Insights
Can AI detectors actually tell if text was written by AI?
- AI
- Content marketing
- Small business
The short answer
AI detectors estimate how statistically predictable a piece of text is, which is not the same as knowing who wrote it. They flag plain, formal and second-language writing as machine-generated often enough that a score should prompt a conversation, never a rejection. Humans and models simply do not leave different enough fingerprints.
No, not reliably. An AI detector produces a probability that a piece of text resembles the statistical patterns of machine-generated writing, and it will sometimes give that verdict to text a person typed from scratch. The score is closer to a smoke alarm than to a lab test: useful as a prompt to look, useless as a verdict.
How does an AI detector actually work?
A language model writes one token at a time. A token is a chunk of text, usually part of a word, and English averages roughly three quarters of a word per token. At every step the model assigns a probability to every token it could produce next and then picks one.
A detector runs the finished text back through a similar model and measures how predictable each choice was. Predictable wording gives low perplexity, which is simply the technical name for how surprised a model is. Text with low perplexity and little variation in sentence length is what detectors call machine-like.
Most commercial tools then feed those signals into a classifier. A classifier is a program trained on thousands of labelled examples, some written by people and some by models, until it learns to separate the two. It is guessing from style, not tracing an origin.
Why do AI detectors flag writing that people wrote?
Because the signals detectors measure are habits, and plenty of humans share the habits of a model. A person writing a product specification, a lease summary or a set of terms and conditions chooses predictable words on purpose. That is what clear, formal writing looks like, and it scores as AI.
Writers working in a second language are flagged more often, for the same reason. Textbook phrasing and common collocations are, by definition, the predictable choices. So is any sentence that has been through three rounds of editing by a committee.
Short text makes it worse. Under about 200 words there is not enough signal for the statistics to mean anything, and most tools will still return a confident-looking percentage.
How accurate are AI detectors in 2026?
Nobody publishes a clean error rate, and the honest general understanding is that these tools are unreliable at the level of a single document. Most comparisons that people run publicly put false positives on human writing somewhere in the range of 5 to 20 per cent, with wide disagreement between tools and between repeat runs of the same tool.
Two things follow from that. The same document can read 80 per cent AI on one tool and 15 per cent on another, and neither number is wrong in any meaningful sense. And a detector that catches most machine text while flagging one human document in ten is a bad basis for accusing a freelancer or failing a student.
Results also drift. When a model provider updates its model, the statistical fingerprint changes and the detectors retrain. A threshold that worked in January is not guaranteed to mean anything by the summer.
What does a score of 80 per cent AI actually mean?
The number is not a proportion of words. It is the classifier’s confidence, often an internal score mapped onto a 0 to 100 scale for display. A reading of 80 does not mean four sentences in five came from a model, and it does not mean 80 per cent of the text is unedited machine output.
Some tools now return bands such as likely human, mixed or likely AI instead of a number, precisely because buyers kept reading the percentage as a measurement. That is an improvement in honesty rather than in accuracy.
Can AI text be watermarked instead?
Watermarking is the more solid idea, and it is different from detection. When a model generates text it picks tokens at random according to their probabilities. A watermarking scheme splits the vocabulary into two lists using a secret key, tilts the model toward one list, and then a checker with the key counts how many of those tokens appear. Ordinary text should show about half; watermarked text shows more.
This works in the narrow sense that it can be verified rather than guessed. It fails in practice for three reasons: the provider has to switch it on, the checker has to hold the key, and any edit, translation or paraphrase destroys the pattern. Text pasted through a second tool to make it read more naturally loses the mark entirely.
Can you bring an AI score down by editing the text?
Yes, and that is part of the problem. Detectors reward specificity and variation. Adding a named customer, an exact figure, a date and an odd but true detail reliably moves a score down, whether a person or a model wrote the rest of the piece.
A writer who drafts in their own words and then asks a model to tighten four sentences can end up with a document that scores 60 per cent AI, because the tightened sentences are the predictable ones. Rewriting them to satisfy the tool makes the copy slightly worse and teaches nobody anything.
What does an AI detector cost?
Single-seat subscriptions typically run from about £8 to £40 a month, and bulk checking through an API is usually a fraction of a penny per hundred words. Institutional contracts processing thousands of documents sit in the thousands of pounds a year.
The bigger cost is the decision the score drives. Rejecting a good writer over a 62 per cent reading, or paying an agency a premium for copy certified as human, costs far more than the subscription. That premium is typically 20 to 40 per cent on top of a normal rate, and no supplier can guarantee what a third-party tool will say about the text next month.
What should a business do instead?
- Brief for specifics. Ask for named examples, exact numbers and first-hand detail, because machine output rarely volunteers them unprompted.
- Ask for the draft history. A shared document with visible edits tells you more about provenance than any percentage.
- Put accuracy and originality in the contract, not detection. You are buying work that is true and not copied, and that is a testable standard.
- Check facts and check for copying. Those tools compare against a known body of text, which is a different and more defensible question than style.
- Say what you did. If a model produced the first pass and a person rewrote it, a one-line note is honest and increasingly normal.
When is an AI detector worth paying for?
Mainly at volume, where a human cannot read everything: a publisher screening submissions, a training provider checking coursework, an agency triaging hundreds of freelance samples. Even then the score should sort the pile rather than decide the outcome, and a second reader should confirm anything that matters.
For a business buying a few thousand words of copy a month, it is not worth the subscription. Read the work. Ask who wrote it and how. A five-minute conversation with a writer beats a percentage from a tool that cannot explain itself.
Keep reading
What happens to your website if your web developer disappears?
If your developer disappears you keep the site only if you control the domain, the DNS and the hosting. What to check now, and what recovery costs.
Read the articleWhat are the real costs of a website accessibility audit in 2026
A realistic price range for a website accessibility audit in 2026, what a typical report includes, and when a small business should do one.
Read the articleFind out how visible you actually are
A free written audit of how search engines and AI assistants currently read your website. No obligation, and no sales call required to receive it.
Prefer to talk first? Email contact@luiinteractive.com or message +44 7349 961542 on WhatsApp.