How accurate are AI detectors? What independent studies found

Updated October 31, 2026 · 5 min read

What have independent studies found?

Newest first. Detectors change every few months, so each result describes the tool versions tested at the time.

How often do detectors flag human writing?

• Rarely, for the leading 2025 to 2026 tools on ordinary text: 0 of 495 passages for Pangram and GPTZero (Epoch AI, 2026); 0 to 0.7% (Chicago Booth, 2025); 1.36 to 1.84% on web articles (Graphite, 2026).
• More often on short text. On passages under 50 words, false positive rates reached 3.3% for GPTZero and 2.5% for Pangram in the Chicago Booth study.
• More often on AI-touched human text. Lightly AI-edited abstracts were flagged 38 to 80% of the time in the Notre Dame study. Whether that counts as a false positive depends on whether the editing was allowed.
• Much more often with weaker tools. The open-source RoBERTa detector flagged 90 to 98% of human passages. In 2023, one tool flagged half of the human documents in a test.
• Turnitin's own figure is under 1% at the document level for documents with over 20% AI writing. That's about 1 in 100 fully human documents, by its own description.

Non-native English writers face extra risk with some tools. See the research on that.

How often do detectors miss AI text?

• Plain AI text, normal length: almost never with the leading tools (at most 0.7% missed in Epoch AI's test).
• AI imitating a writer's style: about 10 to 18% missed (Epoch AI, 2026). Gemini 3.1 Pro's scientific writing in an imitated style evaded Pangram 48% of the time.
• Paraphrased or humanized AI: anywhere from a few percent to almost all, depending on the tool and the humanizer. Details on paraphrasing and humanizers.
• Long, research-style AI papers: in the 2026 Brussels study, Turnitin, GPTZero and Copyleaks scored most fully AI papers as human. Pangram caught nearly all of them (97.5%, counting partial detections).
• Newer or less common models can be harder for some tools. Results by AI model.

Why do vendor accuracy claims look better than independent tests?

Vendors test on their own benchmarks. GPTZero says its 4o model catches over 99% of text from GPT-5.6, Gemini 3.6, Grok 4.5 and Claude 5, with a false positive rate under 0.03%. Pangram reports a 0.34% miss rate across 26 models. These numbers may be accurate for the text they chose. Independent tests use different texts, especially edited, imitated and humanized ones, and find more errors. Vendors' benchmarks also rank each other differently on humanized text, which is another reason to rely on independent studies.

Our interest: AI Checker Plus is an AI detector, so we have a stake in this topic. We report each study's numbers as published. None of these studies tested AI Checker Plus.

What does this mean if you use a detector?

• Use texts of at least a few hundred words; results on short passages are much less reliable.
• A high score on long, unedited text is strong evidence. A middling score on edited text is weak evidence.
• A low score doesn't prove a human wrote it, especially if the text may have been humanized or written to imitate someone.
• Read the highlighted passages and compare them with drafts and other work.
• Don't treat any single result as proof. A fair process for teachers.

By the AI Checker Plus team. We research and draft these guides with help from AI tools, then check every claim against the sources listed above.