Brand safety for autonomous AI content: how Kompozy handles prohibited keywords

Published 2026-05-18 · Updated 2026-10-23

Telling a model "never use the word leverage" works most of the time. The rest of the time is the whole brand-safety problem: the post that slips a banned phrase through is the one your audience reads, and on autopilot nobody reads it first.

An earlier version of this page described an output-time regex gate that rejected and regenerated any post containing a banned word. Kompozy does not have that gate. Your prohibited keywords reach the model as an instruction, and nothing scans the result. This page covers how the instruction works, why instructions slip, how to write a list the model follows more often, and where you still need to look yourself.

Why are prompt-level instructions not enough?

The intuitive fix for AI tells is to list your banned words in the generation prompt and trust the model to obey. It mostly works, which is exactly what makes it dangerous — "mostly" is not a standard you can ship autonomous content against. Three structural forces cause base models to violate prompt-level banned-word instructions, and none of them are fixable by writing the instruction more forcefully:

1. Training-data bias. The model has seen "leverage," "delve," and "in today's fast-paced world" in millions of marketing documents. The statistical pull toward the phrases that dominate its training data competes with your specific instruction, and on enough outputs the pull wins.
2. Context drift. As the prompt context grows — Persona Brief, source material, format instructions, reference posts — a single "do not use X" line gets diluted relative to the total weight of everything else in context. The longer the prompt, the weaker any one rule.
3. Synonym substitution. Even when the model obeys the literal ban, it sometimes swaps in a near-synonym that is just as bad. Ban "dive deep" and it writes "explore in depth." Ban "unlock" and it writes "unleash." The instruction was followed; the AI tell shipped anyway.

An output check gets around all three because it reads the finished text instead of trusting the model. Kompozy does not run one for prohibited keywords, so these forces apply to every post it writes. That is the reason to keep reading outputs, or to turn on Auto-Publish only in workspaces whose output you have watched for a while.

How does Kompozy handle your prohibited keywords?

Prompting settings have a Prohibited Keywords field next to the Persona Brief. Here is what happens to it:

1. You type the words and phrases you never want to see. The field is free text.
2. On every text generation, the field is added to the Voice Profile section of the prompt as "Prohibited words/phrases," along with tone, personality, point of view, complexity, formatting and emoji rule.
3. The built-in anti-AI-tell block goes in too. It bans em dashes, stock openers like "In today's fast-paced world," "it's not just X, it's Y" constructions, and a word list that includes delve, leverage, unlock, navigate, journey, robust, seamless and elevate.
4. The model writes the post. The finished text is not scanned for any of these words, and a post that contains one is not rejected or regenerated.

An admin compliance matrix decides whether each format gets the Voice Profile and anti-AI-tell blocks. Both are on for every format by default, for first generation and for regeneration.

Persona video scripts get one output-side check. Persona Short and Persona Frames run a critic pass that scores the script on AI tells, hook, pace and specificity, and swaps in a revision when any score is under 6 of 10. The critic looks for tells like stub sentences, meta-labels and anonymous-expert anchors. It does not look for your prohibited keywords.

Where does the list live, and what belongs on it?

The list lives in the Prohibited Keywords field and applies to every text format generated for that profile. The built-in anti-AI-tell rules already cover the most common AI vocabulary, so use your list for what those rules do not know about. Three kinds of phrase belong there:

1. AI tells the built-in list misses. Phrases you keep spotting in your own outputs, such as "game-changer" or "at the end of the day."
2. Industry cliches. Real estate has "motivated seller" and "passive income"; SaaS has "synergize" and "low-hanging fruit." These mark content as generic inside your niche.
3. Brand-specific bans. Competitor names, claims your legal team flagged, internal jargon that should never reach the public.

Because the list is an instruction and not a filter, the entries that matter most, like competitor names and claims you cannot legally make, also need a human check. A rule the model follows on most outputs is not enough for a claim that can get you in trouble once.

A starter list of your main industry cliches and brand bans is enough to begin with. It grows from the editing you do during the manual ramp, while you still approve every output.

How do you write entries the model follows more often?

Without a filter, nothing does word-boundary matching for you. The model reads your list as language, so plain, specific entries work best:

• List exact phrases. "dive deep" and "deep dive" read as different phrases. If you want both gone, list both.
• Ban the construction, not the fragment. A lone entry like "tech" is vague. "tech-forward" or "cutting-edge tech" tells the model exactly what to avoid.
• Name the synonyms you also dislike. Ban "unlock" and the model may write "unleash." If the swap is just as bad, add it.
• Keep to phrases you actually see. A long list of words the model never used adds noise to the prompt and dilutes the entries that matter.
• Put the reasoning in the Persona Brief. A sentence like "we never use hype verbs; say what the product does" gives the model a direction, which a bare word list does not.

How do you steer the model toward a better replacement?

Kompozy does not regenerate a post because it found a banned word, so the steering has to live in the original prompt. Pair each kind of ban with a direction in the Persona Brief:

Stripping words with no direction is what makes output sound stilted. Give the model a target and the result reads cleaner instead of emptier. If a post still comes out wrong, regenerate it from the pipeline; the new version is written under the same rules.

Which banned-word patterns are industry-specific?

