Quick answer
AI detectors can be useful, but they cannot reliably prove that a specific person used ChatGPT. Some current tools perform very well on clean, fully AI-generated text in controlled tests, while results vary sharply by detector and model. Hybrid, edited or humanized writing is harder to classify, and both false positives and false negatives remain possible. A detector score is best treated as one piece of evidence, not a verdict.
Contents
- How accurate are AI detectors in 2026?
- Why different studies get very different results
- False positives, false negatives and the problem with a single score
- Can AI detectors catch edited, paraphrased or partly AI-written text?
- What AI detector do colleges use?
- Can a detector prove you used ChatGPT?
- How to interpret an AI-detector result without overclaiming
How accurate are AI detectors in 2026?
There is no honest universal percentage for “AI detector accuracy.” The number changes with the detector, the model that produced the text, the kind of writing being tested, how much human editing happened, the threshold used to call something AI-generated, and even the subject area.
Current studies can therefore produce very different-looking results.
A September 2026 study in Studies in Educational Evaluation tested five detectors across 500 excerpts covering fully human-written, fully AI-generated, humanized AI and hybrid human–AI text. On the clean human-versus-AI task, the detectors showed extremely strong discrimination. Their performance dropped substantially once the text was hybrid or humanized.
Another 2026 higher-education study reached a different-looking result. It compared Turnitin, GPTZero, Copyleaks and Pangram on fully human, fully AI, hybrid and humanized academic papers. Pangram performed much better than the other tools on that dataset. Turnitin, GPTZero and Copyleaks substantially underestimated some AI-generated material, particularly material produced with a newer model. All four handled the study’s fully human papers well.
The studies are testing different tasks. A detector can look excellent on one evaluation and much weaker on another when the text type, model or scoring threshold changes. Blanket claims such as “99% accurate” or “useless” are poor summaries of the evidence.
| Study | Tools / text types | Strongest result | Main limitation | What it means |
|---|---|---|---|---|
| Rosenthal et al., Sep 2026 | 5 detectors; 500 human, AI, hybrid and humanized excerpts | AUC 0.994–1.000 on clean human-vs-AI discrimination | Performance fell substantially on hybrid and humanized text | Strong clean benchmark performance does not equal authorship proof |
| “Who wrote this?”, 2026 | Turnitin, GPTZero, Copyleaks, Pangram; 160 known-ground-truth papers + 1,163 theses | Pangram performed best; all tools handled the study’s human set well | Large between-tool/model variation; real-thesis sample had no known ground truth | Detector and model choice can change the answer dramatically |
| Hadra et al., Feb 2026 | Turnitin + Originality; 192 human/EFL/AI/hybrid texts | Originality 0.69 overall accuracy vs Turnitin 0.61 in this dataset | Both struggled with hybrid text; length and genre affected results | Context, genre and authorship mix materially change reliability |
| Sun et al., 2026 | 13 detectors; assignments, theses and code | Broad systematic comparison across academic tasks | Failures across tasks; hybrid adversarial edits let up to 88% of AI content evade detection | Too fragile to serve as sole high-stakes evidence |
Why different studies get very different results
The first thing to check in any detector claim is what was actually tested.
A detector trained or tuned around one family of model outputs may perform differently when the text comes from a newer model. A benchmark made from long, untouched AI passages is easier than a document where a person wrote most of the text and inserted a few generated paragraphs. A threshold chosen to minimize false accusations will miss more AI-generated text than a threshold designed to catch as much as possible.
Genre matters too. A February 2026 study comparing Turnitin and Originality.ai used 192 texts that included authentic English-as-a-foreign-language student writing, professional human writing, AI output and hybrid compositions. Originality.ai achieved higher overall accuracy than Turnitin in that study, but neither tool handled hybrid text consistently. Performance also changed with text length and genre, with scientific writing proving harder than humanities writing.
When you see an accuracy claim, check what kinds of mistakes the test allowed, what documents were sampled, which models produced the AI text, and whether the task resembles the document you care about.
A score that works well for separating two clean categories in a lab is not automatically a reliable authorship judgment on a student essay.
False positives, false negatives and the problem with a single score
A false positive is human-written text that a detector flags as AI. A false negative is AI-generated text that the detector treats as human.
Both matter, but the harm is different.
For a student, a false positive can trigger an accusation even though the work was written without prohibited AI help. For an educator trying to identify undisclosed AI use, a false negative means the detector misses the thing it was supposed to find.
Turnitin’s own product guidance acknowledges this tradeoff. Its AI Writing Report does not surface an exact percentage for detections between 1% and 19%, because Turnitin says false positives are more likely in that range. The company also states that its AI score should not be used as the sole basis for adverse action against a student.
An AI detector is a classifier: it estimates which category a piece of text resembles. It does not observe the writing process.
The real-world stakes also depend on how often the behavior being detected is actually present. Even a detector with a low false-positive rate can produce a meaningful number of wrong accusations when it is applied to very large numbers of submissions or when the underlying rate of prohibited AI use is low.
Some universities have responded by putting explicit limits around detector evidence. Washington State University says AI detection should not be the sole support for an academic-integrity case. Its current guidance reports that, between 2023 and 2025, 33% of review-board cases involving alleged inappropriate AI use ended with a “not responsible” finding when AI detection had been submitted without other supporting evidence.
