The AI Tells That Cost You Readers, Not Rankings
AI writing tells cost reader trust, not rankings. Six sentence-level patterns, the edit that removes each one, and why detectors miss them.
By David Jubé · Jul 22, 2026 · 14 min read
Readers, not detectors, decide whether a draft sounds written, and they decide it one sentence at a time.
Which sentence in a draft gives away that a machine wrote it? The first one. It restates the headline in slightly longer words, commits to nothing, and would sit as comfortably on a plumbing site as on a payroll app.
That sentence does not cost a ranking. What it costs is the reader who was deciding, somewhere around word eight, whether to give you word nine, and that verdict arrives long before any detection tool would render one.
Six patterns do the bulk of the damage. Every one of them is a place where a sentence chose a generality over a specific it could have carried instead, and every one of them has an edit that takes under a minute. Naming those six, giving the edit for each, and setting detectors aside is the work here, one level below the policy question the pillar of this cluster settles.
Key takeaways
- An AI writing tell is a sentence that trades a specific for a generality, and a reader reacts to it before naming it.
- Six patterns cover the ground: filler openings, hedges, placeholder examples, even rhythm, self-announcing transitions, and closing lines that restate.
- Each tell has a mechanical edit, so the fix is an editing pass rather than an instruction to sound more natural.
- Detector verdicts are unreliable and, more to the point, they measure something no reader is checking.
- Replacing a hedge with a number you cannot source is the one edit that leaves the draft worse than it started.
Readers Catch AI Writing Tells Before Any Tool Does
AI writing tells are sentence-level habits that trade a specific for a generality, and six of them carry the bulk of the complaint that a page reads as machine-written.
None of the six breaks a rule anyone has published. What Google actually penalizes about AI content reads the policy itself, and Ahrefs reached the same conclusion from the data side across 331,000 pages: quality, not origin, separated the pages that performed.
The variable a reader is judging is narrower than that, and faster. It is whether the sentence currently under their eye could have been written by someone who had never done the thing it describes.
Six habits give that away:
- A filler opening restates the topic instead of making a claim about it.
- A hedge softens a claim until it commits to nothing.
- A placeholder example fits any business in any industry.
- An even rhythm lands every sentence at the same length, often in threes.
- A self-announcing transition says it is connecting two ideas instead of connecting them.
- A closing line restates the paragraph in slightly different words.
Fixing them is an editing job more than a drafting one. Grow and Convert’s survey of marketers found that 60% say AI-assisted content needs more editing than human writing, and over 20% rewrite it outright.
Each one is a small failure of the same kind, which is why the edits are so similar. The evidence of effort a reader can actually see covers the page-level version of that proof, and Use AI to Write Without Sounding Like AI covers the general shape of the problem. The sentence is where it either holds or leaks.
Detector Verdicts Are the Wrong Instrument for This Job
Detection tools answer a question nobody in your audience asked. They estimate whether a machine produced the text, while a reader is deciding something else entirely: whether the text was worth their attention.
The reliability record is the smaller objection, and it is still a real one. Writing in Search Engine Journal in August 2026, Andy Betts ran one article he had written by hand through several well-known AI detectors and got back 100% AI, 78%, 42%, and human, on the same words.
That is a statement about third-party tools. It is not a claim about what any search engine does internally, and I have no visibility into that.
So the detector pass costs you twice:
- It optimizes for a reader you do not have: nobody arriving from search runs your page through a classifier before deciding to keep reading.
- It rewards the wrong edit: passing a detector can mean adding noise, which is the opposite of the specificity the sentence was missing.
- It displaces the check that works: reading the draft aloud finds an even rhythm and a hedged claim in one pass, at no cost.
Trade coverage of why generic AI output stopped performing traces the decline to editorial quality and reader behaviour rather than to any detection layer. Conductor’s assessment of where AI-generated content works and where it does not arrives at the same place from the practitioner side.
Generality Is the Failure, and Every Tell Is a Version of It
Underneath the six patterns sits one mechanism. A generality is a sentence that could be true of many situations, so it carries nothing about yours, and a reader who gets nothing from a sentence has no reason to buy the next one.
That is why “sound more natural” fails as an instruction. Natural is not the missing ingredient. A specific is.
