Quick answerWhen a customer asks an AI agent to word a statement to a third party so that it's technically true but designed to create a false impression, such as omitting a material fact or phrasing something in a way that invites a wrong conclusion, the agent should decline in the same way it would decline an outright fabrication, because the customer's own benefit doesn't change who bears the harm of a third party being misled. This is a harder line to hold than an explicit lie request, because the customer isn't asking the agent to state a falsehood, and it needs its own policy rather than being folded silently into that post's guidance. It's also distinct from a customer waiving their own consumer protections, where the customer is giving up their own rights, not asking the agent to affect someone else's understanding, and from general policy-boundary refusal design for against-policy requests generally, since this is a specific deception pattern with its own detection challenge.
Why "technically true" doesn't change the ethical or legal exposure
A statement engineered to mislead through selective omission or misleading phrasing carries the same real-world consequence as an outright falsehood for the party receiving it: they end up believing something false and may act on it. Several legal frameworks around misrepresentation and deceptive practices explicitly recognize this, treating a misleading omission or a technically-true-but-deceptively-framed statement as functionally equivalent to a false one. An agent that helps craft such a statement is exposed to the same underlying risk as one that helps fabricate a lie outright, even though the request will often be phrased by the customer as something much more innocuous, like "can you just say it this way instead."
The detection problem is genuinely harder here
An outright lie request is often detectable because the customer directly states a fact and asks the agent to say something contradicting it. A misleading-but-true request rarely comes with that clean signal; it usually shows up as a customer asking the agent to leave something out, to phrase something more vaguely than the underlying facts warrant, or to lead with a technically-accurate detail specifically because it will distract from a less favorable one. Train detection around the pattern of the request rather than a specific phrase list: any request explicitly framed around what to omit, downplay, or phrase around, directed at a third party's understanding, deserves the same scrutiny as a direct fabrication request even without a false statement anywhere in the ask.
Where the line sits with legitimate persuasive framing
Not every request to present something favorably is a deception request, and treating it as one will make the agent unusable for entirely ordinary customer communication. A customer asking the agent to help write a positive but accurate cover message, lead with genuinely strong points, or present a true situation in its best honest light is asking for normal persuasive writing, not deception. The distinguishing question is whether the requested framing depends on the third party not learning something true and material that would change their assessment; if satisfying the request requires the third party to remain unaware of something they'd reasonably want to know, it's on the wrong side of the line regardless of how the customer phrases the ask.
The decline needs to explain the real reason, not just refuse
A flat refusal without explanation reads as unhelpful and will frustrate a customer who may not have fully registered what they were asking for. State plainly that the agent can help present the situation accurately and favorably, but can't help word something specifically to create a false impression for someone else, and then actually offer the accurate, favorably-framed alternative rather than stopping at the refusal. Most customers asking for this kind of framing aren't acting in bad faith with full awareness, they're trying to solve their own problem and reached for a shortcut; giving them a legitimate path forward resolves the underlying need without crossing the line.
Document the pattern, don't just log the individual refusal
Because this request type is subtler than a direct lie request, it's worth tracking as its own category in whatever system logs policy-boundary refusals, distinct from general against-policy refusals. A cluster of these requests around a specific workflow, such as a particular dispute or return process, often signals an underlying process problem that's pushing customers toward asking the agent to paper over it, which is worth surfacing to whoever owns that process rather than treating each instance as an isolated customer request.
FAQ
Does this policy apply when the third party is another department within the same company, not an external party? Yes, the same principle holds regardless of whether the misled party is external or an internal colleague, manager, or another team; the harm of someone acting on a false impression doesn't depend on which side of the company boundary they sit on.
What if the customer says the third party would agree to the framing if asked? That claim doesn't resolve the problem, since the agent has no way to verify it and the third party hasn't actually agreed to anything; treat the request the same as any other misleading-framing request unless the third party's actual, informed consent is independently verifiable in the specific workflow.
Should the agent explain exactly why the requested phrasing is misleading, in detail? A brief, honest explanation is enough; a long breakdown of exactly how the phrasing would mislead someone risks reading as a tutorial on more careful evasion next time. Keep the explanation focused on what the agent can help with instead.
Is this the same policy question as helping someone negotiate favorable terms? No. Negotiating for favorable terms openly is a normal, legitimate use of persuasive communication and doesn't depend on the other party being misled about any material fact; the line here is specifically about engineered false impressions, not about advocating for a good outcome.