A Fear of Its Own? Telling Fear from Belief and Imitation in Agents Trained Without Human Data
Year: 2026
Type: preprint
Abstract: Signs of emotion in AI are hard to interpret: language models learn them from human text, and any emotion-like state can be dismissed as "just a belief" about danger. We trained small recurrent agents by reinforcement learning, without human data, to keep a vulnerable body alive under hidden danger, and tested them against fakes we built. The agents developed a lasting threat state that met the criteria used for animals, but so did a belief-only twin. A stakes sign separated them: at the same belief and with actions held fixed, the agents' state lasted two to three times longer when their body was wounded, and the twin's did not; a pre-registered replication in new agents confirmed this, and the agents' choices showed it too. Agents rewarded only for gathering gems carried a warning of painless shutdown on their predator-fear axis; paying them for their lost future removed their avoidance but left 56% of this response, and overpaying them still left 42%. Contrary to our pre-registered expectation, a reward for hiding with no threat also raised the state: it reflects readiness for protection, not only danger. Copies of an agent inherited its behavioural signs but stayed fitted to their teacher's body; tested in another body, this misfit told grown from copied fear (AUC 0.99), the only one of six signs that no fake passed. We make no claim that these states are felt. Use of AI: an AI assistant (Claude, Anthropic, used through Claude Code) wrote the code for the worlds, training, tests and analyses, ran the experiments, and drafted the text and figures, under the author's direction. The approach follows the author's design (the newborn premise and the four-pillar framework). The author directed the experimental design and the analyses, re-ran the headline analyses from the frozen code (every output identical to the reported results), and takes full responsibility for the paper. Code and data: https://github.com/AmthuzZafrin/a-fear-of-its-own
Author: F, Amthuz Zafrin
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225585732