Are AI Detectors Accurate? False Positives Explained

Quick answer: AI writing detectors can catch some AI-generated text, but they also falsely flag human writing and miss AI-written passages, and their reliability varies with the tool, the type of text and the threshold. The base rate fallacy makes flags less trustworthy than advertised accuracy suggests: a hypothetical detector with 95% sensitivity and a 2% false positive rate is right only about 49% of the time it flags, if 2% of essays are AI-written. A detector score is not proof of misconduct, and a Turnitin AI percentage is not the probability that a student cheated.

Facts checked November 22, 2026.

Students have been accused of cheating because software said their essay “looks AI-generated.” How much should a teacher trust that flag? The answer depends less on the detector’s advertised accuracy than on a piece of probability that trips up even experts: the base rate fallacy. Here is the math, with numbers you can check.

How Style-Based AI Detectors Work

Most AI detectors estimate how predictable a text is to a language model, measures such as perplexity and “burstiness” (variation between sentences). AI text tends to be smooth and predictable; human text is often more varied. But plenty of human writing is also predictable, especially simple, formal or second-language writing, and that is where false positives come from. This is different from a watermark, which is a secret signal deliberately embedded by the AI company.

Three Numbers You Need

TermMeaningExample value
Sensitivity (true positive rate)Share of AI essays the detector flags95%
False positive rateShare of human essays wrongly flagged2%
Base rateShare of essays that are actually AI-writtenVaries by class

The Base Rate Fallacy, Worked Out

Imagine 1,000 essays. Bayes’ theorem asks: given a flag, how likely is the essay AI-written?

Base rate of AI essaysAI essays flaggedHuman essays flaggedChance a flagged essay is AI
10% (100 AI, 900 human)\(0.95 \times 100 = 95\)\(0.02 \times 900 = 18\)\(\frac{95}{95 + 18} \approx 84\%\)
2% (20 AI, 980 human)\(0.95 \times 20 = 19\)\(0.02 \times 980 \approx 19.6\)\(\frac{19}{19 + 19.6} \approx 49\%\)

Same detector, very different meaning. When cheating is rare, most flags can be false alarms, because the large number of honest essays produces many false positives. In formula form:

\[ P(\text{AI} \mid \text{flag}) = \frac{P(\text{flag} \mid \text{AI}) \, P(\text{AI})}{P(\text{flag} \mid \text{AI}) \, P(\text{AI}) + P(\text{flag} \mid \text{human}) \, P(\text{human})}. \]

These are hypothetical numbers, not measured performance of any product, and the formula assumes a clear yes-or-no task; essays that mix human and AI writing make things murkier. This is the same math as a medical test for a rare disease (see Bayes’ theorem explained).

Small False Positive Rates Add Up

A detector with a 1% false positive rate sounds safe. But in a class of 30 students who all wrote their own essays, the chance that at least one is wrongly flagged is

\[ 1 – 0.99^{30} \approx 0.26, \]

about 26%, assuming each essay is classified independently. Across a school year with many assignments, wrongful flags become almost certain somewhere. At a 2% rate, \(1 – 0.98^{30} \approx 45\%\).

Is a Turnitin AI Score the Probability of Cheating?

No. Turnitin’s AI writing percentage estimates how much of the qualifying text its model classifies as likely AI-generated; it is not the probability that a student used AI, and whether any AI use broke the rules is a separate policy question. To reduce false positives, Turnitin does not show a numerical score or highlights when its model estimates between 1% and 19% AI text; it displays an asterisk instead. Turnitin itself advises against using detection as the sole basis for action against a student.

What the Research Shows

  • Bias against non-native writers: a 2023 Stanford study, which did not test Turnitin, by Weixin Liang, James Zou and colleagues (published in Patterns) tested seven popular detectors on 91 human-written TOEFL essays. The average false positive rate across the seven detectors was 61.22%; 89 of the 91 essays (97.8%) were flagged as AI by at least one detector, while essays by US eighth graders were classified almost perfectly. Asking ChatGPT to enrich the wording of the human essays cut the false positive rate to 11.77%.
  • OpenAI withdrew its own detector (historical): in early 2023, OpenAI’s AI text classifier correctly identified only 26% of AI-written text while mislabeling 9% of human text; the company withdrew it on July 20, 2023, citing its low accuracy.
  • Easy to evade: light paraphrasing often fools style-based detectors, so they miss determined cheaters while still flagging honest writers.

Falsely Accused of Using AI? What to Do

  1. Stay calm and ask which policy applies and what evidence is being used.
  2. Gather your draft history, version history (for example in Google Docs or Word), notes, outlines, sources and timestamps.
  3. Offer to explain your work, reasoning and sources in a meeting.
  4. Point out the known limits of detectors, including the base rate problem and documented bias.
  5. Do not keep rewriting authentic work just to satisfy a detector.

What Teachers Can Do

  1. Treat a detector score as weak evidence, never proof.
  2. Use process evidence: drafts, version history, notes and a short conversation about the work.
  3. Design assignments that show thinking, such as in-class writing and oral explanations.

For broader guidance, see is using AI for homework cheating?

Common Misconceptions

“99% accurate means a flag is 99% reliable.” No. Reliability of a flag depends on the base rate, as the table shows.

“Detectors can prove AI use.” They estimate style; only watermarks or process evidence give stronger signals.

“Good writers never get flagged.” Clear, formulaic and second-language writing are among the most often misflagged.

Watch: Base Rates and False Positives

If you learn better by watching, these videos are worth your time.

The medical test paradox, and redesigning Bayes’ rule (3Blue1Brown). The base rate fallacy explained visually, the same math as AI detectors.
False Positives & Negatives | Using Bayes’ Theorem to Estimate Probabilities (Dr. Trefor Bazett). Worked examples of false positives with Bayes’ theorem.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. With 95% sensitivity, 2% false positives and a 10% base rate, what share of flagged essays are AI?
Found a bug? Let us know!
2. Same detector, base rate 2%. What share of flagged essays are AI?
Found a bug? Let us know!
3. A 1% false positive rate is applied to 30 honest students. Chance at least one is flagged?
Found a bug? Let us know!
4. In the 2023 Stanford study, what was the average false positive rate on TOEFL essays?
Found a bug? Let us know!
5. A Turnitin report says 40% AI. What does that number mean?
Found a bug? Let us know!

Frequently Asked Questions

Are AI detectors accurate?

Not reliably enough to prove cheating. They estimate style, produce false positives, are biased against non-native English writers, and can be fooled by paraphrasing.

Why do AI detectors give false positives?

They flag text that is predictable or uniform, which also describes much human writing, especially formal or second-language writing. When AI use is rare, even a small false positive rate means many flags are wrong.

What is the base rate fallacy?

Ignoring how common something is when interpreting a test. A detector with 95% sensitivity and 2% false positives is right only about half the time when it flags, if just 2% of essays are AI-written.

Can AI detectors be biased?

Yes. A 2023 Stanford study found seven detectors flagged an average of 61% of human-written TOEFL essays by non-native writers as AI-generated.

What does a Turnitin AI score mean?

It estimates the share of qualifying text that Turnitin’s model classifies as likely AI-generated, not the probability a student cheated. Scores from 1% to 19% are shown only as an asterisk to reduce false positives.

What should students do if falsely flagged?

Share drafts, notes and document version history, and offer to explain the work in person; process evidence is far stronger than a detector score.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "Are AI Detectors Accurate? False Positives Explained - Effortless Math"?

No one replied yet.

Leave a Reply