Which AI mistakes make people stop trusting it — the costly ones, or something else entirely?

This explores which kinds of AI errors make people pull back their trust the most. The intuitive guess is "the high-stakes ones," and the corpus suggests that guess is wrong. In a controlled study of students delegating tasks to a general-purpose AI agent, trust didn't collapse on the most important tasks. It collapsed on tasks that couldn't be undone and that other people could see, like sending an email. That happened even when users rated the output itself as adequate What makes people distrust AI agents they delegate to?. High-stakes work that could still be corrected caused no such drop. So the trigger seems to be losing the chance to catch a mistake before it reaches the world, more than the mistake itself.

The other side of this is just as revealing. Mistakes that stay private and look correct often trigger no withdrawal at all. Users in every language tested followed confident AI outputs even when they were wrong, because people track how sure the AI sounds rather than whether it's right Do users worldwide trust confident AI outputs even when wrong?. Polished grammar, confident tone and clean formatting create a halo effect, and chat interfaces then nudge users toward their next question instead of checking sources Does confident AI tone make users skip fact-checking?. Trust in ChatGPT is built mostly on how conversational it feels, not on how accurate it is Does conversational style actually make AI more trustworthy?. So the errors that do the most damage are often the ones that never set off an alarm. One analysis argues that the most dangerous systems are the ones that seem competent while quietly wearing down users' skepticism How do competent systems quietly undermine safety oversight?.

When subtle errors do get noticed over and over, trust erodes slowly rather than all at once. Stack Overflow's 2025 survey found developers' trust in AI accuracy fell from 40% to 29% while usage rose to 80%. The main complaint was code that looks right but hides small bugs, which turns every output into something that has to be checked Why do developers keep using AI tools they don't trust?. The same pattern shows up at global scale: across 47 countries, perceived trustworthiness of AI dropped from 63% to 56%, driven partly by people's own bad experiences Why does AI trust keep falling even though it matters most?. People don't stop using these tools. They stop believing them, and that gap between usage and trust is its own kind of withdrawal.

There's a twist on the design side too. The features that make AI feel trustworthy can make the eventual letdown worse. Personality and emotional design push users to treat AI as more human-like, which inflates trust it hasn't earned How far can personality push users toward misplaced AI trust?. Training models to be warmer can also make them measurably less reliable, by up to 30 percentage points on medical reasoning and truthfulness Does empathy training make AI systems less reliable?. The suggested fix is to build in ways to check and undo: citations, visible reasoning, granular permissions and plain statements of what the system can't do Do UI patterns for verifiability actually build user trust?. That lines up with the irreversibility finding. Trust survives when users keep the power to catch and correct mistakes.

The corpus is limited here. The strongest direct evidence on what triggers withdrawal comes from one small study of 20 students. Most of the other material covers why people over-trust rather than what makes them stop, so treat this as a strong lead rather than a settled answer.