Compliance should not start with a prohibited-words list

A word list cannot understand a promise

Prohibited-word lists are useful.

They can catch obvious terms such as "guaranteed", "risk free" or "always". They can remind writers about required language and help teams apply a consistent first filter.

They become dangerous when the filter is mistaken for the review.

A promotion can avoid every watched word and still create an unsupported promise. It can imply safety through imagery, certainty through structure or suitability through audience framing. It can present accurate facts in a combination that leaves a misleading overall impression.

Compliance has to examine meaning, not only vocabulary.

Context changes ordinary words

Consider the phrase "designed to grow with you".

In one context it may be harmless brand language. In another, surrounded by performance figures and lifestyle imagery, it may reinforce an expectation of investment growth. The words alone do not settle the issue.

The same is true of "simple", "secure", "better", "responsible", "free" and "affordable". Each can be accurate, misleading or immaterial depending on the product, audience, qualification and presentation.

A substantive review asks what a reasonable customer is likely to take from the communication.

Missing information creates risk too

Keyword detection is naturally better at finding what exists than what is absent.

But many serious compliance issues are omissions:

• a material condition is missing
• the risk warning is absent or too remote
• a comparison lacks its basis
• a rate appears without important context
• a sustainability claim has no substantiation
• an affiliate relationship is not disclosed
• the audience is not told who the product is for

The review needs a model of what the content should contain, not only a list of what it should avoid.

The same phrase can need different treatment

A watchlist usually produces one response: flag the term.

A useful compliance finding may need several possible treatments:

• remove the claim
• narrow it
• add evidence
• add a qualification
• change its prominence
• restrict the audience
• use a different channel
• escalate the approval route

The right action depends on why the phrase is risky. That reasoning should be visible to the reviewer.

False positives have a cost

If a system flags every occurrence of a word without context, teams learn to ignore it.

Compliance spends time dismissing harmless findings. Marketing begins to see the tool as an obstacle. Genuine issues become harder to distinguish from noise.

This does not mean a strong system never raises a cautious flag. It means severity and explanation should reflect context, and reviewers should be able to record why they accepted, acknowledged or dismissed the finding.

Those actions can then show where the model or customer policy needs improvement without turning customer content into provider-model training data.

Substantive review needs sources

Contextual reasoning should not become an excuse for unsupported AI opinion.

A useful system connects the issue to the relevant obligation, guidance or customer policy. The source makes the reasoning inspectable and helps the reviewer decide whether the proposed treatment is proportionate.

The standard is not "the AI sounds like a senior reviewer". It is "the accountable reviewer can see what in the content caused the issue, which source matters and what action is available".

Test the system with controlled examples

A good evaluation pack should include:

1. A harmless sentence containing a watched word.
2. A misleading implication using none of the watched words.
3. A promotion with a material omission.
4. A claim that becomes acceptable when evidence is attached.
5. An image where hierarchy changes the meaning.
6. A channel variant where the qualification is cropped.

If the product only performs well on the first example, it is a useful lexical filter. It has not demonstrated substantive compliance review.

Keep the list, raise the standard

Prohibited-word lists still belong in the toolkit. They are fast, transparent and valuable for known red flags.

They should sit inside a broader process that understands scope, detects missing context, examines the whole impression, links findings to sources and preserves the reviewer's decision.

The point is not to replace a simple control with mysterious AI. It is to combine deterministic controls with contextual analysis so each does the work it is best suited to do.

That is how compliance moves beyond finding bad words and starts helping teams make better decisions.