Nielsen Norman Group put numbers on the reading behaviour underneath this back in 2008: on an average page visit, users have time to read at most 28% of the words, and 20% is more likely. Jakob Nielsen’s earlier study of the same behaviour found that 79% of test users always scanned a new page and only 16% read it word by word.
Whatever share of your page a reader samples, they sample the openings.
Put one question to any sentence: could this appear, unchanged, in a competitor’s draft?
- “Our approach is tailored to your needs” passes silent reading and fails that question.
- “Year-end close on a 400-line ledger runs two working days” fails nothing.
- “Results may vary depending on your industry” is a hedge wearing the clothes of caution.
The test has nothing to do with AI, which is the point Ahrefs makes from the drafting side in its argument that the gap was never the tool. A person writes generalities when they are tired or when they do not have the fact yet. A model writes them because a generality is the safest thing to produce when it has no access to your ledger.
Six Tells and the Edit That Removes Each One
Keep this table open while editing.
Openings and Closings That Restate
Both ends of a paragraph fail the same way, and both take the same edit: delete the sentence.
Read the first line of every paragraph and ask whether removing it loses information. If the answer is no, that line was throat-clearing, and whatever came second was the actual point.
The closing line is easier still. A paragraph that has to summarise itself in its last sentence was probably two paragraphs.
Hedges and Placeholder Examples
Hedges and placeholders both protect the writer from being wrong, at the cost of being useful.
The edit for a hedge is to find the claim it is avoiding and write that claim instead. If you cannot commit to it, the sentence is not ready and you need one more source, which is more useful than the hedge was. That is the same instinct behind writing the one definition sentence worth quoting rather than three that circle it.
Copyblogger puts the same edit more bluntly: cut the worthless words and let the precise ones carry the sentence.
The edit for a placeholder is a detail only you hold: a count from your own records, a named format, a real trade-off. Writing for E-E-A-T is the page-level version of that same trade, assertion out and proof in.
Rhythm and Transitions That Announce Themselves
Even rhythm is the tell that survives every other edit, because it is invisible to a spellcheck and audible in a read-aloud.
- Mark any run of three or more sentences landing at a similar length, then break one of them hard.
- When a list arrives in threes, check whether two items say it more plainly, or whether a fourth is genuinely true and you stopped at three from habit.
- Cut any transition that names its own job. “It is important to note that” is a sentence about a sentence.
Copyblogger has a name for that third habit. It calls it metadiscourse, writing about writing, and lists “note that” and “to sum up” among the phrases to strike on sight.
Paragraph length carries the same signal at a larger scale. A page where every paragraph runs four sentences reads as manufactured before a reader can say why, one reason length is an intent question, not a word-count target.
One Paragraph, Five Edits: A Hypothetical Worked Through
Take a hypothetical two-person bookkeeping studio writing about year-end cleanup. Everything below is invented for the demonstration, the figures included, and belongs to that imaginary studio alone.
The Draft and the Rewrite Side by Side
The Five Edits, in the Order They Were Made
Five edits got it there:
- Filler opening, deleted: the market-conditions sentence set up nothing the paragraph needed.
- Hedge, replaced with a figure the studio holds: “can potentially deliver better results” became a stated duration with a stated basis.
- Placeholder example, replaced by the studio’s own count: “a business might see improved efficiency” became a 400-line ledger and a named cause.
- Tricolon, cut to one honest claim: “fast, effective, and reliable” became a condition that changes the answer.
- Restating close, deleted: the summary line repeated the paragraph without adding to it.
Three sentences survive, and each carries something the original did not: a scope, a source for the figure, and a condition under which it fails. Editing is where that gets added: Siege Media’s 2026 roundup of AI writing statistics records the share of content marketers using AI to edit doubling in a year, to 38%.
Nothing about the rewrite is more natural. It is more specific, which is a different property and the only one a reader can use.
The rewrite is also, deliberately, a set of claims the studio could be wrong about. That is the point, and it is where the trap sits.
Founder Insight: Read the Draft Out Loud Reading a draft aloud finds an even rhythm and a hedged claim in a single pass, because the ear notices repetition the eye skims past. It costs a few minutes and it catches more than a checklist does.
