Is it true?
Is it true you can tell whether a text was written by ChatGPT?
By the end, you'll know what AI text detectors can really do, why they fail, and what difference "watermarks" make: invisible signatures that some companies are starting to add to their texts.
Translated with AI from the Spanish original. Nobody has reviewed this translation; if anything seems off, the Spanish version is the reference.
Partly. Today's detectors often get it wrong and accuse innocent people. Watermarks help, but they can be partly removed, only work for one company's texts and aren't available to everyone.
Where the claim comes from
Since ChatGPT took off in late 2022, many teachers, editors and HR managers have been asking the same question: did a person or a machine write this? Tools have appeared that promise to answer it. The best known are GPTZero and the detector that Turnitin, the anti-plagiarism software used by many universities, added in 2023.
It's worth separating this from plagiarism. A plagiarism checker looks for whether your text matches one that already exists. An AI detector can't do that, because every ChatGPT answer is new. Instead, it calculates probabilities: it checks whether the words and sentences look like what an AI tends to write (very predictable sentences, a very even rhythm) and gives a score.
The strongest version of the claim is reasonable: AI-generated text pasted in as is usually has recognisable traits, and a tool trained on millions of examples should spot them. On top of that, some companies are starting to mark their texts with invisible signatures, which in theory would make it possible to check without guessing. Let's see what the data says.
What the data says
The clearest evidence comes from OpenAI itself. In January 2023 it released a detector for AI-written text. By its own figures, it identified only 26% of AI-written texts as "likely AI-written", and wrongly flagged 9% of texts written by people. In July 2023 it withdrew the tool because of its low accuracy.
Independent studies point the same way. In 2023, a team of European researchers tested fourteen detectors, including the most widely used ones, on texts written by people, by AI, and by AI and then edited by hand. Their conclusion, published in the International Journal for Educational Integrity, was that none of them was accurate or reliable, and that rewriting the text slightly was enough to make them fail far more often.
There's another, more sensitive problem: false positives, meaning accusing someone of using AI when they haven't. A 2023 Stanford University study ran TOEFL English exam essays written by non-native students through seven detectors. More than half were flagged as AI-written. The reason: people writing in a language they haven't mastered use simpler, more predictable sentences, just like an AI.
Turnitin, for its part, says its false positive rate is low. But that figure comes from the company itself, and even if it were accurate, at a university with thousands of assignments a small percentage still means many students wrongly flagged.
Watermarks: another way to tell
There's an alternative to guessing: having the AI itself leave an invisible signature as it writes. This is called a watermark. The AI picks certain words over others following a pattern a person won't notice, but a program with the key can recognise. In 2024, Google DeepMind published its system for doing this with text, called SynthID, in the journal Nature.
In October 2026, OpenAI announced that it will watermark ChatGPT's texts this way in the European Union, to comply with the transparency rules of the EU AI Act. But there are important caveats. According to TechCrunch, the watermark detector will, for now, only be available to researchers and organisations approved by OpenAI. In other words, a teacher won't be able to check on their own.
Watermarks can also be partly removed. In a test by OpenAI itself, reported by TechCrunch, replacing 10% of the words with synonyms cut detection from around 92% to 66%. And a watermark only works for texts from the company that adds it: a text written with another AI, or with one that doesn't mark anything, leaves no trace.
The SynthID detector that Google opened to the public in October 2026 works, as of that date, for images, video and audio, not for checking ChatGPT texts.
What we don't know
We don't yet know how OpenAI's watermark will work in practice in Europe: when it will be active on all accounts, who will be able to access the detector later on, or how well it will hold up against people deliberately trying to remove it.
Nor are there recent, large-scale independent studies showing that today's commercial detectors have overcome the problems seen in 2023. Companies publish their own figures, but outside testing is still scarce, and AI models change every few months, so the tests have to be repeated again and again.
What is clear, as of November 2026, is that a detector score is not proof. It can be a clue for starting a conversation with the author, but it isn't enough to accuse anyone.
Watch out: If a detector flags your text as AI-written and it isn't, ask for it to be reviewed using other evidence: drafts, the document's version history or a conversation about the content. Studies show these tools make mistakes, especially with people writing in a language that isn't their own.
Sources
- OpenAI: nuevo clasificador de textos escritos con IA (actualizado al retirarlo)
- Liang et al. (Stanford), «GPT detectors are biased against non-native English writers»
- Weber-Wulff et al., «Testing of detection tools for AI-generated text»
- Nature: «Scalable watermarking for identifying large language model outputs»
- OpenAI News
- TechCrunch
- Google Blog