Trust Calibration: The Make-or-Break Skill of Human-AI Collaboration

In our first article we described the AI Adoption Paradox: 92% of companies are increasing AI investment, yet only 1% call themselves 'mature.' In the second we showed that psychological safety is the hidden variable that decides whether people even try the tools.

But trying isn't the finish line. Once your team is using AI, a harder question appears — and it's the question that quietly determines your return on the entire investment: are they using it well?

Here's the uncomfortable answer from the research. The single largest meta-analysis on the subject found that, on average, human-AI teams perform worse than the better of the human or the AI working alone. Not better. Worse. The value doesn't usually leak because the model is weak. It leaks because people trust it wrong.

The skill that closes that gap has a name: trust calibration. This article is about what it is, why it's the bottleneck, and how to build it.

What trust calibration actually means

Most conversations about 'trust in AI' treat trust as a dial that should be turned up. Get people to trust the tools more, the thinking goes, and adoption follows. That framing is wrong, and it's expensive.

The goal is not more trust. The goal is calibrated trust — relying on AI when it is likely to be right, and overriding it when it is likely to be wrong. Researchers studying human interaction with automation have called this appropriate reliance for decades (Lee & See, 2004; Parasuraman & Riley, 1997). It is a match between how much you trust a system and how trustworthy that system actually is, for the specific task in front of you.

Two ways to get it wrong:

Under-reliance — rejecting good AI output. The team second-guesses, re-does, or ignores recommendations that were correct. Over-reliance — accepting bad AI output. The team defers to a confident-sounding answer that was wrong.

A well-calibrated team minimizes both. A miscalibrated team can be enthusiastic or skeptical and still destroy value — just in opposite directions. This is why 'is the team bought in?' is the wrong diagnostic. Buy-in without calibration is just faster mistakes.

Failure mode one: algorithm aversion

In a now-classic set of five studies, Dietvorst, Simmons and Massey (2015) gave people a forecasting task and let them choose between their own predictions and an algorithm's. The algorithm was measurably better. People still abandoned it — specifically after watching it make a mistake.

The sharp finding wasn't that people dislike algorithms. It's the asymmetry: people lose confidence in an algorithm far faster than in a human after seeing the same error. A human who slips up gets the benefit of the doubt. An algorithm that slips up gets fired.

This is algorithm aversion, and in an organization it looks completely reasonable from the inside. An analyst sees the AI get one forecast wrong, concludes 'it's not reliable,' and quietly goes back to the spreadsheet. No drama, no resistance — just a slow, invisible drift back to pre-AI ways of working, while the license fees keep auto-renewing. It is one of the most common reasons an AI rollout posts strong usage numbers in month one and flat ones by month six.

Failure mode two: over-reliance

The opposite error is just as costly and far less discussed, because it doesn't feel like a problem. It feels like productivity.

Logg, Minson and Moore (2019) ran six experiments and found the mirror image of aversion: algorithm appreciation. People often gave more weight to advice when they believed it came from an algorithm than when they believed it came from a person. So the picture isn't 'humans distrust AI.' The picture is messier — and more useful. Appreciation is strongest in domains people see as opaque or technical, and weakest in domains they see as needing human judgment, like character or ethics.

The danger surfaces when appreciation hardens into deference. A fluent, confident, well-formatted AI answer is extraordinarily persuasive — and modern generative tools produce wrong answers with exactly the same fluency and confidence as right ones. When a team stops checking because 'the AI is usually right,' it has stopped collaborating with the tool and started rubber-stamping it.

This shows up in the most counter-intuitive result in the field. Bansal and colleagues (2021) tested whether having the AI explain its reasoning would help human-AI teams. Intuitively, explanations should help people catch errors. They didn't. Explanations often increased people's tendency to go along with the AI — including when the AI was wrong. A persuasive rationale doesn't sharpen judgment by default. It can simply make a bad recommendation more convincing.

Over-reliance is harder to catch than aversion because the team looks fast, aligned, and confident. Right up until a wrong answer ships.

The evidence that should reframe your AI strategy

Put the two failure modes together and you get the central result every leader funding an AI transformation should know.

Vaccaro, Almaatouq and Malone (2024) published a preregistered meta-analysis in Nature Human Behaviour covering 106 experimental studies and 370 effect sizes. The headline finding:

On average, human-AI combinations performed significantly worse than the best of the human or the AI alone (Hedges' g = −0.23).

