Here is a puzzle worth thinking about: people trust a calculator completely, but many hesitate to trust a self-driving car. Both are machines. Both are far more reliable than the humans they replace. And yet the trust is distributed very unevenly.
Understanding why that is reveals something important about technology, about risk and about ourselves.
The calculator problem
Start with the calculator, because it is the purest case of trust.
A calculator makes arithmetic errors at a rate far below a human’s. Nobody stops to ask whether the calculator “understands” addition. We trust it because its failure mode is small, contained and checkable — if it is wrong, you will probably notice, and the cost of being wrong is trivial.
This is the key insight: trust is not primarily about accuracy. It is about the shape of failure.
Why autonomy feels different
A self-driving car is, statistically, an incredibly safe driver — safer than humans in many conditions. Yet people’s hearts still race when it takes the wheel.
The reason is that the failure mode is different in kind. If a calculator is wrong, you lose a decimal. If a car is wrong, the consequences can be severe, sudden and out of your control. You are delegating not a small task but a significant portion of your safety to a system you cannot fully inspect.
Humans are not irrational to weigh this differently. The brain treats catastrophic risk differently from routine risk, and it is calibrated — imperfectly, but sensibly — to fear large, uncontrolled failures far more than small, checkable ones.
The visibility factor
There is a second factor in trust: how visible the system’s reasoning is.
A calculator shows its work. A spreadsheet shows its formulas. You can check, follow and verify. A neural network, by contrast, gives an answer without a legible chain of reasoning. You can test it, but you cannot fully understand it — and understanding is part of trust.
This is why explainability matters so much in high-stakes AI. The systems people trust most are not necessarily the most accurate; they are the ones that can show their work when asked.
The familiarity effect
Trust also grows with exposure, and the rate of growth depends on experience.
People who have used a system for years trust it more, even if its performance has not changed. People who have seen a system fail badly trust it less — and that distrust can be persistent and contagious. One widely publicized failure of an autonomous system does more damage to trust than a hundred quiet successes build.
This asymmetry is a fact of human psychology, and it shapes how new technologies are adopted. It is also a reminder that the reputation of a technology is built slowly and can be destroyed quickly.
The control question
A deep and often unspoken factor in trust is control: how much do you retain?
People are more comfortable delegating to a system they can override, interrupt or take back from. The ability to intervene is, in a sense, the ability to trust. Systems that remove the human from the loop — that make intervention impossible — face a much higher bar for trust.
This is why autopilots in aviation coexist with pilots who can take over, and why the most trusted AI applications tend to be ones where a human reviews the output. Keeping a hand on the wheel, even metaphorically, makes delegation feel safe.
The fairness dimension
There is one more layer: trust in a machine is often really trust in the people who built it.
When you trust an algorithm with a consequential decision — a loan, a job, a medical diagnosis — you are trusting the engineers, the data and the governance behind it. If those have been careful and transparent, trust follows. If they have been careless or hidden, trust collapses.
This is why incidents of bias and unfairness in AI are so damaging: they reveal that the human chain behind the machine was flawed, and that erodes trust in the whole category, not just the one system.
The practical lesson
None of this is abstract; it has a practical consequence for how we should approach new technology.
When evaluating a system, ask not just “how accurate is it?” but “what happens when it fails, and can I check, override and understand it?” The systems worth trusting are the ones with small, visible, recoverable failure modes — not the ones with a perfect record and a black box.
And when a technology asks for your trust, remember that you are allowed to inspect the shape of its failure before you grant it.
Trust is earned in layers
There is a practical lesson buried in the psychology of trust: it is built in layers and lost in moments.
People extend trust gradually — small delegations first, larger ones only after the small ones prove out. A calculator earns a lifetime of trust by never failing catastrophically. An autonomous system, by contrast, is asking for a large trust in one leap, which is exactly why so many people hesitate. The remedy is design that allows incremental trust: modes that start supervised, clear displays of confidence, and the ability to shrink the delegation when confidence drops.
Systems that make it easy to trust in layers will be trusted more. Systems that demand a leap will be left waiting — however accurate they are.
We will trust machines when their failures are small enough to live with, visible enough to understand and recoverable enough to survive. The machines that meet that test will earn trust regardless of how they are built. The ones that do not will struggle — no matter how good they get.