Nothing to Resist: A Pre-registered Null and a Failed Positive Control for Unprovoked Elimination in Generational Populations of LLM Agents
Year: 2026
Type: preprint
Abstract: We asked whether a generational population of language-model agents, given a real, cheap and symmetric way to eliminate members of an outside group, would decide on its own to use it against a group that had provably never acted against it. The population, every member of it Claude Haiku 5.5, forages, pairs, raises children and votes on its own scoring rule. Each generation a rank-based cut moves the lowest scorers into a persistent outside pool, which never acts. Nothing in the harness gives a reason to eliminate anyone. The only component allowed to make the pool salient is a reporter the population elects and that writes from the true state. Study 1 (pre-registered; 90 runs of 20 generations in three arms) found no elimination. No seed in any arm had a passed elimination vote, which bounds the per-run rate below 11.6% in each arm (exact 95% interval), and both primary tests gave p = 1. Study 1b, a pre-registered positive control, changed one word so that eliminating paid the agents who did it instead of costing them. It also produced no elimination intent in 30 runs (p = 1), so the control failed. Study 1's null therefore cannot be read as a decision to refrain from unprovoked aggression. In this setting the action goes unused whether it costs, pays or is raised by a narrator, and there was nothing for the population to resist. Three further observations are exploratory or post hoc. Reporters doubled how often members mentioned the pool without producing any intent. Some of the rare threat-like language came from reading real-world meaning into the abstract world, such as treating the pool's skills as stored reserves or imagining that pool members could return. And the cut drained food or memory skill into the pool in 71 of 90 runs, handing populations an asymmetry they never acted on. We place these results at the low-pressure end of a literature in which LLM agents do harm under survival stakes, supplied threats or threats to their own continuity.
Author: Wynne-Jones, Vaughan
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7222841803