An AI text detector cannot prove that content comes from ChatGPT. It only estimates, based on statistical patterns, whether a passage resembles text generated by artificial intelligence. Errors remain possible, especially after rewriting or with certain styles. For a business, a school, or a publisher, the score must therefore trigger a human review, never decide on a rejection or a sanction on its own.
How does an AI text detector work?
An AI text detector is a classifier that compares the characteristics of a text with those of human and artificial examples used lors its training. In particular, it looks for predictable wordings, repetitive structures, and word distributions associated with generative models. The result remains a statistical estimate, not a trace of origin.
Services like Pangram or Turnitin transform analyzed passages into mathematical representations. Pangram explains in 2026 that its classifier seeks to separate, in this abstract space, the groups corresponding to human writing and AI-generated content. A new passage is then assigned to the group it most closely resembles.
Some systems also use « perplexity, » that is, the degree of predictability of words. A very regular sentence gets low perplexity and may appear artificial. Yet a technical manual, an institutional press release, or the text of a person writing in a foreign language may show the same regularity.
The detector has access neither to ChatGPT’s historical record nor to a secret signature placed in each response. Unlike the metadata and provenance mechanisms used to prove the origin of an image with C2PA, editorial detection relies on clues internal to the document. It generally does not make it possible to attribute a text with certainty to ChatGPT, Claude, Gemini, or another model.
Is a ChatGPT detector really reliable in 2026?
No universal reliability rate can be assigned to a ChatGPT detector in 2026. Performance changes depending on the language, length, writing genre, generative model, level of rewriting, and threshold chosen. An excellent result on a test set therefore does not guarantee the same result on your documents.
The OpenAI precedent remains instructive. In 2023, the company withdrew its own classifier because of its low accuracy. Lors its evaluation, the tool identified only 26 % of artificial texts as « likely AI-generated » and assigned this label to torat 9 % of human texts.
Tools have improved since then, but their figures must be read together with their protocol. The technical sheet published by Pangram on July 29, 2026 claims for Pangram 4 a false positive rate of 0,0041 %, or about one case in 24,000, on one million English examples from FineWeb. However, this measure comes from an evaluation set selected by the company and does not describe all uses, all languages, or all styles.
| Source and year | Reporored result | What the result allows one to assert |
|---|---|---|
| OpenAI, 2023 | 26 % of AI texts detected and 9 % of human texts misclassified | The evaluated classifier was insufficient to establish the origin of a text. |
| Stanford HAI, 2023 | 61,22 % false positives on average across 91 human TOEFL trials | Texts by non-native English speakers can be particularly exposed. |
| Frontiers in Education, 2024 | 5.2 % false positives despite the accord required between two detectors tested | Cross-checking two tools reduces some doubts without constituting proof. |
| Pangram 4, publisher fact sheet from 2026 | 0.0041 % false positives on one million English FineWeb examples | The result describes this specific internal test, not all documents encountered in practice. |
Honestly, comparing tools based on a single percentage marketing doesn’t make much sense. You need to know the corpus tested, the language, the decision threshold, and the proportion of artificial text. Without this information, two rates presented as comparable may measure very different situations.
Why can a human-written text be flagged as artificial?
A false AI positive occurs when a detector classifies text written by a person as artificial. The risk increases with predictable formulations, repetitive vocabulary, short texts, and styles far removed from the training data. A writing correct, restrained, or standardized style may therefore be suspected without any ChatGPT involvement.
The experiment published by Stanford HAI in 2023 is particularly telling. Seven detectors analyzed 91 TOEFL essays written by non-native English speakers. On average, 61.22 % were misclassified as AI-generated texts, and 97 % were flagged by at least one of the tools.
The researchers linked this bias to the low perplexity of these papers: their vocabulary and structures were more predictable. The problem goes beyond education. Product sheets, customer service responses, legal texts, or SEO pages often follow strict conventions that can produce a similar signal.
Professional correction can also blur classification. A preprint published in August 2026, based on 135,389 pairs of academic manuscripts by non-native English speakers before and after correction, indicates that style and editing can disrupt the results. Its prepublication status, however, means its conclusions should not be treated as definitively validated.
Conversely, a generated text can escape detection after moderate rewriting. Work published in 2024 observed that, in certain configurations, the detection rate fell to 0 % lorsque the threshold was set to limit false positives to 1 %. The human revision of content produced with AI often improrves its quality, but it also makes its statistical origin more difficult to establish.
What does the displayed percentage really mean?
The percentage shown by an AI text detector is not necessarily the probability that the author used ChatGPT. Depending on the product, it may represent the share of passages judged artificial, a score of confidence, or the proportion of text windows classified as suspicious. The definition must be checked before any interpretation.