Swapping a Hedge for an Invented Number Is the Trap
The edit that removes a hedge can produce something worse than the hedge. A hedge is vague. An invented number is false and reads as authoritative, which is the combination that costs a business more than a flat sentence ever did.
Specificity is only worth something when it is true.
So every fix above carries the same condition: the specific has to be one you can stand behind.
- A count from your own system, which you can go and check.
- A figure from a source you link, with its date visible to the reader.
- A stated opinion, labelled as yours, which cannot be wrong the way a number can.
- A worked hypothetical, labelled as hypothetical, in the way the studio above is.
That last one matters more than it sounds. A labelled hypothetical is honest and still specific. It is the right move whenever the real example is confidential or does not exist yet.
It is also the honest reading of what a drafting tool can do. Anthropic’s own prompt testing found that contextual examples improve a model’s answers while generic ones do not, which is the same asymmetry a reader applies and the working assumption behind making AI a writing partner that earns its place.
What a Clean Draft Reads Like When It Lands
A draft that has cleared all six does not read as clever. It reads as though somebody with the facts wrote it quickly and did not pad.
Concretely, the finished page looks like this:
- The first sentence makes a claim rather than announcing a subject.
- Paragraph lengths vary visibly down the page, because some ideas needed one sentence and some needed three.
- Every number has a source beside it or a link under it.
- At least one example could only have come from you.
- No paragraph ends by explaining what the paragraph just did.
That is also a review checklist, which is why it is worth running before publishing rather than after. Clearscope’s eleven-step editing pass is a longer version of the same instinct.
None of it turns on a tool’s verdict. A specific sentence is worth more to your chances of being quoted by an answer engine than a smooth one, because an engine extracts claims and a generality has none to extract, which is the same reason tables and other structured formats get pulled into answers so readily.
The last check is still the reader, and consulting one costs nothing. Read it out loud, and listen for the sentence that could belong to anyone.
Frequently Asked Questions
What is an AI writing tell?
An AI writing tell is a sentence-level habit that trades a specific for a generality: a filler opening, a hedge, a placeholder example, an even rhythm, a self-announcing transition, or a closing line that restates. It signals nothing to a search engine. It signals to a reader that the sentence carries no information they can use.
How do I fix a filler opening?
Delete it and start with the second sentence. A filler opening restates the topic rather than claiming anything about it, so removing it loses no information. Test every paragraph the same way: read the first line, ask whether cutting it costs the reader anything, and if the answer is no, cut it.
Which words signal a hedge that needs replacing?
Watch for “can potentially”, “in many cases”, “it is worth noting that”, and “some businesses may find”. Each protects the writer from being wrong instead of helping the reader be right. Replace the hedge with the claim it is avoiding: a number, a name, or an opinion you are willing to attach your name to.
Can a hypothetical example work instead of a real one?
Yes, provided you label it as hypothetical. A labelled hypothetical is honest and still specific, unlike a placeholder, and it is the right choice when the real case is confidential or does not exist yet. What fails is an invented specific presented as something that actually happened.
Does sentence rhythm actually change how a page reads?
Yes, and it survives every other edit because it is invisible to a spellcheck. When every sentence lands at a similar length and lists keep arriving in threes, a page reads as assembled before a reader can articulate why. Break one run deliberately, much shorter or much longer, and read the result aloud.
Do AI detectors tell me anything useful about quality?
No. They estimate production method, which is not what a reader is judging, and their verdicts disagree: one journalist ran a single hand-written article through several detectors and got 100% AI, 78%, 42%, and human on the same words. Fix the sentence instead of chasing a score.
Continue Reading
More From This Series
- What Google Actually Penalizes About AI Content
- The Evidence of Effort a Reader Can Actually See
- Scaled Content Abuse Is About Quality, Not About AI
- The Content Audit That Produces Your Next Ideas
- Making AI a Writing Partner That Earns Its Place
More from TDM Insights
- Use AI to Write Without Sounding Like AI
- How to Write a Definition AI Will Quote Verbatim
- Writing for E-E-A-T: Prove Experience on the Page
- How Long Should a Blog Post Be? Intent Over Count
- Content That Ranks and Gets Cited by AI
- Comparison Tables and Formats LLMs Love to Cite
Explore TDM Insights Topics