120+ phrases that flag content as AI-written — hedge words, tricolons, marketing speak, vague authority, closing summaries — with category-organized copy-paste lists.
The complete AI banned-word library contains 120+ phrases organized into 8 categories: hedge words, AI bridge phrases, closing summaries, marketing speak, vague authority, em-dash overuse, tricolons, and "not just X but Y" constructions. Each category is a copy-paste list ready for your Persona Brief banned-words section. Banning these phrases at the output-time gate (not just at the prompt level) is what makes AI content stop sounding like AI.
Generic AI output has fingerprints. Every base model — GPT-5, Claude 4, Gemini 2 — has the same handful of phrases it falls back on when it cannot decide what to write. Banning those phrases is the single highest-leverage move in any Persona Brief.
This is the complete library, organized by failure category. Copy-paste the categories that match your industry into your brief.
Models hedge because hedged claims are harder to prove wrong in training data. Humans who actually know the topic do not hedge.
Replacement strategy: drop the hedge entirely. "Most teams" not "Many teams tend to." If a hedge is truly required, replace with a specific qualifier: "In the 4 cases we tested..."
Transitions models use when they cannot find a real connection between two ideas.
Replacement strategy: cut the bridge and start the next sentence with a noun or verb. The reader does not need to be told a new thought is coming.
Models close with a recap because their training data rewards it. Human writers usually close with a concrete line — a callback, a provocation, or a specific next action.
Replacement strategy: delete the entire closing paragraph if it summarizes. Replace with one concrete sentence — a callback to the opening, a specific next action, or a provocation.
The biggest category. These phrases are corporate jargon that AI inherited from training on marketing copy. Audiences tune out within 3 seconds.
Replacement strategy: replace with the plain-language alternative. "Leverage AI" → "use AI." "Streamline workflows" → "make workflows simpler." If the plain-language version sounds empty, the original was empty too — cut the whole sentence.
Phrases that sound authoritative but cite nothing.
Replacement strategy: cite a specific study, name a specific expert, or drop the claim. Vague authority is worse than no authority — it signals laziness.
Em-dashes are not banned outright — they have legitimate uses. But models overuse them at 5-10x human rates. Cap them in the brief:
Em-dash limit: 1 per 300 words.
Replacement strategy: convert most em-dashes to periods or commas. Reserve em-dashes for parenthetical asides that would otherwise need full parentheses.
Three-item lists are a base-model reflex. "Fast, reliable, and scalable." "Easy to use, easy to learn, and easy to love." Sprinkled occasionally these are fine. Three tricolons per paragraph is a dead giveaway.
Tricolon limit: 1 per 500 words.
Replacement strategy: convert most tricolons to two-item lists ("fast and reliable") or four-item lists ("fast, reliable, scalable, and observable") or single punchy claims ("fast — that's it").
The single most over-used AI sentence structure. Ban the literal phrase "not just."
Replacement strategy: drop the "not just" entirely and assert the second clause directly. "It is fundamentally different." "It is a workflow." "We are a partner." Direct claims always beat escalation patterns.
Add these based on your industry:
The list is a living document. A mature Persona Brief typically ends up with 150-250 banned words after 6 months of refinement.
No, if you replace banned phrases with concrete alternatives. The stilted version comes from prompting around the banned word without replacement. The right approach: ban the phrase, give the model a structural rule for what to do instead (e.g. "cut closings entirely, replace with one concrete sentence").
Yes — Kompozy ships a default banned-word list per Persona Brief that you can edit. The full library above is the starter set. Your custom additions stack on top.
Some do, some do not. AI detection is unreliable across the board. The right framing is not "evade detection" but "remove generic AI feel." Readers detect AI long before tools do — banning these phrases improves reader experience whether or not it affects detection scores.
Yes — add competitor brand names (when you do not want them surfaced), legal-sensitive terms (per industry compliance), and any internal jargon you do not want public. The brand-safety gate enforces all of them equally.
Aggressive. Over-banning makes the gate reject more outputs (which trigger regeneration). The cost of over-banning is regen latency. The cost of under-banning is shipped slop. Always over-ban.
← Back to AI Brand Voice & Persona overview · Start a free trial → · See pricing