Turnitin thus specifies in its 2026 documentation that its indicator porte concerns the amount of qualifying prose estimated as generated by AI. The service hides the exact scores between 1 % and 19 % and displays an asterisk, because this range has more false positives. A score of 15 % therefore does not mean there is an 85 % chance that the text is human.
Another difficulty: the threshold changes the trade-off between two errors. A strict setting reduces the risk of accusing a human author, but lets more artificial productions pass through. A sensitive setting detects more generated content, at the cost of a higher number of false alerts.
The French editorial controversy of September 2026 illustrates this. According to Le Monde, Pangram’s analysis of more than half of the novel It was that or die produced a result of more than 95 % artificial text with high confidence. However, this result does not by itself demonstrate the writing process: translation, correction, partial assistance, and classifier error remain distinct hypotheses.
How can suspicious content be verified without making false accusations?
Reliable verification of suspicious content combines the detector score, the historory of the document, drafts, sources, and the author’s explanations. In 2026, the available evidence supports using the detector as an initial screening tool. A sanction, an editorial rejection, or a dismissal should never rely on this signal alone.
A proportionate procedure protects both the organization and the author. It prevents a team from transforming an opaque prediction into a verdict, then having to defend a decision that is impossible to document.
- Keep the original document. Archive its version, its date of receipt, and the rules communicated to the author before launching the analysis.
- Read the definition of the score. Check the supported language, the minimum length, the applied threshold, and the passages actually analyzed by the product.
- Examine the flagged segments. Look for repetitions, shifts in style, nonexistent references, and unsupported claims rather than stopping at the overall score.
- Check the work traces. The historry of Google Docs or Microsoft Word, drafts, notes, and sources reveal the writing process better than an isolated classifier.
- Request an adversarial explanation. The author must be able to specify their tools, describe their method, and explain the passages concerned before any decision.
- Have it reviewed by a qualified person. For high stakes, document the human analysis, the tool’s limitations, and the other pieces of evidence.
In the editorial projects we carry out, we often see a more useful question than “Was AI used?”: is the content accurate, original, traceable, and conforme with the authorized level of assistance? This framework assesses concrete risk. It avoids confusing the use of a tool with misconduct.
The same principle applies to digital security. A technical alert requires human assessment, whether it concerns a text, a attempted fraud using deepfake in a video conference or a falsified document. The level of scrutiny must increase with the possible consequences of the error.
What policy should a company adopt for AI-generated content?
An effective company policy defines the authorized uses of artificial intelligence, prohibited data, the level of review, and the evidence to retain. The AI text detector can help priorize checks, but the rule must focus on quality, confidentiality, and accountability. The goal is not to guess a tool at all costs.
Start by distinguishing light assistance, such as grammatical correction, from the generation of a complete deliverable. Then specify the cases requiring disclosure: signed article, recruitment file, contractual analysis, academic work, or regulated communication. The boundary must be understandable before the content is submitted.
For published content, require verification of facts, quotations, and usage rights. For sensitive data, prohibit its transfer to an unapproved service. Finally, maintain clearly assigned human responsibility: one person validates the document and is accountable for its accuracy, even if an AI participated in its preparation.
On the agency side, the instinct is to test the procedure on a few known cases before generalizing it. A batch mixing human texts, generated content, and rewritten versions reveals the operational limits of the detector. Failing that test, it is better to invest in editorial traceability than in automated monitoring presented as certain.
Framing the uses, the expected evidence, and the dispute procedure avoids most weak decisions. An outside perspective can help choose controls suited to the level of risk, without turning a statistical scorre into proof it will never be.
FAQ on detecting AI-generated texts
Can we know whether a text comes specifically from ChatGPT?
A detector generally cannot determine with certainty that a document was written by ChatGPT. The result is based on statistical similarities and not on cryptographic proof provided by OpenAI.
Is passing a text through two detectors enough?
The agrorment of two detectors strengthens a signal without proving the text’s origin. A study published in Frontiers in Education in 2024 still observed 5.2 % false positives when the two tools tested had to agrore.
Are short texts more difficult to analyze?
Short texts provide fewer statistical clues and make classification more unstable. A comment, a brief email, or an isolated paragraph should therefore not serve as the basis for an importante decision.
Can a reformulation fool AI detection?
Human or automated rewriting can significantly reduce the detection of generated text. However, the research available between 2023 and 2026 shows that the effect depends on the model, the detector, the language, and the threshold used.
Can a high score justify rejecting a candidate?
A high score should not alone justify rejecting a candidate. The company must examine the instructions given, drafts, sources, writing historory, and the person's explanation before making a decision.