Calibrated Trust

Show people when to trust an AI answer, when to check it, and how to override it.

Show users when to trust, verify, and override AI output.

Calibrated Trust, side by side

Show sources and honest confidence, and give a fast override. Users can then trust selectively instead of blindly.

How to apply it

Always: cite sources, expose confidence honestly, give a quick override path.

Calibrated Trust in practice: a concrete example

An AI assistant that shows its receipts

When Perplexity answers a question it puts a numbered citation on every claim, so you can click through and check the one line you doubt instead of trusting or dismissing the whole answer. Compare a chatbot that states a wrong flight time in the same confident tone it uses for everything else. The first earns calibrated trust: you lean on it where it is strong and verify where it matters. The second teaches you to either over-trust it and get burned, or ignore it entirely.

Common Calibrated Trust mistakes

Presenting every output in the same confident tone, so users can't tell a solid answer from a guess.
Hiding sources and reasoning, which forces an all-or-nothing choice: trust blindly, or don't trust at all.
Making the override path slow or buried, so when the AI is wrong the user can't easily take back control.

When Calibrated Trust doesn't apply

For low-stakes, easily reversible output (a draft caption, a first-pass colour suggestion) heavy confidence UI and citations are just noise. Reserve the full trust scaffolding for decisions that cost money, touch real data, or are hard to undo, where a wrong answer accepted silently does real harm.

Practise it on a real challenge

AI Summaries — the ones that quietly hallucinate

An AI feature generates summaries, and users occasionally get a confident, wrong one with no way to tell — design the loop that catches it before they trust it.

AI Travel Assistant — Wrong Train Booking

An assistant booked a non-refundable 06:12 train after an ambiguous request for a 'flexible morning trip'.

Health Assistant — Unsafe Advice

A health assistant gives overly confident advice without explaining limits or directing the user to professional help.

Recruitment Copilot — Biased Shortlist

An AI tool silently filters candidates using unclear criteria.

Smart Calendar — Unexpected Auto-Reschedule

A calendar agent moves an important meeting based on incomplete context.

Email Assistant — One-Tap Reply, Wrong Tone

A one-tap smart reply sends instantly, in a tone that doesn't fit the thread.

Chat App — Your Messages Train the Model by Default

A new AI feature is on by default and trains on private messages; opt-out is buried.

Research Assistant — A Confident, Fake Citation

An AI answer includes a specific-looking source that doesn't exist.

Support Bot — No Way to Reach a Human

A support bot loops the user through canned answers with no path to a human.

Marketplace — AI Quietly Sets a Different Price for You

An AI raises a user's price based on their behaviour, with no disclosure.

Coding Assistant — Auto-Accepts Its Own Suggestions

A coding assistant auto-applies its completions, so the developer stops reviewing what lands in the file.

Photo App — AI 'Cleanup' Deletes the Wrong Shots

An AI 'cleanup' permanently removes photos it judges duplicate or blurry — some the user wanted.

Content Moderation — Removed With No Reason

An AI removes a user's post and gives no specific reason and no way to appeal.

Put Calibrated Trust into practice

AI Prompt Engineering Kit — A designer's AI toolkit — prompts that assist, not replace.