An overly polished job application email, an assignment submitted without a single typo, a blog article with an oddly smooth tone: these situations increasingly lead readers to wonder if a text was generated by ChatGPT. Identifying content produced by artificial intelligence does not rely on a single clue, but on a bundle of signals that can be learned to recognize, even without specialized tools.
Stylistic Analysis: What the Regularity of the Text Reveals
Have you ever noticed that a text seemed “too clean” without being able to say why? This is often the first signal. A human naturally varies the length of their sentences: a short one, then a long one, then a medium one. ChatGPT, on the other hand, tends to produce sentences of comparable length, paragraph after paragraph.
Also, pay attention to punctuation. The complete absence of typos or hesitations is a strong marker. A human writer might let a misplaced comma or a forgotten dash slip through. A text that is perfect typographically, especially if it is long, deserves closer examination.
Another concrete clue: the linking vocabulary. ChatGPT frequently resorts to the same connectors (“moreover,” “furthermore,” “it should be noted that”). If you find more than three occurrences of the same connector in a 500-word text, the likelihood of automatic generation increases. To delve deeper into this type of analysis, you can discover Au Top’s articles on the subject.
An AI text almost always lacks personal bias. It presents, it lists, it nuances, but it does not take a stance. If an 800-word article on a controversial topic contains no strong opinions, that’s an additional signal.
AI Content Detectors: Real Reliability and Technical Limits
Several tools promise to spot a text generated by ChatGPT: GPTZero, ZeroGPT, Turnitin, and Draft & Goal. Their principle is based on statistical analysis. They measure “perplexity” (the degree of surprise of a model in response to each word) and “burstiness” (the variation in complexity between sentences).
A human text exhibits high and irregular perplexity. An AI text shows low and homogeneous values. That’s the theory.
In practice, detectors achieve only about 40% accuracy on unmodified AI text, according to a 2024 study involving 805 evaluations of texts produced by GPT-3.5 and GPT-4. Even more concerning: this accuracy drops to just over 20% as soon as the text undergoes simple modifications (rephrasing, minor additions, partial rewriting). By comparison, human evaluators achieve about 67% in the same protocol.
What Detectors Don’t Tell You
A “0% AI” result guarantees nothing. A “95% AI” result doesn’t either. False positives regularly affect writers whose second-language French produces more regular structures. Highly standardized academic texts, written without AI, also trigger alerts.
Relying on a single tool for a decision would be a mistake. Reliable detection always combines software with attentive human reading.
Human Reading: The Clues That Machines Miss
A detector measures statistical probabilities. An attentive reader picks up on something else: the absence of lived experience, personal reference, or deliberate awkwardness. Here are the most reliable signals to check manually:
- Examples remain vague or generic. ChatGPT writes “let’s take the example of a company” without ever naming that company, without specific industry context.
- Numbers are often rounded or presented in broad ranges. A human cites “37%” because they have read a source. ChatGPT rounds or invents a range (“between 30 and 50%”).
- The text contains no internal contradictions, no nuances that would weaken its own argument. A human writer sometimes admits a limitation. ChatGPT constructs a systematically balanced argument.
- Transitions between paragraphs are mechanical: each section starts with a sentence summarizing what follows, like an essay outline.
This last point is the most discriminating. A human text progresses through the association of ideas, not through systematic announcements. If each paragraph begins with an introductory sentence followed by three arguments, the pattern is typically that of a language model.
Legal Obligations for AI Marking in Europe Starting in 2026
The question “is this an AI text?” may soon find a technical answer mandated by law. Regulation (EU) 2024/1689, known as the “AI Act,” stipulates in Article 50 that generative systems like ChatGPT must mark their content (text, image, audio, video) as “artificially generated” in a machine-readable format.
These transparency obligations come into effect on August 2, 2026 for systems newly placed on the market. In practice, a text produced by ChatGPT should include metadata identifying its origin, accessible by compatible software.
This marking will not solve everything. It only covers content generated after the application date. It also assumes that the end user does not manually delete the metadata, which remains technically simple. Human detection therefore retains its relevance, even within a regulated framework.
What This Changes for Daily Detection
For a teacher, recruiter, or editor, the useful reflex remains the same: cross at least two approaches. Run the text through a detector for a first signal. Then reread while looking for the stylistic clues described above.
- Check for suspicious regularity in structure and vocabulary
- Test the text on two different detectors and compare the results
- Look for the absence of personal voice, anecdotes, or specific references
None of these clues, taken in isolation, prove that a text comes from ChatGPT. However, their accumulation constitutes a sufficient bundle to justify a thorough check. The best detection remains that which combines a statistical tool with a critical eye, capable of spotting what no algorithm yet measures: the imprint of a singular thought in writing.
