Trust isn't a setting. It's a skill.

The real risk in AI adoption isn't the machine — it's how we calibrate trust in it. A breakdown of automation bias, algorithm aversion, and the judgement we quietly lose along the way.

Most conversations about AI adoption focus on capability. Can it write the email? Can it build the model? Can it replace the analyst? The conversation that actually matters is different: do we know how to trust it?

Two failure modes

Over-trust, or automation bias, shows up when a fluent, confident answer feels right — so we stop checking. Aviation and clinical research have shown this for decades: skilled professionals accepting wrong suggestions they would have caught on their own. Fluency is not accuracy. It's just good delivery.

Under-trust, or algorithm aversion, is the opposite problem. The moment a tool that's been right ninety-nine times gets something wrong once, we write it off entirely and go back to gut feel. A human having a bad day gets grace a machine never does — even when the machine is right far more often.

The quiet third problem

Underneath both of these sits cognitive offloading. The more we hand off drafting, summarising, and first-pass thinking, the less we practise the judgement we need to actually review what AI gives us. GPS didn't just replace turn-by-turn directions — it replaced our mental map. AI is starting to do the same to our thinking. You can only meaningfully review AI output in a domain where you've kept your own judgement sharp.

Designing for calibrated trust

The fix isn't 'trust AI more' or 'trust AI less.' It's calibrated trust — task by task, matched to stakes rather than mood or convenience. That means explainability where it actually matters, deliberate friction on high-stakes outputs, a named human owner for every AI-assisted decision, and treating AI literacy as a governance control rather than an HR nicety.

Five things to change this week

Map your trust: for each AI tool you use, write down what you verify and what you wave through. Match checking to stakes: tie verification effort to consequence, not convenience. Keep one muscle deliberately: choose a core skill you'll keep practising without assistance. Name the owner: before any AI-assisted decision ships, be able to say who's accountable in one sentence. Reward the catch: publicly credit whoever questions an AI output and finds the flaw.

The measure of good AI was never how much it can do without us. It's how much better we become with it.