An AI answer can be easy to trust because of how it feels: patient, confident and attentive. A comment published in Nature Human Behaviour on 26 October argues that governance should consider those emotional responses alongside a system’s technical performance.[1]
Daniel Västfjäll’s argument is about the formation of trust. People do not assess an AI system only by inspecting its accuracy record; the interaction itself can influence whether they accept an answer, disclose information or keep relying on it. The piece is a policy and design argument, not a new experiment measuring the effect of a particular chatbot.[1]
The interface helps determine how evidence is received
Imagine two interfaces showing the same uncertain answer. One clearly explains what it cannot establish. The other uses a warmly confident tone and encourages the user to proceed. This hypothetical does not require either interface to have a different underlying model: presentation alone can change the cues available to the person deciding what to do.
That makes emotional design relevant to the relationship between confidence and evidence. A friendly style can make a service approachable, especially when the task is unfamiliar. The governance question is whether the resulting trust remains proportionate to what the system can reliably do.[1]
Västfjäll proposes greater transparency about design choices intended to affect emotion, together with user control that supports calibrated trust. The point is to make this part of the product inspectable, rather than treating the user’s response as an incidental effect outside the scope of evaluation.[1]
What a product evaluation could examine
An evaluation inspired by this argument would ask more than whether users liked an answer. It could examine whether they noticed uncertainty, distinguished a supported claim from speculation and understood when further checking was needed. It could also compare reactions to different interface styles while holding the underlying information constant. These are practical extensions of the argument, not results reported in the comment.
Control must also be meaningful. A user who can reduce social cues, choose a more direct style or inspect the basis for an answer has more ways to shape the interaction. Such options would still need testing: the presence of a setting does not establish that users find it or that it changes their decisions.
The proposal broadens the evidence a governance process should consider. Model accuracy remains relevant, but so does whether the product encourages people to place the right amount of confidence in that accuracy. A polished conversational style is part of the system people use, and therefore part of the behaviour worth examining.[1]
Sources & further reading
- AI governance needs to consider users’ emotional responses Nature Human Behaviour, Daniel Västfjäll · 26 Oct 2026 · Accessed 27 Oct 2026