Beyond the uncanny valley: make AI-assisted experiences feel trustworthy

The uncanny valley usually describes discomfort with something that appears almost human but feels noticeably wrong. In digital products, similar discomfort can also come from a mismatch: a warm voice making an inaccurate claim, a face with unnatural movement, or a chatbot implying a relationship it cannot actually sustain. Generic visual style is a different problem, though both can weaken trust.

Choose the right level of realism

Ask what a human likeness contributes to the task. A clearly illustrated character may explain a service better than a near-photorealistic presenter. For product support, a clear interface and useful answer may matter more than a face. Do not introduce a synthetic person simply to make the system seem human.

If you use an avatar, establish a coherent style across face, voice, expression and movement. Stylization can make the intended fiction clear. Test the result with the audience rather than assuming your team finds the same details comfortable.

Inspect temporal consistency

A still image can look convincing while motion exposes mismatched gaze, teeth, hands, blinking or lip synchronization. Review scene transitions and close-ups, not just the opening frame. Listen for unnatural emphasis and changes in voice identity. Check product and environment details for unexplained movement.

Reduce unnecessary close-ups or motion when they do not serve the message. Simpler production can improve coherence while making defects easier to identify.

Be clear about the interaction

Do not present generated people as actual customers, staff or experts. Use appropriate disclosure where the synthetic nature affects understanding. In a conversational product, make clear what the assistant can do, when it is uncertain and how a person can correct or stop an action.

A friendly tone cannot compensate for a missing escape route. Provide useful status, editable drafts, confirmation for consequential actions and a path to human help when the service requires it. Avoid claims of human feelings or personal knowledge the system does not have.

Evaluate more than whether people like it

Ask viewers what they think they are seeing, what they believe the system knows and what they expect it can do. Observe task completion, hesitation and mistaken assumptions. Separate visual discomfort from confusion about capability.

Include people with relevant access needs and different familiarity with AI. Offer captions, transcripts, reduced-motion options and alternative ways to obtain essential information as appropriate. Do not force someone to interact with an avatar to complete an ordinary task.

Keep a trust review in the workflow

Record misleading implications, inconsistent details, unclear boundaries and recovery failures. Assign an owner to each fix. A coherent, clearly artificial experience can be more trustworthy than an elaborate simulation that creates false expectations.

The aim is not to make synthetic material impossible to recognize. It is to make the experience honest, understandable and appropriate to the relationship with the user.