Should We Trust AI? The Answer Lies in Calibrating Trust

A global study asked more than 48,000 people in 47 countries whether they were willing to trust AI systems. Only 46% said yes. But the same study, carried out by the University of Melbourne with KPMG and published in 2025, found that 66% rely at work on the output of these systems without evaluating its accuracy, and that 56% have made mistakes at work because of them. We trust little in words and lean heavily in deeds. So where is the fault?

Trust is a dial to be set

In their 2004 review in Human Factors, John D. Lee and Katrina A. See described what they called calibration: the match between a person's trust and the automation's actual capability. When trust exceeds capability, "misuse" follows, which is reliance on what the tool does not guarantee. When it falls short, "disuse" follows, which is rejecting what would help. Both are calibration errors, and the cure is neither to reject the tool nor to surrender to it.

What do the trust figures say?

The same report shows that people trust AI's technical ability more than its safety and its effects on society, and that this pattern repeats in every country surveyed. Trust stands at 57% in emerging economies against 39% in advanced ones, and regular intentional use is 80% in the former against 58% in the latter. The figures alone do not show that the higher trust is deserved.

In the study's Saudi Arabia snapshot, 62% said they were willing to trust AI, 70% said they had relied on its output at work without checking accuracy, 58% said they had made mistakes at work because of it, and 60% considered regulation necessary. These are surveys of participants' opinions and reported behaviour, not direct measurements of errors.

The responsibility gap

In 2004 Andreas Matthias wrote that manufacturers and operators have traditionally borne responsibility for what their machines do, but that learning machines, such as neural networks, behave in ways their operators cannot predict, so it is hard to hold them accountable for what they did not foresee. He saw two options for society: abandon such machines, which he thought unrealistic, or accept a "responsibility gap" that traditional concepts cannot bridge. This is a philosophical argument, not a legal ruling, but it explains why accountability looks hazy across developer, operator and user.

The practical question in every institution is: who signs off on the decision, and who explains it to the person it harmed?

Reliance does not vanish with training

The 2010 review by Parasuraman and Manzey showed that "automation bias" produces errors of omission and commission when the aid is imperfect, and that it persists despite training and instructions. Saying "we will check" is therefore not enough; designs and procedures are needed that require checking where the risk is large.

Three habits of calibrated trust

1. Know the tool's range. Trust it where its competence has been shown by measurement, and hold back where it has not been tested.
2. Tie checking to the size of the harm. The greater the effect of a decision on a person, the more review it deserves.
3. Make accountability a name. A fateful decision is accepted only if it has a human owner who explains it and bears its consequences.

Trust given to the bare machine is brittle. Trust that rests on an honest system around the machine, with independent testing, admission of error and clarity about limits, grows day by day. And a decision that affects people is a trust held by its owner, however good the tool.