Do AI Content Detectors Work? What They Actually Measure

Short answer: no, not in the way the name implies. A detector cannot determine how a piece of text was produced. It measures how statistically predictable the wording is, and predictable writing is common in both machine drafts and careful human prose — which is why technical documentation, non-native English and plain instructional writing are frequently flagged. The scores are probabilistic, vary between tools on the same text, and have no standing with search engines, which have said repeatedly that they judge usefulness rather than authorship. Treat a detector score as noise and spend the time on accuracy instead.

The question behind this one is usually practical: will publishing machine-assisted content get my site penalised, and can anyone tell? The detector part of that has a clear answer.

These tools do not detect authorship. They estimate predictability, and they then present that estimate as a percentage, which invites people to read it as a verdict.

What a detector actually computes

Language models assign probabilities to what word comes next. Detectors run text through a model and ask how surprising each choice was.

Text where nearly every word is the expected one scores as likely machine-written. Text with unusual constructions, abrupt shifts and idiosyncratic word choices scores as likely human. That is the whole mechanism, dressed in a percentage.

The consequence follows immediately. Any writing that is deliberately plain and conventional — a manual, a recipe, a legal summary, an article written to be easy to follow — looks machine-written by this measure. And any generated text that has been edited for rhythm looks human.

Who gets wrongly flagged

The false positives are not random; they fall on predictable groups.

• People writing in a second language, who tend to use common constructions carefully. This is the best documented bias in the research on these tools.
• Technical writers, whose job is to be unsurprising.
• Anyone following a style guide, which is what a style guide is for.
• Short texts. Under a few hundred words the estimate is close to worthless.
• Formulaic passages, such as a paragraph of definitions, inside an otherwise original article.

Run the same paragraph through three detectors and you will often get three materially different scores. That alone tells you what kind of instrument this is.

What search engines have said

This is where the real question sits, and the published position is not ambiguous: what matters is whether content is helpful, original and made for people, not how it was produced. Automation used primarily to manipulate rankings is against the guidelines; automation used to produce useful pages is not.

There is also no evidence that search engines run consumer detectors, and good reason to think they would not, given the error rates. what search engines actually say about AI content goes through the guidance in detail, including what it does say you should worry about.

The practical translation: the thing that gets a page demoted is being thin, duplicated, inaccurate or pointless. Those are properties of the page, and you can assess all of them yourself without a tool.

When a score still tells you something

A high machine-likelihood score is not a verdict, but it is sometimes a symptom worth reading.

If a draft scores as highly predictable and you read it and find no specifics, no position, no examples and nothing a reader could not get from the first result they clicked, the score and your judgement are pointing at the same thing. The problem is not detectability; it is that the article says nothing.

Used that way — as a weak prompt to reread — a detector is harmless. Used as a gate that content must pass, it makes you rewrite good sentences into worse ones.

Do not write to defeat detectors

There is an industry of advice about humanising text, and following it makes pages worse.

The last two are the only items on that list with a return, and neither has anything to do with detection.

If a client or platform requires a detector score

Some agencies and marketplaces do require one, and arguing about methodology rarely helps.

Two things make it manageable. First, ask what the requirement is protecting against, and offer to meet that directly: a named author who stands behind the piece, sources for every claim, and original material the client can verify. That is a stronger guarantee than any percentage.

Second, if a number is non-negotiable, meet it by adding substance — your own data, quotes, specific examples, a real opinion — rather than by scrambling the prose. Genuine original content scores lower on predictability anyway, and unlike the alternative it also makes the page worth reading.

The disclosure question, separately

Whether to tell readers that content is produced with automation is a different question from detection, and it is a real one.

Search guidelines do not require a disclosure. Some jurisdictions and some industries do, and some audiences care a great deal. The defensible position is to be accountable rather than coy: a named author or organisation stands behind the page, claims are checkable, and you would not be embarrassed to explain your process if asked.

That is also what experience, expertise and trust is about in practice, and it is a far better use of attention than a percentage that changes depending on which tool you paste into.

Related reading

If this was useful, these cover the questions that usually come next.

• What search engines actually say — the guidance these tools are often confused with
• Editing a draft — what to do with the editing time instead
• Experience, expertise and trust — the signals that genuinely matter

The bottom line

Detectors measure predictability, not authorship, and they misfire on exactly the writing that is clearest. No search engine has said it uses them, and the published guidance is about whether a page is useful. Ignore the score, check your facts, add something only you know, and be willing to say who stands behind the page.

FAQ

Can AI detectors tell if content was written by a machine?

No. They estimate how statistically predictable the wording is. Predictable writing is common in both generated text and careful human prose, so the output is a probability with a substantial error rate rather than a determination.

Does Google use AI detectors to rank pages?

There is no indication that it does, and the published guidance is about whether content is helpful and original rather than how it was produced. Automation aimed at manipulating rankings is against the rules; automation used to make useful pages is not.

Why did my own writing get flagged as AI?

Because it was clear and conventional. Technical writing, anything following a style guide, short passages and English written by a second-language speaker are all flagged disproportionately.

Should I rewrite text to lower a detector score?

Not by adding unusual phrasing or errors, which makes the page worse. If you must move the number, add original material: your own data, specific examples, a clear opinion. That lowers predictability and improves the article at the same time.

Do different detectors agree with each other?

Frequently not. The same paragraph can score very differently across tools, which is a good reason to treat any single score as noise rather than evidence.

Do I have to tell readers that I use automation?

Search guidelines do not require it, but some jurisdictions, industries and audiences do care. The safer position is accountability: a named author or organisation behind the page, checkable claims, and a process you would be comfortable describing.