One failure weakens trust more than one success strengthens it

What it is

A diagnostic assistant gets one case right, and trust steps up a little; it gets one wrong, and the step down is longer. The unit is “one” both times; the weights are not equal. One failure weakens trust more than one success strengthens it.

This is asymmetry in the update rule, not collapse in which one severe miss empties long-built trust, and not the recovery count in which many successes are needed to make up. Here the comparison is one against one, paired.

Why it happens

People are more sensitive to losses than to gains. Handover is handing consequence to the system; failure is a loss, success is “nothing went wrong.” Nothing-went-wrong is hard to encode as evidence of equal strength. The interface further amplifies: failure gets a modal and has to be handled; success walks silently into the next item. Failure is then easier to extract from memory, and the next estimate is pulled harder by it.

Lee and See’s calibration allows the estimate to move with experience; if the step size is larger for failure than for success, after the same number of positive and negative events the net position is biased negative. This can coexist with a streak pushing the estimate over the base rate: the streak is a window with no negative events; once one appears, the asymmetric update shaves extra off the bit just pushed up.

Studying it

In a sequence of known base rate, insert one success and one equally visible failure at symmetric positions, and measure the change in handover and estimate before and after each insertion. Independent variables: whether failure and success are matched in visibility, whether consequence sits in the same band. Dependent variables: step up, step down, the ratio of the two.

Consequence must be in the same band. Comparing a penalty-level failure with a silent success measures severity, not unit asymmetry.

Where it stops holding

If success is designed to be equally visible (“this one was right, on these grounds”), the asymmetry narrows and does not fully vanish. In gambling or games people may chase a streak the other way. A high-stakes incident is collapse; step size is no longer something “one against one” can describe. Generalising a failure in one domain to all capability is the next scope error, not the step size itself.

Applying it

• Do not only make failure loud and success quiet. Give success an equally short visible receipt so positive samples can enter the estimate.
• Right after a failure, give one observable success on a same-band task, so the asymmetric step does not sit in memory alone.
• When measuring, report separately how much handover rises after one success and how much it falls after one failure; do not only report a net trust score.
• Check: after one plus and one minus in the same band, is net handover clearly negative. If yes, the asymmetric update is working; then see whether you only put heavy prompting on the failure side.

Related

• Same group: L5.10.2 Users generalise a failure in one domain to the system's whole capability · L5.10.3 An early failure weighs more than an equivalent late one, because there is no success history to offset it · L5.10.4 How a failure is handled can partly offset the damage; admitting the error beats downplaying it · L5.10.5 Restoring trust takes far more successes than the failures that caused the damage
• Nearby: L5.04 Collapse of Trust · L5.09 Overtrust and Trust Collapse · L5.03 Trust Calibration
• Search terms: asymmetric trust update · negativity bias · loss versus gain in trust