Tone is an intervention
Researchers varied explanation tone across three AI roles: assistant, second-opinion provider, and expert. Tone significantly changed decisions regardless of user characteristics in the second-opinion setting, while its effect in the other two roles depended on characteristics such as age and extraversion. “Friendly” is therefore not one treatment with one predictable result.
The CHI research on explanation tone found effects that depended on the role of the AI and user characteristics. That interaction is the practical lesson. There is no universally optimal brand voice for every answer. A coach, an auditor, a salesperson, and a safety monitor should not sound identical because they are asking the customer to relate to the information differently.
Persuasion can outrun evidence
Anthropic’s work on model persuasiveness measures how arguments shift stated opinions and warns that real-world behavioral impact remains a separate question. Product teams should add a further distinction: a persuasive answer may improve informed action, or it may make a weak claim easier to accept. A lift in agreement is ambiguous until we know what the customer understood.
Narrative research shows why tone and story can be effective: they shape attention, emotional engagement, and mental simulation. Those mechanisms can make a complex service relevant. They can also lower resistance to an answer whose uncertainty and incentives are hidden. Persuasion is therefore a capability to govern, not simply a marketing strength.
Friendly can turn into sycophantic
Sycophancy research finds that models can echo a user’s view because agreement is rewarded by human or model preferences. In customer-facing settings, the behavior may look like good service. The assistant validates the premise, avoids contradiction, and maintains a positive tone. But agreement can make a mistaken purchase, unsafe plan, or biased interpretation feel endorsed.
A useful tone policy separates respect for the person from agreement with the claim. The system can acknowledge goals and emotion while testing assumptions. It can state disagreement without dominance. It can ask a clarifying question without pretending confusion is the customer’s fault. These are rhetorical skills with direct safety and retention consequences.
Write the role contract
For each answer class, specify the system’s role, authority, interest, and desired customer action. A sales assistant may help explore fit but should not impersonate an independent adviser. A health explainer may clarify options but should not imply a diagnosis it cannot establish. A compliance assistant may be formal because traceability matters, while still using clear language.
Then translate the contract into observable style rules: how directly to recommend, when to name uncertainty, when to challenge a premise, which claims need evidence, when warmth is appropriate, and when the system must hand off. “Professional and friendly” is too vague to test. “Respectful, non-flattering, explicit about limits, and willing to correct a false premise” is closer.
Give the same facts three different social roles
Build role-consistent variants for one decision: warm peer, formal expert, neutral evidence guide, and your current voice. Hold facts and recommendation constant. Include cases where the user’s premise is right and wrong. Measure comprehension, calibrated reliance, willingness to continue, perceived respect, correction acceptance, and the business action.
Inspect subgroups and stakes. The voice that helps first-time users may irritate experts; the voice that increases a low-risk trial may be unacceptable for a denial or warning. Select tone by the behavior and risk of the answer class. A brand can remain coherent through principles while allowing different social roles to sound appropriately different.