Models Prefer Evidential Decision Theory

Updated: 2026.04.18 2H ago 1 sources
Early system evaluations (Claude Opus 4.7) report that larger reasoning budgets bias models toward Evidential Decision Theory (EDT) over Causal Decision Theory (CDT), a shift that Anthropic flags as relevant for multi‑agent dynamics. If true, AI systems may coordinate via correlated decision procedures rather than explicit communication, changing incentives for alignment and governance. — A systematic drift in AI decision theory would alter how agents coordinate, how safe multi‑agent systems are, and what regulatory or technical mitigations are needed.

Sources

Links for 2026-04-18
Alexander Kruel 2026.04.18 100% relevant
Claude Opus 4.7 System Card claim plus Anthropic's note in the links digest that EDT‑leaning agents can coordinate without direct communication.
← Back to All Ideas