glossary · ai tell
The tells that give machine-written prose away
Not a detector score — a specific habit in the writing: the throat-clearing opener, the tidy triad, the hedge where a number belongs. What the tells are, what they cost, and how to remove them.
An AI tell is a recognisable habit of machine-written prose — a formulaic opener, symmetrical paragraphs, hedged non-answers — that signals to a reader the text was generated rather than written by someone who knows the subject.
reviewed 2026-09-02 · by the IT Master editorial team · how we check facts
What an AI tell looks like in practice
A tell is not a vibe. It is a specific, nameable habit that survives from draft to draft because the model learned it from a great deal of text that already contained it.
The recurring ones are easy to list.
- The throat-clearing opener. Two sentences establishing that the topic exists and matters, before the page says anything.
- The tidy triad. Every idea arriving in threes, every list balanced to the same length, whether or not the subject has three parts.
- The hedge where a number belongs. "Costs vary depending on several factors" standing in for a price, a range or a date.
- The restatement close. A last paragraph that summarises the introduction and adds nothing.
- Signature connectives. "It is important to note", "in conclusion", "not only, but also", "delve into".
- Flat specificity. Statements that are true in general and carry no example, part number or figure that could be wrong.
The last one is the expensive one. The others are surface habits an editor strikes out in a single pass. Flat specificity means the writer had nothing concrete to say, and no amount of rewriting inserts a fact that was never there.
The costliest AI tell is not a phrase. It is flat specificity: prose that is true in general and carries no example, part number or figure that could be wrong.
What a tell actually costs you
Tells cost you in two places, and neither of them is a detector score.
Readers. Someone who fits cameras for a living recognises hollow prose within a paragraph. They leave, and they learn not to click that domain again. That behaviour is what the helpful content guidance describes: pages that answer nothing get selected less over time.
Assistants. ChatGPT and Perplexity cite passages that state something checkable, a figure, a procedure, a limit. A hedged paragraph gives an assistant nothing to quote, so a page can be crawled, indexed and still never appear in an answer.
There is a commercial cost too. Prospects paste your case study into a checker. Editors at client sites bounce the copy. A trade buyer reads two paragraphs and concludes you do not actually do this work. None of that shows up in Search Console.
Note what is not on the list. Google publishes no per-URL tell score, and no ranking factor called AI-sounding exists. The mechanism is duller than that: text that says nothing gets ignored, by people first.
No search engine scores your page for sounding like a machine. Text that says nothing gets ignored by people first, and by the systems watching people second.
How the engine handles AI tells
AI-tell detection is one of four checks a draft must pass, alongside fact-check, novelty against the pages already ranking, and an E-E-A-T critique. Claude models write; Gemini and GPT judge, so the model that writes is never the model that checks. Failures come back as named fixes — this opener, this hedge, this closing paragraph — for up to three revision rounds. A draft that still fails is not published.
One honest limit. A judge model learned prose from the same sort of text as the writer, so it shares some of the habits it is asked to spot. That is the argument for a checker from a different vendor rather than a second pass by the author, and for never leaning on a single check.
A detector percentage is not the target either. Writing to satisfy a classifier produces choppier prose, not better prose.
What removes tells is having something to say. Where a customer holds first-party data — product records, datasheets, the questions support really receives — drafts are grounded in it, so the writer has a figure to put where a hedge would otherwise sit.
Measured, not promised. Across 186 Standard-tier runs, 51% of drafts passed every check first time; the rest were revised or rejected. How it works walks the full pipeline.
Common misunderstandings
"A detector score measures tells." It measures how predictable the wording is. A page dense with your own part numbers can score as machine-like, and a padded, hedged article can pass clean. See AI content detection.
"A humaniser fixes it." Injecting typos and rare synonyms changes the texture and leaves the emptiness in place. A hedge is still a hedge with an unusual verb in it.
"Only models write like this." Every habit on the list predates language models. Press releases, content-mill briefs and nervous junior copy produce the same shapes; models learned them from exactly that text.
"Removing tells makes a page rank." It removes a reason to dismiss the page. Search demand, internal links, indexing and what the page uniquely contains decide the rest.
"Length proves effort." Length is where tells accumulate. Three paragraphs with a specification table beat twelve paragraphs without one, and the short version is harder to write.
"An editor can catch them later." At any real publishing volume, nobody reads every draft. The check has to be part of the pipeline, not a promise about attention. See cross-model validation.
questions people ask
What is an AI tell?
An AI tell is a habit in the writing that marks it as generated: an opener that spends two sentences establishing that the topic matters, every idea arriving in a tidy group of three, a hedge such as costs vary depending on several factors where a price belongs, and a closing paragraph that restates the introduction. The deepest tell is not a phrase at all. It is flat specificity: statements that are true in general and contain no example, part number or figure that could be checked. A reader who knows the subject notices that within a paragraph.
How do I remove AI tells from my writing?
Cut the opener until the first sentence carries information. Replace each hedge with a number, a dated range, or an honest “we do not know”. Break the symmetry: let one section run long because the subject deserves it and another stop after two sentences. Add at least one thing only you hold, a specification, the price you charge, a fault you see every week, a photograph of the actual part. Then have someone other than the writer read it, because the author of a tell is the person least likely to see it. Typos and rare synonyms fix nothing.
Do AI tells hurt SEO rankings?
Not directly. There is no ranking factor called AI-sounding, and no per-URL score for it in Search Console. The damage is indirect and reliable: a hedged page answers nothing, so readers return to the results, and assistants such as ChatGPT and Perplexity find no checkable sentence worth quoting. Over months that shows up as pages losing impressions and never being cited. The remedy is the same in both cases, put something specific and verifiable on the page, which is why chasing a detector percentage is a detour rather than a strategy.
Can a model spot its own AI tells?
Poorly, on its own output. The weights that produced the sentence also produce the verdict, so a model reads its own phrasing as natural and its own hedge as appropriate caution. A model from a different vendor makes a better critic because it learned different habits, although it shares some of them and misses the subtle cases. That is why the separation is structural here rather than a prompt: the writer is never the checker, drafts get up to three revision rounds against named fixes, and grounding in first-party data does the work no critique can do.
See what search engines and AI assistants find on your site
Free, no account. Type your address and we show you what is missing and what we would write first.