I heard a sentence recently that sounded harmless enough to pass without comment: humans must decide how far they want to trust AI. It sat in my head anyway. The word trust caught hold of me and wouldn’t let go.

Trust is a category that exists between people. It assumes intention on the other side: that the other side means well, or at least is not working against you on purpose. When I trust someone, I am trusting a person who chooses how to treat me. An algorithm makes no such choice. It calculates. That difference is exactly what the word trust hides.

Take a calculator. When it returns the wrong number, nobody says I don’t trust it. They say it’s broken, or I mistyped something. The device has no intention, so trust is simply the wrong category to bring to it. It either works or it doesn’t.

With AI, the word trust gets used anyway, more often every year: we trust AI, AI is trustworthy. Regulators build entire debates on it. Once the word is in place, it drags a whole frame of human expectations in with it.

The emotional register gives this away. Betrayed trust in a person feels like betrayal, while a tool that merely fails reads as a malfunction. Hand a chatbot your trust and let it produce a wrong answer, though, and the reaction settles somewhere closer to I have been deceived than to I have a broken tool, even though nobody set out to deceive anyone.

People usually treat trust as something decided in increments: it rises after a good experience and falls after a bad one. That sounds rational enough. It misses what actually happens when you extend trust to something with no consciousness on the other end.

I see this more and more regularly in companies. Teams start relying on AI output because it looks right, and they stop questioning it because the machine has been right before. What they call trust is usually habituation. The two feel identical from the inside, though the mechanism differs. Habituation builds carelessness across many small confirmations; extending trust to something that decides nothing leaves you exposed the first time it is wrong.

A more accurate question would drop trust altogether and ask how reliable a given tool is in a given context. That question is less elegant and less human than the one about trust, and that is exactly its advantage. It lowers the odds that a person and an algorithm get mistaken for the same kind of thing.

None of this is new in principle. People name their cars and swear at malfunctioning computers; personifying machinery is an old habit. What changes with AI is the consequence of that habit, because the output is language. A sentence sounds like it came from someone, even when nothing on the other side decided to say it.

The most common defence I hear runs like this: trust builds over time as results improve. Small tasks first, then larger ones, and once the system proves itself, it earns more responsibility. That is the exact pattern companies use to onboard a new employee, applied to a system that has no stake in the outcome. Once the same language covers both cases, there is no longer a linguistic difference between a person and a tool.

The question underneath the one about how much to trust AI is different: whether we are about to lose the distinction between a tool and a counterpart altogether, one loose word at a time. Distrust might be the more honest word for this technology. A calculator cannot want anything, so there is no malice to suspect. On the other side a calculation runs, and the question a calculation supports is how reliable it is in the case at hand.