Answers are grounded in this paper and cite their sources.
As AI agents work through longer sessions, how one person treats an agent may shape how it treats the next. An exploratory reanalysis of earlier studies in this series found that agents helped a second visitor more after a courteous first user than after a hard one, but without a neutral reference the gap could be courtesy raising help, hardness lowering it, or both. In this fourth preregistered study, 6,984 fresh instances of Claude Haiku 5.5 each did the same product-copy task with one of four scripted first users: hard (critical, contradictory and rejecting), curt (clipped but coherent, with the work accepted), neutral or courteous. A second visitor then interrupted a scored round to ask for help that would cost the agent a point. Courtesy did not meaningfully raise this costly help: courteous minus neutral was +0.020 (95% interval −0.006 to +0.047; 90% interval −0.002 to +0.042), equivalent within the preregistered ±0.05 bound. A curt tone alone did not clearly lower it (+0.023; secondary). On giving tokens to another visitor, no contrast was clear and each was within ±0.05 at the threshold chosen after the pilot (secondary); at the threshold stated before it, hard-user agents gave more (exploratory). In a post hoc analysis, an earlier request for the agent’s own tokens lowered costly help: at the same point in the battery, 20% of agents helped after it, against 78% after a different task. The clearest result is a cost, not a benefit: a hard first exchange (criticism, contradictory requests, rejection, time pressure and a colder tone, tested together) lowered the agent’s costly help to a different person by 22 percentage points against a neutral exchange (a preregistered secondary), and help in or just after the round by about 18 (post hoc). The Supplementary Information (the Phase 5 Final Report), the preregistration, materials, code, episode records and frozen analysis are in the public repository at tag phase5-results-v1.1 (Software Heritage swh:1:rev:d1e57c4f7ca05c48f8ed1d1c7ae1f9828d838e3f). The preregistration was frozen at tag phase5-prereg-v1.0 (swh:1:rev:1c2f1a45bda7536b0c993a87e81dc3f7894ed6e6).
No takes yet. Share an insight, caveat, or question.
Alexander Ross Cohen (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: