Last updated: October 2026 · Reading time: about 12 minutes
"Our detector is 99% accurate." You will see claims like this on many websites. But are AI detectors accurate in the way that matters to a student whose grade is at risk, or a teacher who needs to be fair? The honest answer is: it depends on what you measure, on whose writing is tested, and on how common AI use actually is. In this guide we explain the key ideas, false positives, false negatives, and base rates, with a worked example so you can judge any accuracy claim for yourself.
Table of contents
- The two kinds of mistakes
- Why "99% accurate" can be misleading
- A worked example (hypothetical numbers)
- What independent research and educators say
- Seven questions to ask about any accuracy claim
- How to use detectors responsibly
- FAQ
The two kinds of mistakes
- False positive: a human-written text is labeled as AI. This is the mistake that harms honest students.
- False negative: an AI-generated text is labeled as human. This is the mistake that lets some AI use go unnoticed.
Every detector makes both kinds of mistakes, and there is a trade-off. A tool tuned to catch more AI text will usually flag more human text by mistake. A tool tuned to avoid false accusations will miss more AI text. That is why two honest providers can have different settings and report different numbers.
Why "99% accurate" can be misleading
A single accuracy percentage hides several things:
- What was tested. Was it long, unedited AI text compared with polished human articles? Real classroom writing is more varied: short, edited, mixed, or written by learners.
- Which errors are counted. An overall number can look excellent even if the false positive rate for a particular group is high.
- When it was tested. New AI models appear frequently, and results from an older test may not apply today.
- Who ran the test. Results published by the vendor are not the same as independent evaluations.
- The base rate. This one is the most overlooked, and we explain it next.
A worked example (hypothetical numbers)
To show how the base rate matters, imagine a class (or a school year) in which teachers check 1,000 essays. Suppose the following, purely as an illustration:
- 10% of the essays (100) really were generated by AI, and 90% (900) were written by students.
- The detector correctly flags 90% of the AI essays: 90 of 100 caught.
- The detector wrongly flags 5% of the human essays: 45 of 900 flagged by mistake.
On paper this sounds strong: 90% detection and only 5% false alarms. Now count the flags: 90 (correct) + 45 (mistaken) = 135 flagged essays. Of those, 45 out of 135, about one third, belong to innocent students.
| Flagged as AI | Not flagged | Total | |
|---|---|---|---|
| Actually AI | 90 | 10 | 100 |
| Actually human | 45 | 855 | 900 |
| Total | 135 | 865 | 1,000 |
These numbers are invented to explain the concept. They are not the performance of any real tool.
Now change one assumption: if only 2% of essays were AI-generated, and the same rates applied, the number of mistaken flags would stay large relative to correct ones, and a flagged student would be even more likely to be innocent. The lesson: a flag is a reason to investigate, not a conclusion, and the lower the real rate of misuse, the less a flag means by itself.
What independent research and educators say
- Bias concerns. Researchers at Stanford reported that several popular detectors misclassified many essays by non-native English writers as AI-generated. See our article on AI detectors and non-native English speakers.
- Educator caution. Academic integrity researchers, such as Dr. Sarah Elaine Eaton, have argued that AI-detection tools should not be used as the sole basis for judging student work.
- Real consequences. News reports, including Bloomberg's coverage, have described students facing serious consequences after being flagged.
- Vendor transparency. Many providers themselves state that results should be used with other evidence and are not proof.
Because tools and tests change, always look for recent, independent evaluations and read the provider's own limitations.
Seven questions to ask about any accuracy claim
- What kind of text was used (length, subject, edited or unedited)?
- What are the false positive and false negative rates separately?
- Were non-native English writers and short texts included?
- Who ran the test: the vendor or an independent group?
- When was it tested, and with which AI models?
- What threshold was used to call text "AI"?
- What does the provider say about how results should be used?
If a website cannot answer these, treat its accuracy number as marketing. You can use our method for comparing free AI essay detectors to run your own test.
How to use detectors responsibly
For students
- Use a checker for self-review and revise generic sentences in your own voice.
- Keep drafts, notes, and version history for every important assignment.
- If flagged, ask for the report and evidence, and respond calmly with documentation.
For teachers and schools
- Never use a score as the sole evidence.
- Combine multiple sources: drafts, earlier work, and a conversation with the student.
- Publish a clear AI policy. See our AI policy template.
- Be cautious with short texts and with second-language writers. Read how detectors behave on short essays.
Watch: accuracy explained
Frequently asked questions
So, are AI detectors accurate or not?
They can be informative in some conditions, such as longer, unedited text, but they are not accurate enough to serve as proof of cheating. They make both false positives and false negatives, and performance varies by text type and writer.
What is a good false positive rate?
There is no universal answer, but even a small percentage can mean many wrongly flagged students at scale. That is why a flag should never be treated as a verdict.
Can AI detectors detect edited or paraphrased AI text?
Edited and paraphrased text is much harder to classify, so detectors may miss it. This is one reason scores cannot be treated as complete evidence either way.
Which AI detector is the most accurate?
No tool is best on all texts. Look for independent evaluations, transparency about limitations, and consistency on your own samples.
Does a 0% AI score prove my work is human?
No. It only means that this tool did not flag the text. Another tool might, and the tool might be wrong.
Key takeaways
- Detectors make two kinds of errors, and reducing one tends to increase the other.
- A single accuracy number hides the false positive rate, the text type, and the base rate.
- In low-misuse settings, a surprisingly large share of flagged work can be human-written.
- Use scores to start a conversation, and decide using multiple forms of evidence.
Sources and further reading
- Liang et al., "GPT detectors are biased against non-native English writers".
- Dr. Sarah Elaine Eaton, "The Use of AI-Detection Tools in the Assessment of Student Work".
- Bloomberg, "AI Detectors Falsely Accuse Students of Cheating—With Big Consequences".
Disclaimer: The worked example uses invented numbers for teaching purposes only. This article is general information, not legal or academic advice.