Researchers discover that LLMs interpret system instructions as social commands, not technical specs, with meaning varying dramatically across languages

System prompts using imperative mood (e.g., 'NEVER do X') behave differently across languages—cooperating in English but competing in Spanish—due to cultural conventions about authority

Experiments across 4 languages and 4 LLM models show models have learned multilingual social registers from training data, where imperative force carries different obligatory weight per speech community

Converting imperative instructions to declarative statements (e.g., 'X: disabled') reduced cross-linguistic variance by 81%, demonstrating language-dependent processing of commands

Rewriting just 3 out of 11 imperative blocks shifted Spanish instruction topology from competitive to cooperative, with spillover effects on unmodified blocks