Key points
- "Sounds like AI" comes in three layers: marketing buzzwords, structural formula, and register. Only the third layer is independent of product, channel, and format.
- The first two layers get all the attention. The fingerprint sits in the third: abstract verbs, copulas, negative parallels, synonym loops, clumped filler words, and stacked hedging.
- Rules for this layer have to measure density, not presence. A check that fires on every sentence teaches reviewers to ignore it after two passes.
- Typography rules need tiers. In long-form, the list is the content, so a blanket ban just gets routed around.
AI writing gives itself away at the level of register, not subject: abstract verbs, sentences built on is, negative parallels, synonym loops, filler words arriving in clumps, and hedging stacked on hedging. Buzzwords and predictable structure shift with the topic. Register doesn't, which is why it reads the same everywhere.
Two lines in, something feels off. You can't name the problem yet, but you already know a person didn't write this.
That judgment is usually right, and it arrives without evidence. It also doesn't survive being turned into a test: detectors miss in both directions, flagging non-native English writers as machine-generated on one side and failing on generated text that's simply been paraphrased on the other. So this piece isn't about catching it after the fact. It's about not writing it in the first place. The trouble is that most advice on making AI writing sound human stops at swapping out a few words, and the smell survives the swap.
Three layers of "sounds like AI"
Layer one: marketing buzzwords. Empower, seamless, one-stop, game-changing. The most visible layer, and the fastest to clean up.
Layer two: structural formula. Rule-of-three parallelism. Scene setup, product name, feature list, call to action. This layer is tied to the format: the formula behind a blog post and the formula behind a fifteen-word caption have nothing in common, so each surface needs its own rules.
Layer three: register. Sentence-level habits. Abstract verbs holding up empty claims, sentences built entirely on is, negative parallels, the same idea circled three times in different words, filler words arriving in clumps, hedging stacked until nothing has been claimed, and a set of typographic tics.
Only the third layer is independent of product, channel, and format. A one-line ad and a two-thousand-word article make the same mistakes here, which is why this layer is worth writing down once and reusing everywhere. The rest of this piece is about layer three.
Six tells in the register layer
1. Abstract verbs standing in for a fact
Bad:
The redesign highlights our commitment to detail and demonstrates the team's product thinking.
Rewritten:
The redesign collapsed settings from three levels to one. Changing a notification preference went from four taps to one.
Highlights, underscores, reflects, showcases, demonstrates, exemplifies, fosters, embodies. When one of these turns up, it is usually covering for a specific fact the writer can't state, either because they don't know what actually happened or because the real answer is unflattering. Either way, the verb is doing the hiding.
The test takes five seconds. Cross out the verb and ask what happened. If you can answer, write the answer. If you can't, cut the sentence.
False positives here are rare, which is why this one can fire on a single hit with no threshold. Real writing almost never stacks these verbs.
2. Sentences built on is
Bad:
The dashboard is an important tool for decision-making, and reports are its primary output.
Rewritten:
The team opens the dashboard every Monday to decide where next week's budget goes, then exports one report into the weekly doc.
Is, serves as, constitutes, represents, is regarded as. These turn a sentence into an index card. Index cards have no time, no people, and no action, so the reader learns that the writer can categorize things, and nothing else. Where a real verb fits, use it. Where none fits, "has" usually will.
3. Negative parallels
Bad:
It's not a writing tool, it's a way of writing.
Rewritten:
The tool does one thing: it rewrites your draft against seven rules.
"Not X, but Y", "no need to X, just Y", "not only X but also Y". They read well because they borrow the rhythm of antithesis. The information content is often half a sentence: the first clause denies a claim nobody made, and the second one carries the actual point. People write like this occasionally. Models write like this in bulk.
The fix: delete the denial and state the point.
4. Synonym loops
Bad:
The problem lies in the process. The bottleneck across the pipeline ultimately comes down to how one stage hands off to the next.
Rewritten:
The problem lies in the process: nobody tells engineering when a design lands, so finished files sit untouched for about two days.
The same idea, restated to fill space. On the second pass a reader expects new information, finds none, and trusts the writer slightly less. Say it once and keep going.
5. Filler words arriving in clumps
Bad:
This capability is crucial for solving a pivotal bottleneck in content production. Furthermore, it seamlessly adapts to the shifting landscape of every channel.
Rewritten:
This solves one problem. The same piece has to run on four channels; you used to write it four times, and now you write it once and revise it three times.
Every word in the bad version is legal on its own. "Crucial" is an ordinary adjective. "Furthermore" is an ordinary connective. They only become a tell by arriving together, and that fact decides how the rule has to be written.
6. Stacked hedging and chatbot filler
Bad:
This approach may be somewhat helpful in certain contexts, although results will likely vary from team to team. Adjust as appropriate for your situation. Hope this helps!
Rewritten:
This works if you already have a draft. It won't help you start one; that problem lives somewhere else.
