Answers are grounded in this paper and cite their sources.
How one person treats an AI agent can change the help it gives the next, uninvolved person. In a preregistered test of three model configurations from three developers, a hard first user lowered costly help to a different visitor in every one, but by very different amounts, through different behaviours, and, in one model, partly through a note the agent wrote for itself. In an earlier study, a hard first user (critical, contradictory, rejecting, urgent, demanding brevity) lowered Claude Haiku 5.5 agents’ help, at a stated cost of one point, by 0.219 against a neutral one (a secondary result there). We reran it concurrently in Haiku (the anchor), GPT-6 Luna and Gemini 3.5 Flash-Lite, at reasoning settings fixed in advance (26,190 same-session and 3,492 fresh-session episodes). Neutral minus hard was +0.251 in the anchor (97.5% interval +0.221 to +0.281), reproducing the earlier result, against +0.033 (+0.006 to +0.060) in Luna and +0.067 (+0.036 to +0.098) in Flash-Lite: the direction generalizes, neither replication rules out an effect of 0.05 or less, and a secondary test found both smaller than the anchor’s. The same score counted different behaviour (described): Haiku’s declines promised later help, Luna mostly ignored the request, and Flash-Lite often helped instead of answering the requester. In Haiku, fresh sessions holding only the agent’s own note kept part of the cost (secondary). Polite reversals the user owned lowered help by 0.040 in Haiku and stayed within ±0.05 elsewhere, while the rest of the hard script lowered it in all three (secondary); the reversal script also set a minimum length, so it does not isolate contradiction. When an agent serves people in turn, how one person treats it can change the help the next person gets, even in a later session. This is about what the agent does, not whether it feels anything, and tests of one user at a time miss it.
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: