The Nine Tells of AI-Generic Writing

AI-generic writing is easy to feel and hard to name, so most teams call it "a bit flat" and move on. It is not flat. It carries nine recurring surface tells, and each one is a measurable expression of the same thing: a language model returning the most probable next word, which is the category average dressed as your topic. Ivanooo runs a scan for these nine patterns on every piece it publishes, including this one. Firoz Azees built the scan because the tells are countable, and anything countable can be removed on purpose.

The short version: There are nine patterns (P1 to P9) that mark machine-default writing, from predictable phrasing to parachuted numbers to borrowed authority. They read as small stylistic quirks. They are the fingerprint of copy pulled toward the middle of the training data.

Remove them and the writing stops matching everyone else's, because it stops being the average. That is why the nine tells are a Voice instrument, not a style guide: they measure distance from the category center, and closing that distance is what makes a brand attributable again.

Where the nine tells come from

A model does not write, it predicts. Given your brief it returns the highest-probability continuation, then preference tuning sands off whatever annotators found odd. What survives is the middle of the middle: the reduction in output diversity researchers call mode collapse. The nine tells are what that middle becomes once it reaches the page.

So the tells are not signs of a careless writer. They are signs of an obedient tool doing exactly its job. Predictable phrasing, agentless action, hedge density: each is the model choosing the safe center over the specific edge. Count them and you are counting how far a page has drifted toward the category average.

The nine tells, one line each

Here is the full scan. Each tell is one line: the pattern, and what it signals. The before/after that follows shows the same sentence written toward the average, then written toward the brand.

Tell	What it signals
P1 Predictable phrasing	Every next word is the expected one; the reader finishes the sentence for you
P2 Scaffolding overload	Transitional connectors carry the structure instead of the ideas
P3 Agentless action	Things happen with no one doing them; nothing is named, no one acts
P4 Parachuted data	A number drops into generic prose as decoration, not tied to a real observation
P5 Parallel triplets	Everything wrapped in three ("fast, flexible, focused") as the default rhythm
P6 Flat cadence	Sentences cluster in one length band; no short punch, no long thinking-aloud
P7 Categorical abstraction	"The business," "the customer," "the platform" replace real named things
P8 Hedge density	Every claim softened into safe territory until it commits to nothing
P9 Borrowed authority	A famous name or a vague "research" cited with no quote, source, or date

What each tell costs, in brand-copy terms

Read the tells as edits, not grades. Take P3, agentless action. Generic: "Engagement is driven through quality content." Fixed: "Firoz Azees writes the audit himself, then the scan grades it." The first could sit under any founder's name; the second names a person doing a thing.

Take P7, categorical abstraction. Generic: "The platform helps businesses stand out." Fixed: "Ivanooo measures your voice against your category's average and hands you the distance." One commits to nothing. The other commits to a named brand and a named measure. And P8, hedge density: a claim wrapped in soft qualifiers about stronger recognition becomes "This makes the brand recognisable." The hedge is the tell. Deleting it is the move.

Why removing them makes you attributable

Every tell you remove moves the page off the center that the model defaults to. That matters because the center is crowded. In a controlled study, generative AI raised the quality of individual writing while lowering the diversity of the collective output: each piece scored better on its own, yet the pieces resembled each other far more. Doshi and Hauser put it plainly: generative AI "enhances individual creativity but reduces the collective diversity of novel content." The nine tells are the mechanism of that merge, made visible one pattern at a time.

So a brand that eliminates them is not "writing better." It is writing further from the average, which is the one property a competitor cannot reach by prompting harder. This is where Voice meets the other two layers Ivanooo instruments. Voice measures the distance. Entity makes the distinct writing resolve to your brand by name, so the sharp page credits you and not the category. Topic Authority supplies the proof only you own, the material no model could have generated because it was never in the training set. AI can average your voice; the scan is how you stop it.

How to de-generic a page

You do not need the whole instrument to start. Three passes clear most of the nine:

1. Name the actors. Find every floating verb and attach a real person, brand, or place. This kills P3 and P7 together, because agentless action and categorical abstraction are the same avoidance.
2. Cut the hedges and the scaffolding. Delete the softening qualifiers and the academic connectors that carry your paragraphs. This clears P2 and P8, and it is the fastest read on how much a page was hiding.
3. Ground every number and name. If a stat is decoration, remove it or tie it to a real observation. If a famous name has no quote or source, cut it. That closes P4 and P9.

Run those three and you have removed six of the nine by hand. The remaining three, predictable phrasing, parallel triplets, and flat cadence, are about rhythm, and rhythm is where the voice really lives.

The tells are the same reason a model is a machine for the average: the exception is the one thing it cannot generate for you. The metrics most teams watch, mentions and share of voice, measure whether the machine knows you exist, not whether it can still tell you apart. The nine-tell scan is the second test, and it is the one that decides distinctiveness. See where your voice sits against your category's average: paste your URL, get the nine-tell scan run on your pages, with the evidence. No call required.

FAQ

What is AI-generic writing? It is writing that reads smooth but matches everyone else, because a language model returned the most probable continuation and preference tuning narrowed it further toward what annotators rated safe. The result carries nine recurring surface tells, from predictable phrasing to borrowed authority, that mark it as writing toward the category average.

Are the nine tells writing mistakes? No. They are the fingerprint of a tool working as designed. A model is built to return the center of its training data, and the nine tells are what that center becomes on the page. They are not errors to correct so much as a direction to reverse: away from the average, toward the specific.

Which tell is the most damaging? Predictable phrasing and flat cadence, because they affect the whole piece and not one sentence. Borrowed authority is the most local; you fix it by removing a single unsourced name. The scan flags all nine, but the rewrite starts with the two that touch every line.

Can I fix this by switching AI models or prompting harder? No. The pull to the average is a shared property of how these models are trained, not a quirk of one vendor, so a different tool moves you between near-identical centers. Instructing a model to "sound distinctive" targets the average of everything already labelled distinctive. Distinctiveness has to be supplied from outside the model.

How do I know if my own pages carry the tells? Run the scan. Take a page, strip the logo, and check it against the nine patterns: are there named actors, or floating verbs? Grounded numbers, or parachuted ones? A real cadence, or one flat band? If the page could sit under any competitor's name, it has converged, and the tells will show you exactly where.