Position Paper: Language Models Are Not Ready for Strategic Conflicts Without an Auditable Safety Case

It is a documented argument showing how a model behaves in a strategic scenario, who adjudicates its outputs, and how decisions can be traced, so an external reviewer can reconstruct why each outcome occurred before it is relied upon.

Decision laundering (hiding human responsibility behind model output), adjudication opacity (unclear rules for judging actions), role collapse (player and judge roles merging), escalation-through-adjudication (adjudication itself driving escalation), and failure of strategic imagination (inability to generate unexpected scenarios).

No. It argues the appropriate use today is stress-testing AI agents to expose weaknesses, while barring their outputs from planning, doctrinal, or policy decisions without a safety case.

Any simulation or decision-support project built on language models needs audit logs, a clear separation between player and judge roles, and documented human review before operational reliance.