AI detectors: what the evidence says
Tools such as Turnitin, GPTZero and Pangram estimate whether text was written by AI. Studies from the past year show the tools are improving but still fall short of the standard needed for misconduct decisions.
- Newer detectors rarely flag fully human text but struggle with edited or disguised AI text.
- A score can start a conversation; it cannot decide a case.
- Redesign assessment and collect process evidence instead.
All research and institutional sources on this page date from November 2025 to November 2026.
Bottom line
Newer detectors rarely flag fully human writing. They still miss disguised AI text and misread lightly edited work, which is common in courses that allow some AI use. Use a score as a reason to talk with the student, and decide the case on other evidence.
3 of 4
detectors produced no false positives on fully human-written papers, a clear improvement on earlier research. An October 2026 correction adjusted some detection-rate figures.
Van Vlasselaer et al. (2026) ↗~10%
of AI text written to imitate real authors was missed by Pangram and GPTZero, even though both rarely flagged human writing.
Epoch AI (July 2026) ↗64–80%
of human-written abstracts lightly edited by an AI model were flagged by one leading detector.
Why AI Detection Fails (2026 preprint) ↗>96%
of AI rewrites passed through a commercial “humanizer” went undetected by the two detectors tested.
Why AI Detection Fails (2026 preprint) ↗What recent research shows
Detectors now rarely mislabel fully human text. Mixed, edited and disguised text still trips them up.
Van Vlasselaer, Van Droogenbroeck & Spruyt (2026)
International Journal for Educational Integrity · correction published October 2026
Compared Turnitin, Copyleaks, GPTZero and Pangram on 160 papers of known origin: human, AI, hybrid and “humanized.” Three tools had zero false positives on human papers, but detection rates dropped considerably for hybrid and humanized text.
Read the study ↗Epoch AI (July 2026)
Independent evaluation
Tested Pangram, GPTZero and Originality.ai. On AI text from basic prompts, misses were near zero, but roughly one in ten AI passages imitating a real author’s style got through.
Read the evaluation ↗Why AI Detection Fails for Academic Integrity (2026)
arXiv preprint, not yet peer reviewed
Leading detectors flagged some original human abstracts, flagged most lightly AI-edited human text, and missed nearly all text run through a humanizer tool.
Read the preprint ↗What a score can’t tell you
World Assessment Council (2026)
A low false-positive rate still cannot tell an instructor how likely one flagged paper is to be AI-written. And where a course allows AI for drafting or editing, a high score does not show misconduct.
Read the analysis ↗How institutions have responded
Three stances appear in guidance published over the past year.
Disable the detector
Curtin University switched off Turnitin’s AI detection across all campuses from 1 January 2026, keeping text-matching originality checks and framing the move around trust and fair assessment.
Scores never decide alone
Where detection is allowed, a score is one piece of evidence. Dartmouth’s conduct committee looks for other evidence before reaching a finding.
Allowed under strict conditions
Dartmouth’s Arts and Sciences faculty (Sept 2026) authorized only Pangram. Faculty must tell students in writing if their work may be checked, and work is submitted de-identified.
Base rates: why a flag is weaker evidence than it looks
Accuracy on a test dataset is not proof about one student. How often a flag is wrong also depends on how common misuse is in the class.
1,000 essays; 100 (10%) used AI without permission. The detector catches 90% of AI essays and wrongly flags 1% of human ones.
Result: 90 correct flags and 9 wrong flags.
About 1 in 11 flags is wrong.
Same detector, but only 20 essays (2%) used AI without permission.
Result: 18 correct flags and 10 wrong flags.
About 1 in 3 flags is wrong.
Illustrative numbers, not results from a study. Real error rates vary by tool, writing style and how much text was edited (Van Vlasselaer et al., 2026).
Due process for students
If an institution allows detectors, a flag should start a fair process. These steps draw on Dartmouth’s guidelines and standard academic integrity procedure.
- 1. Notice in advance Students are told in writing, before submitting, if their work may be checked by a detector. Dartmouth guidelines ↗
- 2. See the evidence A flagged student sees the score and any other evidence before being asked to respond.
- 3. A chance to respond The student can explain how the work was made, using drafts, notes or version history.
- 4. Human review of all the evidence A person weighs the full picture. A score is never the sole basis for a finding. Dartmouth guidelines ↗
- 5. Institutional procedure and appeal Cases go through the academic integrity process, including any right to appeal.
Why it matters for teaching & learning
Allowed use can get flagged. If a course allows AI for editing, a detector may flag work the policy permits.
2026 preprint ↗Disguised AI text gets through. Humanizer tools and style imitation defeat detectors, so students who try to hide AI use are the least likely to be caught.
Epoch AI (2026) ↗Trust suffers. Curtin cited trust and fair assessment when it turned detection off.
Curtin University ↗What to do instead
Redesign assessment. Use in-class, oral or staged tasks, and set an AI level for each assignment.
Policy builder ↗Collect process evidence. Drafts, version history and a short conversation show how the work was made.
World Assessment Council ↗If detectors are allowed, tell students in writing beforehand, use only approved tools, and never act on a score alone.
Dartmouth guidelines ↗Detector accuracy shifts with each model update. Vendor figures often differ from independent tests.