How accurate are AI detectors? Limitations explained

AI detectors are often presented as if they can tell with certainty whether a text was written by a person or a machine. They cannot. Every detector, including ours, makes an estimate based on patterns, and estimates can be wrong. Understanding why helps you use a detector sensibly and avoid drawing conclusions that the result does not support.

How AI detectors work

A detector looks at features of the text: how predictable the word choices are, how much sentence length and structure vary, and how often typical AI phrasing appears. From these signals it estimates how likely it is that the text was machine-generated.

The key word is likely. The detector has no access to how the text was produced. It compares the text with patterns, and a human can write in a pattern that resembles AI just as AI output can be edited to resemble human writing.

False positives: human text flagged as AI

A false positive happens when a person's own writing is marked as likely AI. This is the most harmful mistake, because it can lead to unfair suspicion. Formal, neutral and highly structured writing is more exposed, as is writing by people using a second language, who often choose simpler and more predictable phrasing.

False negatives: AI text that passes as human

The opposite also happens. AI text that has been rewritten, mixed with the writer's own sentences or produced with instructions to vary the style can receive a low score. A low score therefore does not prove that a person wrote the text either.

What makes a result less reliable

Several things increase uncertainty. Short texts contain few patterns to analyse. Mixed texts, where a person and a tool both contributed, sit between the categories by nature. Specialised genres such as legal or technical writing are often formulaic regardless of who wrote them. And new AI models change the patterns detectors rely on.

Be careful with accuracy claims

Some tools advertise very high accuracy figures. Such numbers depend heavily on which texts were tested, how long they were and which models produced them, and they rarely carry over to the texts you actually check. We do not publish an accuracy percentage, because a single number would suggest a certainty the method does not have.

How to read a result responsibly

Treat the score as an indication, not a verdict. Read the passages the detector highlights and judge for yourself whether the reasoning holds. Combine the result with other evidence, such as drafts, sources and the writer's earlier work. If an important decision depends on it, talk with the writer before concluding anything.

Frequently asked questions

Is any AI detector 100% accurate?

No. All AI detectors estimate probabilities from patterns in the text and can be wrong in both directions. Be sceptical of any tool that promises certainty.

Why was my own text flagged as AI?

Formal, structured or very predictable writing can resemble AI patterns. This is a known limitation called a false positive, and it does not mean you did anything wrong.

Can edited AI text avoid detection?

Sometimes. Rewritten or mixed text can get a low score, which is why a low score is not proof of human authorship either.