In these times, when the fear of artificial intelligence plays notes that oscillate between legitimate fears and absurdity, AI-generated text detectors have begun to produce results that look straight out of a Valle-Inclán novel. Writer Pedro Torrijos demonstrated and posted on Twitter how much fun it can be to blindly trust these systems. His experiment left many scratching their heads: the detector rated the first paragraph as Don Quixoteby Miguel de Cervantes, as “86% generated by artificial intelligence”. So, according to this tool, Cervantes was a visionary or… a robot.

Of course, this was not an isolated case. Torrijos also presented his own 2013 text, entirely human and written long before the rise of generative artificial intelligence. The machine’s verdict was implacable: “99% AI”. Instead, when he inserted a Gemini-generated paragraph with only three words changed, the detector decided it was “48% AI”. Are we facing a world where too much good writing is suspected of being robotic, and robots can pretend better than humans? All indications are that yes.

There are serious problems behind the joke.. Proliferated in industries such as education and business, these detectors are used to assess the originality of texts and even decide on sanctions in cases of alleged plagiarism. However, experiments like Torrijos reveal its lack of reliability. If a text from over 400 years ago is “86% AI” and the current paragraph by a human is “99% AI”, What criteria do they use to issue a sentence? And what is most disturbing: who gave them such authority?

The problem is not only technical; It is philosophical and practical. These systems are based on probabilistic models that detect statistical patterns associated with the production of artificial intelligence. But when these patterns match human literary styles, especially in elaborate texts, the detectors tend to fail miserably. What looks “too perfect” or “too uniform” is labeled as robotic.ignoring centuries of literary tradition that precisely values ​​this stylistic excellence.

On the other hand, These detectors can also be easily fooled. With minimal editing, AI-generated text can pass as human and vice versa. In this context, the tool becomes more of a puzzle than a reliable resource. And meanwhile, users—students, writers, and professionals—end up being judged by a system that doesn’t even understand the meaning of what it’s evaluating.

These types of situations clearly show that we are still far from understanding and controlling the tools we have developed. In the meantime, writers and students will have to live with the shadow of suspicion and face algorithms that see robots where there are none. And who knows Maybe one day the detector will declare the same message to be 100% AI. If that happens, I can only say: «Cervantes, welcome to the club«.