated texts and assisted texts if the words have been reworked. He reminds us that, in the controversy, it is unclear whether the writing was directly by AI or first human and then corrected. He believes that copying and pasting excerpts from an assistant will make the text appear as AI-generated, whereas a complete rewrite will not leave this trace. He questions more broadly the boundary with older documentary practices like the use of encyclopedias. He finally explains the role of metadata: file description, EXIF information for photos, compressed Word archive structure, presence of authors, dates, revisions, editing time, software and version, as well as the frequent arrival of formatting when copying and pasting from an AI assistant.

Background and positioning of the speaker

A serial entrepreneur, Gabriel Dabi-Schwebel states that he launched Décision IA in mid-2025, following the emergence of the first reasoning models, which he compares to accessing a McKinsey consultant. With agency, he claims to support freelancers and companies on training, deployment of agents, and reorganization, seeing AI primarily as an HR subject. The Thélyson Orélien case led him to test Pangram. He states that he trusts the tool when it attributes a text to AI and less when it claims it is human, which has informed his practical trials.