The universal AI-tells library is the same for everyone, but the industry layer is where most workspaces win or lose on voice. The cliches that mark content as generic are field-specific, and a model trained on the internet will reach for your industry's most overused phrases unless you ban them explicitly. Starting points by vertical — extend each with the phrases you personally keep cutting:

SaaS / B2B tech

• synergize
• synergistic
• circle back
• low-hanging fruit
• value-add
• cross-functional
• best-in-class
• move the needle
• mission-critical

Real estate

• motivated seller
• investment opportunity
• wealth-building
• passive income (when overused)
• cash-flowing asset
• build generational wealth

Coaching / consulting

• mindset shift
• limiting beliefs
• next-level mindset
• breakthrough
• transformative journey
• unlock your potential
• show up as your best self

Health / wellness

• wellness journey
• holistic approach
• natural solution
• gentle yet effective
• nourish your body
• self-care ritual

Two cautions on the industry layer. First, some of these phrases are only bad when overused — "passive income" is a legitimate real-estate term and a cliche when it shows up in every post. Ban the ones you are tired of seeing, not the entire vocabulary of your field. Second, the industry list is the part you should expect to revise most often, because field cliches rotate; the phrase everyone overused last year reads as dated this year, and a fresh AI tell takes its place. The monthly tuning loop is where that revision happens.

What does Kompozy enforce, and what does it miss?

What the current setup does:

• Puts your prohibited keywords in front of the model on every text format.
• Adds the built-in anti-AI-tell rules, which cover the most common AI vocabulary and sentence patterns.
• Runs a critic pass on persona video scripts that scores AI tells and revises once when a score is low.
• Holds batch-generate output for your approval, even with Auto-Publish on.

What it cannot catch:

• A prohibited keyword the model used anyway. No format scans, rejects or regenerates for it.
• Tone that misses the brief in ways the rules do not name.
• Misleading framing or weak strategy. A clean post can still make the wrong point.
• Invented numbers, which are handled by separate rules covered on the fact-anchor page.

If a phrase must never appear, a word list in the prompt is not enough. Keep Auto-Publish Generated off for that workspace, or read scheduled posts before their slot comes up.

What are the common list problems, and how do you fix them?

• Over-banning. A very long list dilutes every entry and can push the model into stiff phrasing. Trim it to phrases you actually see.
• Under-banning. A thin list lets generic phrasing through. The sign is that you keep editing the same phrases out by hand. Anything you delete twice belongs on the list.
• Mixed signals. A phrase banned in the keyword field but used in your Persona Brief examples tells the model two things at once. Take it out of the examples too.
• Vague entries. Single short words are easy for the model to treat as a suggestion. Prefer specific phrases.

The list needs the most work early on, while you are still reviewing every output. Once you approve most posts untouched, it has done most of its job.

How should you tune the list over time?

1. Review recent outputs. Every phrase you keep cutting by hand belongs on the list. Your own edits are the best source of new entries.
2. Check whether entries hold. If a listed phrase still shows up, reword the entry, add the variants the model uses, or add a direction to the Persona Brief.
3. Watch your edit rate. The share of outputs you change before approving should fall as the list and brief settle. If it climbs, look for what changed, such as a new source type, a new format or a Persona Brief edit.
4. Prune stale entries. Drop phrases that no longer appear, and add new tics as models change.

For the full map of what Kompozy checks and where, see quality gates. For building the list while you still approve every output, see the manual-vs-autopilot ramp. For how one source fans out across platforms, see content repurposing. Autopilot is included on Kompozy's Pro, Agency, and Enterprise plans; see pricing.

What does this look like in Kompozy?

A Persona Tweet generated in the Kompozy workspace: short copy in a brand voice, the kind of output a banned-word list is checked against.

Who is this for, and who should skip it?

Best for:

• Brands with a clear list of AI tells and industry cliches to avoid
• Teams that keep deleting the same phrases from published posts
• Operators deciding whether to auto-publish a workspace

Not ideal for:

• Anyone expecting automated enforcement of a banned-word list
• Catching tone or strategy problems, which no word list can see
• Compliance-grade filtering, which still needs human review

Frequently asked questions

How do we know this?

We checked every product claim on this page against Kompozy's generation, autopilot and publishing code on 2026-10-23, and plan details against kompozy.io/pricing. The earlier description of a regex gate with rejection and regeneration was removed because the code has no such gate; the page now covers how prohibited keywords reach the model.

Sources

1. Kompozy — Glossary: Quality gates (read 2026-10-23)
2. Kompozy — Pricing (plans, credits, workspaces, autopilot, HeyGen BYO) (read 2026-10-23)

Which other autopilot guides should you read next?

• AI content quality gates: what Kompozy checks before a post ships — Which checks run on which formats: prompt rules everywhere, a three-layer newsletter gate, a persona script critic, publish limits, and what is not checked.
• Fact-anchoring AI content: how Kompozy limits invented stats — Kompozy has no output-time fact check. How its prompt rules and script critic limit invented numbers, and how to check the rest yourself.
• Posting cadence by platform for autonomous content — Kompozy has no per-platform cap: you set posting days and posts per day for each format. Typical rhythms per platform to use when you set those limits.
• AI content autopilot: what it actually means in 2026 — Definition, mechanism, failure modes, and the brutal honest assessment of which AI tools actually achieve true autonomy.
• Founder-led marketing autopilot: scale your voice without losing it — How founders run daily content output across 9 platforms with 15 minutes of review per day.
• The 14-day manual-to-autopilot ramp methodology — How to graduate from full manual review to fully autonomous shipping in 14 days without losing voice control.

Which related topics connect to autopilot?

• AI Brand Voice & Persona — Without a Persona Brief, every AI output averages to the LLM default voice. This is the 5-section methodology that makes 100+ AI-generated posts feel like one human author wrote them.

← Back to Autonomous Content Creation overview · Get started →