The best AI interface is honest about uncertainty.

Users quickly learn to recognize polished generative responses. They do not always know whether a response was grounded, which part was inferred or whether the system already changed a record. The interface, not just the model, determines whether the user can make a sound decision.

The decision behind the technology

AI UX is often reduced to typing indicators and confidence chips. Neither explains the provenance of a specific finding. Confidence values may be poorly calibrated, and citations can link to irrelevant passages. Designers need to map the different states of the task, including when an answer is unavailable or an action requires review.

A better way to frame the work

Treat observed facts, model suggestions and committed actions as visually distinct. Provide the exact evidence behind significant claims. Make edits and rejection natural. Use side-by-side review for documents, structured fields for transactions and conversation only when discovery or exploration is the real job.

Put the idea under realistic pressure

A user reviews a contract amendment. The application highlights suggested changes, links back to clauses, states what could not be verified, and requires explicit approval before posting an update. If the source is unavailable, it does not create a fake citation. The system has a meaningful “cannot complete” state.

When the recommendation changes

Low-stakes drafting tools can use lighter interfaces. High-consequence systems need more explicit review. Overloading a harmless task with enterprise controls can make the product unusable; hiding controls in a consequential task is worse.

What to decide before moving forward

Design five states for every AI feature: ready, working, complete with evidence, needs human correction, cannot safely complete. Then test whether users understand the distinction.

References and context

NIST AI RMF Generative AI Profile.