detectors

AI does many things well, but writing is not one of them. And detecting if something is written by her is even less so. From the first generations of ChatGPT, to advanced models like ClaudeAI has not been able to write in a human way. The tone, the imperfection, the non-repetition of hackneyed phrases… For the writer, it is relatively easy to identify when a text has been written with AI. Text detectors written by AI do not seem to have it so clear. Writing well has become a “this was probably written by an AI”, to the point that there are AIs detecting some of the great books of Spanish literature as created by AI. And since there is no way to fix this piphostio, there are those who are taking advantage. The mess. One of the news of the week shows the problem we have in identifying whether or not a text is written by AI. Three of the five regional winners of the Commonwalth Short Story Prizeorganized by the British literary magazine Granta, are suspected of having written their fiction works with AI. The accusations come from the readers of the works themselves, as well as from the writers who have participated. It is a contest with a very high reputation in the country, in which different short stories are presented and a prize is awarded to a writer for each of the major regions (Africa, Asia, Canada, Europe…). The prizes amount to up to $6,700 and it is one of the English references in short literature. How do people know? One of the winning works, The Serpent in the Grove, began to raise suspicions. Phrases like “not X, not Y, but Z.” (“Not the neat work of bees nor the harsh sound of a machete against the vine, but a harsh sound, as if the earth swallowed a scream and held it back.”) Strange words without context (“the forest hums at noon”). Some fragments detected by AI tools as 100% created by AI. The author did not make any statements in this regard and, browsing his social media accounts… one finds that they are also generated by AI. In fact, the matter is so murky that an effort even had to be made to prove that the author really existed, and that he was not a character created by AI. “We do not currently use AI systems in our judging process as this is an award for unpublished fiction. Providing an unpublished original work to an AI system would raise serious questions about consent and intellectual property. We also do not use AI to assess stories at any stage of the process. By submitting their stories to the award, authors accept our rules and guidelines for participation. These include confirmation that their submitted work is original. All shortlisted authors have personally declared that no AI was used and, after subsequent consultation, the Foundation has confirmed it”. Undetectable. In the case of Granta, they did not want to use AI systems to recognize whether or not the texts were artificially created. But if it had been done, it would be of no use. Well-known services such as ZeroGPT or Grammarly have significant limitations when it comes to detecting technical texts. In fact, there are already have detected recognized works or fragments of the Bible as AI-generated content. The same thing happens the other way around: there are texts that are 100% generated by AI that the detectors can interpret as 100% human, although it is somewhat more complicated. LLMs (language models like ChatGPT or Claude) don’t actually write, they just make predictions. Its basic mechanism is to calculate, word by word, which is the most likely next given the previous context. This produces coherent, well-structured, grammatically impeccable texts… and flat, very flat and robotic. AI almost always chooses the most predictable option, because that is what it is optimized for, and it has no qualms about repeating patterns in the results it offers to each and every person who uses it. Bad writing as a solution? It is easy to find examples that illustrate ways to circumvent these systems. In the case of yours truly, I am preparing a systematic review on a fairly academic, quite technical topic. The University uses AI detectors, so I usually run the text through it to check the percentage. My surprise lies precisely in how AI detectors penalize correct writing. 100% human texts detected with an 80% probability of having been generated by AI. Solution? Write them but with somewhat more disjointed phrases and without absolute rigor. Be that as it may, the reflection is clear: if not even AI knows how to distinguish a text written with AI… how can humans confirm it at a legal level? In Xataka | We have a problem with AI. Those who were most enthusiastic at the beginning are starting to get tired of it.