Read that carefully, because it is easy to misread. It does not mean AI is useless. The same analysis found human-AI teams did beat humans working alone (g = 0.64). The authors' precise term for this is augmentation without synergy: the combination helps the person, but the combination is still worse than just deploying whichever party — human or AI — was stronger for that task.

In plain terms: most organizations are leaving value on the table not because their people lack AI, and not because the AI is bad, but because the hand-off between the two is miscalibrated. The team overrides the AI when it should defer, and defers when it should override.

The meta-analysis also points straight at where to act. Performance losses clustered in decision-making tasks. Performance gains clustered in content-creation tasks. And the pattern of who-should-lead was clear: when the human was stronger than the AI, the combination produced gains; when the AI was stronger, the combination produced losses — because people kept inserting weaker judgment into a process the AI was handling better.

That is not an argument against human-AI collaboration. It is a precise specification of the engineering problem: calibration is the variable, and it is addressable.

Trust calibration is a capability, not a personality trait

Here is the genuinely good news, and the reason this is a growth story rather than a cautionary one.

Calibration is not a fixed trait that some employees have and others lack. It is a capability — and capabilities can be designed into workflows and built through practice. Three levers, drawn from the same research base:

1. Make the AI's reliability legible — per task, not in general

Aversion and over-reliance both thrive on vagueness. 'The AI is good' and 'the AI is bad' are both useless. Teams calibrate when they know where the tool is strong and where it is weak: which task types, which data conditions, which edge cases. Map your AI's competence to your actual workflow and make that map visible. Notably, Bansal's work suggests that simply showing the AI's confidence can support calibration better than elaborate explanations that mainly increase persuasion.

2. Give people authority to modify, not just accept or reject

A follow-up to the original algorithm-aversion work found something practical: people were far more willing to use an imperfect algorithm — and performed better — when they could adjust its output, even slightly. Binary 'accept the AI or don't' interfaces push teams toward one failure mode or the other. Workflows that let people keep agency and edit, override, or refine the AI's contribution produce more calibrated reliance and more durable adoption.

3. Train calibration with feedback, the way you'd train any judgment

People calibrate by seeing the consequences of their reliance decisions. When did deferring to the AI pay off? When did overriding it save you? Without that loop, the team learns from anecdotes — one vivid error becomes 'never trust it,' one lucky save becomes 'always trust it.' Build the loop deliberately: review where AI-assisted decisions landed, separate the AI's contribution from the human's, and treat miscalibration as a learning signal rather than a verdict on the tool.

None of these are technology problems. They are work-design and capability-building problems — which is exactly why they are solvable, and exactly why they are usually skipped.

How to know if your teams are calibrated

Calibration is measurable. A few signals worth instrumenting:

• Override accuracy. When people override the AI, are they usually right? When they accept it, was it usually correct? Low accuracy in either direction is miscalibration.
• The aversion fingerprint. Usage that spikes at launch and decays over the following months — especially after a visible AI error — is the signature of algorithm aversion, not normal 'novelty wearing off.'
• The over-reliance fingerprint. AI-assisted decisions that ship with near-zero edits, revisions, or recorded disagreement. Frictionless is not the same as good.
• Task-type fit. Are you pushing AI hardest into decision-heavy work, where the meta-analytic evidence shows losses cluster — instead of content-creation work, where gains cluster?

If you can't answer these, you are flying an AI transformation without instruments. Adoption dashboards count whether people use AI. Calibration is about how well — and that is the number tied to value.

The takeaway

The story of AI in organizations is often told as a contest: human versus machine, with one eventually winning. The evidence tells a quieter, more demanding story. The winner is neither the human nor the AI alone — it is the team that has learned to hand work back and forth at the right moments.

That skill is trust calibration. It sits directly downstream of psychological safety: a team has to feel safe enough to use AI honestly before it can learn to use it well. And it is the difference between 'augmentation without synergy' — the disappointing average — and genuine human-AI complementarity, where the team really does beat both of its parts.

The organizations pulling ahead aren't the ones with the most advanced models. Everyone has access to those. They are the ones treating calibration as a core capability — designed into workflows, measured like any other performance metric, and built through practice.

Sources

• Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126.
• Dietvorst, B. J., Simmons, J. P., & Massey, C. (2018). Overcoming algorithm aversion: People will use imperfect algorithms if they can (even slightly) modify them. Management Science, 64(3), 1155–1170.
• Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90–103.
• Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8, 2293–2303.
• Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M. T., & Weld, D. (2021). Does the whole exceed its parts? The effect of AI explanations on complementary team performance. CHI '21.
• Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80.
• Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253.