You read a paragraph and something’s off. Nothing is factually wrong, no sentence is broken, but it doesn’t sound like anyone in particular wrote it.
That reaction is doing real work. Language models produce text through a specific mechanical process, and that process leaves fingerprints: predictable words, uniform rhythm, stock transitions, constant hedging. None of it is random.
This is a walk through those patterns one at a time, and why training a model to predict the next likely word produces exactly this kind of writing. NeverPrompted’s on-device check measures several of these patterns directly, so none of this is theoretical.
- 01AI writing tends toward the statistically expected word, not the specific or surprising one, because that’s what next-token training rewards.
- 02Human sentence length varies a lot within a paragraph; model output tends to smooth that variation out.
- 03A small set of stock transitions shows up constantly in generated text: they’re safe, high-probability connective choices, which is exactly why models lean on them.
- 04Hedging phrases (“it’s worth noting”, “arguably”) show up more often in generated text, because a confident, specific claim is a riskier bet than a qualified, general one.
- 05None of these patterns, alone, proves a sentence was generated. They’re signals to weigh, not a verdict.
The Most Likely Word Problem
A language model is trained to do one thing: given the words so far, predict which word comes next. It learns this by seeing an enormous amount of text and adjusting itself until its guesses match what actually followed, over and over, billions of times.
That training objective has a quiet side effect. The model gets very good at picking the word that’s likely given everything it’s seen, which is close to the average, expected choice across a huge number of writers and contexts. A specific, idiosyncratic word choice is a less likely bet than a safe, common one, so safe and common is what gets rewarded.
That’s why generated text leans on words like “delve”, “boasts”, “robust”, “landscape” and “tapestry” far more than any one human writer would. None of those words are wrong. They’re just the median choice, repeated at a scale no single person’s vocabulary would ever produce.
Sentence Rhythm: Why Every Sentence Feels the Same Length
Human writing is bursty. A short sentence lands, then a long one unpacks it, then a fragment. That variation isn’t decoration, it’s how people actually think, and how emphasis gets carried in prose. Read a paragraph you wrote a year ago and the sentence lengths will jump around without you ever having planned it.
A model generating one token at a time, optimizing for the locally likely continuation, tends to produce sentences that cluster around a similar length and a similar internal structure: subject, verb, qualifier, done. Nothing pushes it toward the sudden short sentence for effect, because “effect” isn’t what the training objective measures.
The result reads smoothly, which is exactly the problem. Real writing has bumps in it. A paragraph with no bumps at all, where every sentence unfolds the same unhurried way, is one of the more reliable tells once you know to look for it.
Stock Transitions: The Connective Tissue That Gives It Away
“That said.” “On balance.” “It’s important to note that.” “On the other hand.” These phrases exist in human writing too, but generated text reaches for them constantly, almost as scaffolding, because they’re extremely common connective phrases that fit an enormous range of contexts.
A model doesn’t know, in any meaningful sense, that a paragraph needs a transition. It knows that after a certain kind of sentence, one of a handful of familiar connecting phrases is statistically a very plausible next few words. So it reaches for the same short list again and again, regardless of whether the piece actually needs a formal signpost there.
Real writers vary their transitions, or skip them, because the connection between two ideas is often obvious from context alone. A text that names every logical link out loud, in the same few stock phrases, is doing more narrating than writing.
Hedging and Over-Qualification
A confident, specific claim is a riskier bet for a model than a qualified, general one. “This always works” can be contradicted by a single counterexample somewhere in the training data. “This can often help, depending on the context” is safe against almost anything.
That asymmetry shows up as a habit: “it’s worth noting that”, “arguably”, “in many cases”, “to some extent”. None of these phrases commit to anything a reader could actually check or disagree with. They soften a sentence just enough that it’s never quite wrong.
A person writing from direct experience doesn’t need to hedge that much, because they’re reporting something specific they actually know. Constant qualification, stacked sentence after sentence, is what a system trained to avoid being provably wrong produces by default.
The Em Dash as an All-Purpose Connector
An em dash can do almost any job in a sentence: it can introduce an aside, replace a comma, stand in for a colon, or link two related clauses. That flexibility is exactly why generated text leans on it so heavily. It’s a punctuation mark that’s rarely wrong, because it can plausibly fit wherever a writer needs some kind of pause.
Human writers tend to pick a specific tool for a specific job: a comma for a mild pause, a colon to introduce something, a full stop to end a thought cleanly. A model doesn’t need to choose that precisely, because the em dash covers most of the cases at once, and covering most cases is what a probability-driven system optimizes for.
That’s the reasoning behind NeverPrompted treating heavy em dash use as one measurable signal among several, not a standalone verdict. One em dash on a page proves nothing. A paragraph that leans on it every third sentence is a pattern worth noticing.
Why This Happens: Next-Token Prediction and the Average Voice
Every pattern above traces back to the same cause. A model trained to predict the statistically likely next token, across a training set drawn from an enormous cross-section of writing, converges toward something like an average voice: the phrasing, rhythm and vocabulary that fits the widest range of contexts with the least risk.
No single human writes like that average, because every person’s writing carries the specific quirks of what they’ve read, how they think, and what they’re actually trying to say in that moment. An average has no quirks by definition; it’s the point every individual voice deviates from.
Understanding this is what makes fixing it possible. The goal isn’t to trick a detector, it’s to push the writing back toward something specific and yours: a real sentence length you’d actually use, a word you’d actually reach for, a claim you’d actually stand behind instead of hedge around.
“None of these patterns are mysterious once you see the training objective behind them. A model rewarded for the likely next word will always drift toward the average sentence. Getting your own voice back means deliberately writing away from that average.”
Common pitfalls
- Assuming a single em dash or one instance of “delve” proves a paragraph was generated.
- Fixing AI-sounding text by swapping in fancier synonyms, which just trades one predictable word for another.
- Cutting out every connective phrase and ending up with flat, disconnected sentences instead of a natural rhythm.
- Treating vocabulary as the whole story and ignoring sentence rhythm, which is a harder pattern to fake by hand.
A detected mark is not proof of authorship, and an absent mark is not proof of human authorship. NeverPrompted's on-device rewrite can reduce detectable evidence but cannot guarantee defeating a vendor's undisclosed watermark, on any tier.