NN/g's 2026 State of UX report identifies trust as the central design challenge of AI-powered products. Here is what calibrated trust means in practice — and three principles every AI product designer should apply now.
Every year, the Nielsen Norman Group publishes their State of UX report. The 2026 edition delivers a sharp message: trust is now the central design challenge of AI-powered products. Not personalization. Not speed. Trust.
As AI becomes embedded in more products — from smart assistants to recommendation engines to autonomous workflows — users face a new cognitive burden: deciding when to believe the system and when to question it. Getting this calibration wrong in either direction creates real consequences: misplaced confidence in AI errors, or blanket rejection of genuinely useful assistance.
The primary UX problem with AI in 2026 is not making it powerful enough — it is making it trustworthy enough for users to rely on in the right situations.
— Nielsen Norman Group, State of UX 2026
Two Ways AI Trust Fails
NN/g identifies two dominant failure modes users fall into with AI systems:
- Over-trust: Accepting AI outputs without critical evaluation, propagating errors downstream — especially dangerous in healthcare, finance, and legal contexts.
- Under-trust: Dismissing AI assistance entirely, missing real value and creating unnecessary manual friction.
The design goal is calibrated trust: users develop an accurate mental model of when the AI is reliable and when human judgment needs to step in.
Three Design Principles for Calibrated Trust
1. Show your work
AI systems that explain how they reached a conclusion — even briefly — perform significantly better on trust metrics. This does not mean exposing model internals; it means giving users enough context to evaluate the output themselves. A sentence like 'Based on your last 30 days of activity...' does more for trust than a confident recommendation with no attribution.
2. Surface uncertainty explicitly
When a recommendation carries lower confidence, say so. Confidence signals, 'based on limited data' labels, and explicit caveats give users the information they need to decide when to verify manually and when to rely on the AI. Hiding uncertainty to appear more capable is a trust debt that compounds.
3. Design for graceful recovery
AI will be wrong. The design question is whether that failure is visible, understandable, and reversible. Undo states, confirmation steps for high-stakes actions, and clear correction flows are not optional in AI products — they are the trust infrastructure that allows users to rely more heavily on the system over time.