When an AI system makes a recommendation, the person in the loop faces a simple question. Do I trust this or not?
Get it right and person and machine outperform either alone. Get it wrong in either direction and the partnership underperforms. Researchers call the goal appropriate reliance.
Two ways to get it wrong
Over-reliance, often called automation bias, means accepting output without enough scrutiny. Fluent AI makes this tempting, because wrong answers sound as confident as right ones.
Under-reliance is real too. People who have been burned override good recommendations and lose the benefit.
The 2026 workslop study shows the cost of over-trust. Among 962 US desk workers, those with high trust in AI were more likely to send low-quality AI work to colleagues.
Why calibration is getting harder
AI reliability varies by task in ways that are hard to see. The same tool can be excellent on one question and wrong on the next.
Stanford's AI Index 2026 found developers are sharing less about how their models work. Its transparency index fell from 58 to 40. Less information makes trust harder to judge.
Most users already sense the risk. Microsoft's 2026 Work Trend Index found 86% treat AI output as a starting point. The question is whether they have the time and habits to check it.
Build this capability in your organisation
Growth Performance designs bespoke programmes that turn insight into measurable behaviour change.
Calibration can be designed
Simple prompts that make people pause before accepting AI advice reduce over-reliance. Buçinca and colleagues called these cognitive forcing functions.
Forming your own view before seeing the AI's answer also helps. A 2026 study in Frontiers in Psychology found that checking your own thinking weakened the slide into dependence.
What this means for leaders
Calibrated trust connects two conditions. People must feel safe to say when they doubt the AI, which supports Adoption. The work must build in checking, which is Adaptation.
Add friction where stakes are high. Require a short written view before anyone sees the AI's answer.
Respect good overrides. When someone overrides the AI with a sound reason, say so in public.
Track disagreements. Note where people and AI disagree. Those cases show where your frontier sits.
Try this
For one important recurring decision, change the order. Ask the person to commit to a view first. Then reveal the AI's recommendation and reconcile the two.
To see whether your team has time to check properly, try the Checking Capacity Calculator.