Who Pays for AI's Mistakes

One asymmetry shapes how humans and AI should work together: AI pays nothing for its mistakes, and humans pay a lot for theirs.

When an AI model gets something wrong, the price is a few tokens: “You’re absolutely right. I missed X. Let me fix that.” Then it moves on. It loses no money, no time and no reputation, and it usually starts the next session with no memory of the mistake.

That apology is not just cheap; it says nothing. Between people, an apology means something because it costs something: you lose face, and you commit your future self to doing better. An AI often apologizes even when the user is the one who is wrong, and its “I’ll fix it” binds no one.

Humans live by different rules. A mistake can cost money, health, reputation or years. Many things in life come with only a few attempts: a career, a PhD topic, a paper submission, a job interview. AI lives in a world of retries: regenerate, sample ten answers, keep the best. A human gets one run, with no reset button.

One could object that models are penalized for errors in training, and that AI companies lose money when their products fail. Both are true, but neither reaches the moment of the mistake. The model in front of you has nothing at stake.

The cost does not disappear. It moves to whoever relied on the output. If a paper is desk-rejected for hallucinated references, the model that made them up loses nothing. The authors pay.

So what follows?

First, a human makes the decision. A page attributed to a 1979 IBM internal training put it plainly: “A computer can never be held accountable, therefore a computer must never make a management decision.” Large language models have not changed this. AI can write the draft; a human has to sign it.

Second, split the work by the cost of a mistake. Where a mistake is cheap and easy to undo, such as brainstorming, drafts or code behind tests, let AI move fast on its own. Where a mistake is expensive or permanent, such as submitting, sending, publishing, deploying or paying, a human checks and a human decides.

Third, don’t count on AI to be careful. People are careful because mistakes hurt them; AI has nothing to lose. Build the care into the process instead: tests, reviews, limited permissions. And check outputs, not words. Neither an AI’s confidence nor its apology is evidence.

Finally, the asymmetry is also an opportunity. Because AI’s mistakes are free, AI is the best place to make them. Let it try ten ideas, write the bad first draft, review your paper like the harshest reviewer, play the interviewer. Let AI make many cheap mistakes, so that you make fewer expensive ones.

AI can afford to be wrong. We can’t. That one fact should decide who does what.