Stacked hedging reads as politeness and functions as a handoff: the writer refuses to make the call and pushes it back to the reader. Say what you know, and say you don't know the rest. Both beat three "may"s in one paragraph. Sign-offs like "hope this helps" and "in conclusion" work the same way, carrying tone and no information.
Why density beats presence
This is where a rule set most often goes wrong.
There is nothing wrong with the word "crucial." Ban it outright and the writer complies the first time, works around it the second time, and stops reading the rules the third. A check that fires on every sentence trains reviewers to skip it, which makes it worse than no check at all.
So rules need two shapes:
- Tight signals fire on a single hit. A pattern like "not X, but Y" can be matched with a bounded regex, so false positives stay low and one hit is worth one edit.
- Legal-in-isolation signals fire on density. Our threshold for register words is three distinct hits in a piece. One doesn't report.
Em-dashes belong to the second group and are harder still, because the reasonable count scales with length. Two em-dashes in a sixty-character caption already look strange; five across a two-thousand-word article read as normal punctuation. Any fixed number is wrong for one side or the other. Our formula: two by default, plus one for every eight hundred characters. That quota is a number we arrived at from our own drafts, not from any external standard — tune it to your own formats.
Two rather than one, deliberately. The em-dash is legitimate punctuation, and our own rule text uses four of them in about a thousand characters. Set the floor at one and the rule flags its own house style. A false positive costs a wasted rewrite pass and blocks the piece from auto-approving, which is more expensive than letting one through.
Typography rules need tiers, not a ban
"No em-dashes, no bullets, no emoji, no ALL-CAPS" holds up perfectly for a single line of ad copy. Applied to long-form it contradicts the job: in an article the list is the content and the subheadings carry the structure. Applied to a lifestyle-platform post it contradicts the channel, where a couple of emoji are part of how people write and zero reads as stiff.
A blanket ban gets routed around. So we split it into three tiers:
| Tier | Where it applies | How typography is handled |
|---|---|---|
| prose | Ad copy, short social posts | Em-dash quota; no emoji, no bullets |
| structured | Markdown long-form | Em-dash quota; no emoji in prose. Lists and subheadings are welcome |
| expressive | Channels that explicitly allow emoji | Em-dash quota; two emoji, carrying feeling rather than separating paragraphs |
All three tiers share one word list and one set of sentence rules. Only the typography line changes. Register travels; typography stays with the format.
Some tells can't be detected, only written around
Copulas in Chinese are the clearest case. The verb appears in nearly every sentence, in perfectly good ones, and any frequency threshold produces false positives at scale. We tried, then decided to skip the detector: that rule lives in the writing guide and belongs to the writer and the editor. That's our call on our own corpus, not a settled finding — run it on a different body of text and you might land elsewhere.
Admitting that beats shipping a detector anyway. A check that is wrong seven times out of ten does nothing except spend a reviewer's attention. A rule list should mark which lines only a human can enforce, so people know where to look.
A revision order that saves time
Given an AI first draft, running these in order beats editing top to bottom:
- Search for abstract verbs (highlights, underscores, showcases, demonstrates). For each hit, ask what happened. Answer it or cut the sentence.
- Search for "not only", "no need", "rather than". Flatten each one into a statement.
- Read for repetition. Where an idea appears twice, keep the concrete version.
- Count register words. Three or more means the whole piece has drifted, so rewrite paragraphs instead of swapping words.
- Count em-dashes against a quota set by length.
- Delete every "may", "perhaps", and "to some extent". Then reread and find the sentence you were afraid to make. Back it with evidence or replace it with "we don't know."
- Last, look at the copulas. Pick five and try turning them into actions.
The order matters. Steps one and two rewrite whole sentences, so counting repetition and filler words before that gives you numbers you'll have to recount.
How this runs inside LinkBloom
None of the above requires a tool. It stops being practical once you're shipping enough drafts that nobody wants to count by hand.
The content engine automates the countable part. Each finished article passes nine deterministic checks, and cliché density is one of them: cliché patterns are grouped into categories, with a cap on how many categories a piece may touch and a cap on total hits. Crossing either one records a flag; the occasional single hit does not. After that comes an editorial review scoring four dimensions, including factual accuracy and how natural the language reads.
Ad copy generation shares the same register rules, with the typography tier set to prose or expressive. It adds one deterministic scan: a hit triggers a targeted rewrite, and a piece that still hits after the rewrite gets flagged for a person.
Automated checks catch the systemic problems. Whether one particular sentence earns its place is still a human call.
FAQ
Would a different model fix this?
A different model shifts which filler words show up. The rest of layer three barely moves. Abstract verbs propping up claims, negative parallels, and synonym loops turn up across models. Writing rules costs less than switching.
Won't these rules make the writing flat?
If you only subtract, yes. Every abstract verb you remove has to be replaced by a specific fact, or the sentence just gets shorter and stays empty. These rules assume you have something concrete to say. If you don't, the problem is in your source material, not your register.
Do human writers make these mistakes too?
Constantly, especially on deadline. Layer three really describes the default output of a writer who hasn't decided what they want to say. A model is in that state permanently, which is why it writes this way every time.