← Accueil actu | Salle du marché

Pentagon reportedly sought OpenAI system with fewer refusals, documents show; officials dispute language

2026-09-13 · Source:TechRadar

Pentagon reportedly sought OpenAI system with fewer refusals, documents show; officials dispute language

Leaked records point to defense requests for a model less inclined to refuse, even as officials insist no such language appeared in final agreements.

Documents reviewed in reporting have surfaced a dispute over whether the Pentagon pressed OpenAI for an artificial intelligence system engineered to be less inclined to refuse user requests. According to the records, defense officials floated language seeking a model that would push back less often, a request that has since become the subject of contention. Officials, for their part, have denied that any such phrasing survived into the final agreements.

The disagreement turns on a deceptively narrow technical point. Refusal behavior is one of the most visible guardrails in modern AI systems: models are tuned to decline certain requests as a core safety measure. A push for a system that "doesn't like saying no" — as the reported request has been characterized — would therefore touch directly on how safety tuning is calibrated for a defense customer. Any adjustment of that kind would carry significant weight for how AI vendors structure their government offerings.

What makes the episode notable is the gap between the paper trail and the official record. Documents indicate the requests were made and disputed; officials counter that the language never appeared in the agreements ultimately signed. That divergence leaves open basic questions of accountability: procurement negotiations can produce drafts, talking points and internal memos that differ sharply from the contract text, and the discrepancy between the two versions is precisely where oversight questions tend to arise.

For AI compute buyers and investors, the episode signals sustained defense-sector appetite for model behavior tailored to mission needs, and it highlights how safety postures may become a negotiated contract term rather than a fixed vendor policy. The reported exchange between the Pentagon and OpenAI suggests refusal tuning is now a live subject in government dealings.

As it stands, the records point one way and official denials point the other. Until the underlying documents and final agreements are reconciled publicly, the dispute remains unresolved.

Mots-clés: OpenAI · Pentagon · AI Safety · Military AI · AI Policy


À lire aussi


萬安算交所 AIXX · AIXX · Bourse compute vérifiée · Accéder à la salle