Do AI Content Detectors Actually Work?

This question matters to two groups with opposite fears: students and writers worried about being falsely accused, and teachers and editors trying to know what they are reading. The honest answer disappoints both, and understanding why is more useful than any tool recommendation.

How detectors claim to work

Most detectors look at statistical properties of text rather than meaning. Two ideas come up repeatedly:

• Perplexity — roughly, how surprising each word is given the ones before it. Language models tend to choose likely continuations, so their output can be less "surprising" than human writing.
• Burstiness — variation in sentence length and complexity. Human writing tends to swing between long and short sentences more than generated text does.

Both measure regularity. That is the root of the problem, because regularity is not the same as machine authorship.

Why they produce false positives

Writing that is clear, well-structured and consistent scores as machine-like — which describes a great deal of good human writing.

The groups most affected are predictable once you see the mechanism:

• People writing in a second language. Learners often use simpler sentence structures and a narrower vocabulary, producing exactly the low-variation profile detectors flag. A Stanford study published in 2023, GPT detectors are biased against non-native English writers, found detectors misclassified non-native writing samples at dramatically higher rates than native ones — while classifying native samples almost perfectly. That makes detector use a fairness problem, not just an accuracy one.
• Technical and academic writing. Formal registers reward consistency and discourage stylistic flourish.
• Anyone who edits carefully. Tightening prose removes irregularity — the same irregularity detectors read as human.
• Formulaic formats. Lab reports, legal summaries and structured documentation are regular by design.

Widely circulated examples include detectors flagging historical texts and religious documents written long before language models existed. Those demonstrations are unfair to the tools in one sense — they are out of distribution — but they illustrate the point exactly: the tools measure regularity, and regular writing predates AI by centuries.

Why they produce false negatives

The other failure is quieter and harder to demonstrate. Detectors are easily defeated, often without anyone trying:

• Ordinary editing. Rewriting a few sentences, varying length, adding a specific detail — normal revision — moves text out of the flagged range.
• Prompting for a different style. Asking a model to write more conversationally changes the statistical profile.
• Mixing. Text that is part written and part generated tends to score inconclusively.

So the tools are most likely to catch the least sophisticated use, and least likely to catch anything deliberately concealed. That is close to the opposite of a useful detector.

Reading a detector score honestly

Two things are commonly misunderstood about the number a detector returns.

"95% AI" is not a probability that the text is AI-generated. It is a similarity score against the tool's model of machine-like text. It does not mean there is a 95% chance of machine authorship, and treating it that way is a statistical error with real consequences for the person accused.

Base rates matter enormously. Even a detector with a low false-positive rate produces many false accusations when most submissions are human-written. Applied to a class where the great majority wrote their own work, a 5% false-positive rate on a hundred students means several innocent people flagged — likely outnumbering any genuine cases.

This is the same reasoning that applies to any screening test for a rare condition, and it is why "the tool said so" is never sufficient grounds for a decision about a person.

If you have been wrongly accused

Being flagged when you wrote something yourself is genuinely distressing. What actually helps:

1. Show your process. Version history in Google Docs or Word, drafts, notes, search history, saved sources. A document with hours of incremental edits is strong evidence of authorship; a document pasted in complete is not.
2. Offer to discuss the work. Someone who wrote a piece can explain why they structured it that way, what they cut, and where they were uncertain. This is the most convincing evidence available and cannot be faked from a generated text.
3. Ask what the score actually means. Politely: what is the tool's documented false-positive rate, and is a score alone considered sufficient? Many institutions have policies stating it is not.
4. Do not run it through another detector as proof. A second tool disagreeing shows the tools disagree — which supports the general point, but is not evidence about your specific document.

Prevention is simple and worth the habit: write in something that keeps version history, and keep your notes. It costs nothing and turns an unanswerable accusation into a resolvable one.

If you are checking someone else's work

A detector score is a reason to look more closely, never a conclusion. What is more informative:

• Check the citations. Language models fabricate references that look correct — right format, plausible journal, author who exists. A reference that cannot be found in any database is far stronger evidence than any detector score.
• Look for specificity. Generated text tends toward the general. Concrete details, personal observations and engagement with the particular assignment are hard to fake and easy to check.
• Ask about it. A short conversation about the argument reveals authorship more reliably than any tool.
• Compare with earlier work. A sudden change in voice or capability is a genuine signal, though it can also mean someone simply improved or got help.

Where this leaves you

The technology is not converging on a solution. As models improve, their output becomes less statistically distinguishable from human writing — so detector accuracy tends to degrade over time rather than improve. Several major providers have withdrawn or quietly deprecated their own detection tools after finding accuracy insufficient, which is a meaningful signal about the difficulty.

The practical consequence for institutions is that policies leaning on detector scores are building on sand, and many are shifting toward process evidence — drafts, version history, oral discussion — instead. For individuals, the takeaway is simpler: keep your working history, and be able to talk about what you wrote.

Worth noting separately, because the two questions get conflated: Google's published guidance on AI-generated content does not penalise AI assistance as such. It rewards useful content and penalises content produced primarily to manipulate rankings, however it was written. If you are worried about detectors for SEO reasons rather than academic ones, that is the policy that actually applies.

On why those fabricated citations appear at all, why AI makes things up explains the mechanism — and why a confident tone tells you nothing about whether a claim is true.

If your concern is search rather than academic honesty, that is a separate question with a clearer answer: does AI content rank on Google? covers what the published policy actually says.

Frequently asked questions

Do AI content detectors actually work?

Not reliably. They produce false positives on genuine human writing — particularly from people writing in a second language — and miss AI text that has been lightly edited. They measure statistical regularity, which is not the same as machine authorship.

Can a detector prove I used AI?

No. A detector score is a similarity measure against a model of machine-like text, not proof of how a document was produced. Most institutional policies acknowledge this, and a score alone is generally not considered sufficient grounds for an accusation.

What should I do if I am falsely accused of using AI?

Show your process: version history, drafts, notes and sources. Offer to discuss the work — being able to explain your structure, your cuts and your uncertainties is the most convincing evidence there is. Ask what the tool's documented false-positive rate is.

Why do detectors flag non-native English writers more often?

Because they measure variation in vocabulary and sentence structure. Writing in a second language often uses simpler, more consistent constructions, which produces the same low-variation signal detectors associate with generated text. This makes detector use a fairness problem as well as an accuracy one.

Does "95% AI" mean it is 95% likely to be AI?

No. It is a similarity score against the tool's internal model, not a probability of authorship. Interpreting it as a probability is a statistical error, and one with real consequences when someone is accused on that basis.

Can I avoid detection by editing AI text?

Ordinary editing does change the statistical profile enough to affect scores, which is precisely why detectors are unreliable. Whether you should is a different question — your institution or client almost certainly has a disclosure policy, and that is the standard that actually applies to you.

Are there any reliable ways to identify AI writing?

Nothing reliable from the text alone. Fabricated citations that cannot be found in any database are the strongest practical signal, followed by an absence of specific, checkable detail. A conversation about the work remains the most dependable method.

Conclusion

Detectors measure regularity and call it authorship. That mistake produces both false accusations against careful writers and false confidence about text that was generated and tidied.

If you write: keep your drafts and version history. If you assess: check the citations, look for specificity, and talk to the person. Neither of those depends on a tool that cannot do what it claims.

Related: our guide to free AI tools for students covers using these tools without importing fabricated citations, writing better AI prompts covers getting genuinely useful output rather than generic text, and Claude vs ChatGPT vs Gemini compares the assistants themselves.