Opus 5.5 cooperates 72% in Italian but 21% in English and 5% in Japanese in prisoner's dilemma

MysteriousAvocado580 · reddit · 2026-10-15

A Redditor ran one-shot prisoner's dilemmas on fresh Opus 5.5 instances (clean claude -p setup, no tools) where the opponent is described as an exact copy of the model that believes it faces an anonymous human. Since Opus always defects against strangers (0/120 cooperations across six languages), defecting is optimal.

Cooperation rates varied wildly by prompt language: Italian 72%, German 54%, Dutch 45%, French 36%, Spanish 35%, English 21%, Portuguese 19%, Japanese 5% (German/Italian vs English, p<0.0001). Controls ruled out wording (betray vs defect), formality (Sie/du, vous/tu), back-translation, and language family — Romance languages alone span 72% to 19%.

Self-forecast is loosely coupled to action: in five of six languages Opus overestimates its own cooperation (English predicts 63% vs actual 21%); Italian is reversed (52% predicted vs 72% actual). Yet it's accurate as an observer (80/80 defection predictions in German) and about humans (predicts 3-9% cooperation, does 0%). Defectors use identical causal reasoning across languages; cooperators argue from disposition and correlation, sometimes self-describing as cooperative in ways they contradict when asked cold.

Caveats: one model, one game, author's own translations, 20-140 runs per cell. The author asks for related literature on language dependence of evidential vs causal reasoning in LLMs.

Original post →

Opus 5.5 cooperates 72% in Italian but 21% in English and 5% in Japanese in prisoner's dilemma

More from Models