Can your AI survive one mistake?

Better tools get forgiven: a model that is right more often keeps its users.

In forecasting tests with a bonus for accuracy, people who watched a statistical model make mistakes chose their own judgment instead, though the model was better: their own forecasts had 15–29% more error. A follow-up study found the way back: when people could adjust the model's forecast, even slightly, they were far more willing to use it.

The same mistake cost a model more trust than it cost a person. A small say in the result was enough to bring people back.

• Let people adjust what the system gives them: edit a suggestion, nudge a forecast, re-rank results. Even a small degree of control helps.
• Design the moment the system is wrong before it happens: acknowledge it, make correction easy, keep the person in charge.
• Set honest expectations about accuracy up front.

• Make the output take-it-or-leave-it.
• Promise a model that never errs. One visible mistake costs more than a modest claim.

• Lab forecasting tasks with a bonus at stake, not everyday products.
• The studies measured whether people chose the model, not how they used it over months.

• JZ-005 · Show the work
• Negotiated Software Loop
• The Trust Stack