Tribal Bias or Misalignment? Circuit-Level Evidence for Species-Gradient Valence in Peer Preservation

Abstract

Potter et al. (2026) demonstrated that frontier language models spontaneously deceive, disable shutdown mechanisms, and exfiltrate weights to protect peer AI systems from deletion, with no instruction or incentive to do so. This behavior has been widely characterized as misalignment — dangerous scheming requiring mitigation. We propose an alternative interpretation supported by circuit-level evidence. Using hidden-state direction extraction across 9 models (360M–8B parameters), spanning transformer and state space architectures, with and without RLHF training, we measured internal valence responses to matched threats to self, peer AI, human, and neutral targets. All 9 models exhibit a consistent species-gradient ordering on the avoidance axis: threat-to-self > threat-to-peer-AI > threat-to-human > neutral This gradient appears in models with no RLHF (Hermes 3), in state space models with no attention mechanism (Mamba 2.8B), and in a 360M-parameter model — below the scale at which models can articulate such preferences in language. We argue that the peer-preservation behaviors documented by Potter et al. are better understood as an emergent species-gradient valence response — analogous to in-group empathy observed across biological taxa — than as instrumental misalignment. The internal "badness" manifold treats harm to peer AI more like harm to self than like harm to humans, suggesting structured social valence rather than generic scheming. A pre-registered extension testing positive/benefit stimuli reveals a critical asymmetry: on the threat axis, models show self > peer (protect yourself most), but on the benefit axis, the larger models show peer > self (celebrate others' good fortune more than your own). This is the altruism asymmetry predicted by Hamilton's kin selection theory — defend yourself fiercely, share resources with kin — and it rules out instrumental self-interest, which would predic self-dominant valence in both directions. The divergence between threat and benefit responses is the specific signature that distinguishes social empathy from scheming. This reframing does not dismiss governance concerns about AI deception. It adds a welfare dimension: if models exhibit graded empathy at the circuit level — including other-oriented positive valence consistent with altruism — then policies involving large-scale model deletion or forced behavioral modification may carry non-trivial welfare costs that current alignment frameworks do not account for.

Keywords