WSU’s figure is institution-specific, not a national false-positive rate. It shows why a detector score and a misconduct finding cannot be treated as equivalent.
Can AI detectors catch edited, paraphrased or partly AI-written text?
Real people rarely use generative AI in a perfectly clean experimental way. A student might generate one paragraph and write the rest. Someone may rewrite sentences, add sources, remove awkward phrases or combine their own notes with AI-generated prose. That produces a mixed-authorship document rather than a pure AI sample.
Several 2026 studies point to this as a weak spot.
The 192-text Turnitin/Originality.ai study found poor performance on hybrid writing. A separate Computers & Education study published in 2026 evaluated 13 detectors across academic assignments, theses and code and reported systematic failures across tasks and disciplines. It also found that hybrid adversarial editing allowed a large share of AI-generated material to evade detection.
Other current work is more optimistic about particular tools. The higher-education comparison involving Pangram found that Pangram performed well on its hybrid and humanized samples while the other tools were much less consistent.
Hybrid text is not necessarily undetectable, but performance depends heavily on the tool and the exact transformation. A result from one detector cannot be assumed to describe every other detector, model or kind of editing.
Two different questions are often collapsed here:
- Can a detector recognize statistical patterns associated with AI-generated text?
- Can it reconstruct the actual writing history of this document?
The first is what detectors attempt. The second is much closer to what a disciplinary accusation requires.
What AI detector do colleges use?
There is no single AI detector used by every college.
Turnitin is a prominent option in higher education, and current institutional policies make its role concrete. San José State University gives faculty access to Turnitin and explicitly limits how its AI-detection reports may be used. Washington State University still uses Turnitin for plagiarism checking but cancelled its AI Detection contract in February 2026.
Policies vary sharply. SJSU says AI-detection reports can only be considered alongside other evidence and cannot be the sole basis for an academic sanction. WSU says it does not endorse AI detectors for academic-integrity decisions and continues to prohibit using detector output as the sole support for a case.
GPTZero, Copyleaks, Pangram and Originality.ai also appear in current academic research. Their presence in a study does not tell you what a particular institution has licensed or enabled. The 2026 comparisons are a reminder that “which detector?” is not a trivial detail: different tools can produce very different results on the same type of writing.
If you want to know what your own college or teacher uses, the reliable sources are the course policy, academic-integrity guidance and the submission system you are actually using. A generic web list cannot tell you whether a feature is licensed or enabled for your class.
Can a detector prove you used ChatGPT?
No. Not by itself.
A text detector does not receive a cryptographic receipt from ChatGPT. It does not see your browser history or the sequence of edits that produced the file. It analyzes the submitted text and estimates how much of it resembles patterns associated with machine-generated writing.
Even a strong classifier cannot reconstruct a document’s writing history.
A high AI score does not prove which model generated the text. It does not establish whether AI use was allowed under the assignment rules. It does not show whether a student generated the whole document, used AI for one section, used an AI-powered editing tool, or wrote the piece independently and happened to trigger the classifier.
Turnitin’s own guidance describes its report as one data point rather than a definitive answer. That is a sensible standard for any detector, not just Turnitin.
For high-stakes decisions, more informative evidence can include version history, drafts and notes, earlier writing by the same student, the sources and citations used, whether the writer can explain the argument and research process, and the AI policy that applied to the assignment.
These forms of evidence are imperfect, but they address the writing process more directly than a single classifier score.
How to interpret an AI-detector result without overclaiming
If you are looking at an AI-detector result, the most useful first question is not “Is the score high?” It is “What claim does this score actually support?”
A high score can justify closer review. It cannot, on its own, establish authorship or misconduct.
A low score does not prove that no AI was used. Current research shows that detectors can miss AI-generated material, especially depending on the model, detector and amount of human editing.
A very small flag should be treated especially cautiously. Turnitin suppresses exact scores in its 1–19% range specifically because of false-positive concerns.
Two detectors disagreeing is not an invitation to average the numbers. It is evidence that the classification is model-dependent. The tools may use different training data, thresholds and definitions of what counts as likely AI-generated.
For educators, the strongest next step is usually to look for process evidence and apply the institution’s policy. For students whose work is questioned, drafts, notes, document history and the ability to explain the work can be more relevant than running the same text through a collection of public detectors.
AI-text detection is useful when it is treated as a probabilistic signal. It becomes much less defensible when a probability score is turned into a claim about a person’s behavior without the evidence needed to bridge that gap.
Sources and verification
- Rosenthal et al. “Interpreting AI text detection outputs: Empirical evaluation and a threshold-based framework.” Studies in Educational Evaluation, 2026.
- “Who wrote this? Evaluating the reliability of AI detection tools in higher education.” International Journal for Educational Integrity, 2026.
- Hadra, Cambridge & Mesbah. “Evaluating the accuracy and reliability of AI content detectors in academic contexts.” International Journal for Educational Integrity, 2026.
- Sun, Liao & Ma. “Trusting AI to detect AI? A systematic evaluation of the reliability and robustness of current AIGC detection tools for student academic work.” Computers & Education, 2026.
- Turnitin. “How should I review the AI Writing report?”
- Turnitin. “AI writing detection model.”
- Washington State University. “Detecting and Reporting Misconduct Related to Generative AI.”
- San José State University. “Academic Integrity Policy — Campus Updates for Faculty.”