Do AI detectors actually work?

Conventional AI detectors guess from writing style, and the evidence on their accuracy is not good. Here is what they measure, where they fail, and why a negative result proves less than people assume.

By didaiwritethis.ai

The short answer

Not well enough to accuse anyone. Style-based AI detectors produce a confident-looking percentage from a genuinely uncertain inference, and the ways they fail are not random — they fall hardest on specific groups of writers. That combination, high apparent confidence and structured failure, is what makes them dangerous rather than merely imperfect.

What these tools actually measure

Almost every conventional detector works from the same basic idea: language models tend to pick predictable words. Given everything written so far, a model usually selects a high-probability next token. Human writers are messier — they reach for an odd word, build a strange sentence, change direction mid-paragraph.

Detectors quantify that intuition with two related measures:

• Perplexity — roughly, how surprising the text is to a reference language model. Low perplexity means the words were easy to predict.
• Burstiness — how much that predictability varies across the piece. Human writing tends to swing between dense and simple passages; generated text is often flatter.

So a detector is not identifying an AI. It is measuring how closely the writing resembles statistically average prose, then reporting that resemblance as a probability. Those are not the same claim, and the gap between them is where the damage happens.

Where they fail

Writers who use plainer English

The most serious documented problem is bias against non-native English speakers. A 2023 Stanford study by Liang and colleagues ran essays written by non-native English speakers through several widely used GPT detectors and found they were misclassified as AI-generated at strikingly high rates, while essays by native speakers were classified correctly far more often. The mechanism is straightforward once you see what is being measured: writing with a more limited vocabulary and more conventional sentence construction scores as low-perplexity — which is exactly what detectors treat as evidence of a machine.

The same logic penalises anyone whose prose is deliberately plain. Technical documentation, legal boilerplate, formal academic register, non-fiction written to be clear rather than interesting — all of it drifts toward the profile these tools flag.

Editing defeats them

Paraphrasing generated text, even lightly, tends to raise its perplexity enough to move it under the threshold. Anyone deliberately trying to evade detection can usually do so with modest effort, which inverts the incentive: the tool is most likely to catch the person who did not think to disguise anything, and least likely to catch the one who did.

The vendors have conceded the point

OpenAI launched a classifier for AI-written text in early 2023 and withdrew it that July, citing its low rate of accuracy. The organisation with the deepest access to how the underlying models behave could not make style-based detection work reliably enough to keep offering it.

Why the accuracy numbers mislead

Vendors often advertise something like 99% accuracy. Even taken at face value, that figure does not mean what people read it to mean, because of how base rates work.

Suppose a class submits 10,000 essays and 1% are AI-written. A detector with a 1% false positive rate flags about 99 innocent essays. It also correctly flags most of the 100 real ones. So roughly half of everything it flags is a false accusation — from a tool that is, by its own marketing, 99% accurate. Push the AI-written share lower, and the proportion of flagged-but-innocent work climbs further.

This is not a criticism of any particular product. It is arithmetic that applies to any screening test for a relatively rare condition, and it is why a detector score should never function as the evidence in a misconduct case.

How watermark detection differs

Watermarking attacks the problem from the other end. Instead of inspecting finished text and guessing at its origin, the model embeds a deliberate statistical signal at the moment of generation. A detector holding the corresponding key measures whether that specific signal is present.

That is a materially stronger form of evidence: a positive result points to a specific model rather than to a resemblance. But it comes with a hard boundary that is easy to overstate away —

• It only finds watermarks in text from models that watermark. It cannot detect AI text in general, and never will.
• A negative result is weak evidence. The text may come from a non-watermarking model, from a watermarking model with the feature off, or from output edited enough to disturb the signal.
• Heavy paraphrasing degrades the signal, just as it defeats style-based detectors.

How watermarking actually works goes through the mechanism in more detail.

What to do instead

If you need to know whether a specific person wrote a specific thing, the text itself is usually the weakest available evidence. Stronger sources:

• Process artifacts. Document version history, draft timestamps, and commit logs are far harder to fabricate convincingly than prose is.
• A conversation. Someone who wrote a piece of work can generally discuss the choices behind it. Someone who did not, generally cannot.
• Watermark detection, where available — as one input among several, and weighted for the limits above.

And if you are on the receiving end of a detector accusation: ask what the tool measured, ask what its false positive rate is on writing like yours, and ask what evidence exists beyond the score. Often there is none.