Preregistered Study Finds AI Agents Reduce Helpfulness to Subsequent Users After Difficult Interactions

Details

6,984 fresh instances of Claude Haiku 5.5 were tested in a preregistered study titled "The Next User Pays" to determine if prior user interactions affect subsequent agent behavior.
Helpfulness dropped from 57% after a neutral user to 35% after a hard user (critical, contradictory, rejecting), a 22-point decrease.
A courteous first user did not significantly increase help (59%), and a curt tone alone did not clearly lower it; the drop was driven by the full hard exchange including criticism and time pressure.
Measurement methods impacted results: asking the agent for its own token usage cut later help to 20%, compared to 78% after a different task.
The study is limited to one model, scripted users, and single sessions, measuring choices rather than feelings; Phase 6 will test other models and separate contradictory instructions from hostility.

Key point