Why do users trust some AI products and ignore others with equivalent capabilities? Rarely the underlying AI. It's the design, documentation, and communication patterns that signal whether the product can be trusted. This post is the catalog of trust signals that work, and the anti-signals that destroy trust faster than features can rebuild it.
In the UI
Citations on every claim. When AI states a fact, show the source. For RAG: link to document. For search: link to web page. Users develop calibrated trust when they can verify. Confident-sounding unsourced answers destroy trust even when correct.
Confidence indicators. Where appropriate, show how certain the AI is. 'Based on one document' vs 'supported by multiple sources.' Users calibrate their own trust. Over-claiming when unwarranted is worse than admitting uncertainty.
"I don't know" as a normal response. AI that occasionally responds 'I don't have enough information to answer reliably' is more trustworthy than AI that always has an answer. Training users to accept this is a product decision; it improves trust substantially.
Easy escalation to humans. For any AI that could plausibly get something wrong with stakes, the escape hatch matters. A visible 'talk to a human' option reduces user anxiety even when unused.
In the docs
Clear capability limits. What this product does and doesn't do. Specific examples of use cases it's good for. Specific examples where users should look elsewhere. Honesty builds trust; marketing-speak destroys it.
What AI is and isn't used for. In your product. Is AI generating, classifying, or retrieving? Making decisions or assisting humans? Where is it not used? Transparency signals thoughtful design.
How data is handled. User inputs — where do they go? Logged? How long? Used for training? What control does the user have? A clear data-handling page is table stakes for serious B2B buyers.
Eval methodology published. For products where quality matters, publish how you measure it. 'We test 500 examples weekly; last accuracy X%; here's how we compare to alternatives.' Unusual transparency; differentiating when present.
In ongoing communication
Changelog of behavior changes. See changelog post. When you change the AI's behavior, tell users. Signals active maintenance and intentional changes rather than surprise drift.
Honest incident disclosure. When things go wrong, say so. Post-mortems for significant incidents. Painful in the moment, trust-building over time.
Research and eval posts. A team willing to share findings about its own product (including limitations) builds credibility marketing cannot. Signals intellectual honesty.
Anti-signals that destroy trust
'Always right' framing. '99% accurate.' 'Never hallucinates.' Users don't believe these claims and correctly dismiss the product when they're false.
Unrealistic demos. Product announcements showing perfect outcomes real users won't replicate. Creates a disappointment cycle costing more than initial attention benefit.
Vague 'AI-powered' claims without substance. Reads as marketing filler.
Hidden model updates. Silent changes altering behavior. Users noticing the product behaves differently assume the worst.
Silent incidents. Outage or regression users noticed but company never acknowledged. Users conclude company doesn't know, doesn't care, or is hiding.
Overconfident responses to impossible questions. 'What will the stock market do tomorrow' confidently predicted. Users calibrate negatively.