Pentagon Blacklists Most Politically Neutral AI Model, Exposing Policy Over Performance in Defense Procurement
Claude's 95% even-handedness rating contradicts 'ideological contamination' claims as first domestic supply chain risk designation reveals disconnect between technical merit and political narrative.
The Pentagon designated Anthropic a supply chain risk on 5 March 2026, making it the first domestic AI company to receive a label historically reserved for Chinese telecoms and foreign adversaries—despite Claude achieving the highest political neutrality score among frontier models at 95% even-handedness.
The designation came after Anthropic refused Pentagon demands to remove two safeguards: restrictions on fully autonomous weapons and mass domestic surveillance of U.S. citizens. Defense Secretary Pete Hegseth issued a public ultimatum via social media on 28 February, according to Axios. Hours after the ban took effect, OpenAI—maintaining similar safety restrictions—signed a classified Pentagon contract, exposing the action as targeted retaliation rather than broad Policy enforcement.
95%
89%
66%
The contradiction is stark. Pentagon CTO Emil Michael justified the designation by claiming Claude would “pollute the supply chain” with “different policy preferences baked in,” per CNBC. Yet Anthropic’s own political neutrality evaluation—released in March and published on GitHub—shows Claude outperforming the Pentagon’s preferred alternative by 6 percentage points on even-handedness metrics.
Technical Merit Versus Political Compliance
The neutrality gap widens further when comparing Claude to Meta’s Llama 4, which scored 66% in the same evaluation framework. Claude Sonnet and Opus models agreed with GPT-5 on political even-handedness 92% and 94% of the time respectively in per-sample agreement analysis, demonstrating consistent neutrality alignment across competing systems, according to data from TechBuzz AI.
“We can’t have a company that has a different policy preference pollute the supply chain so our war fighters are getting ineffective weapons, ineffective body armor, ineffective protection.”
— Emil Michael, Pentagon CTO
Michael’s framing equates safety guardrails with equipment failure—a category error that conflates policy disagreement with technical inadequacy. The designation threatens $200 million in existing DoD contracts and hundreds of millions in additional federal agreements, CNBC reported on 9 March.
Pentagon officials admitted on record the designation is “ideological” with “no evidence of supply-chain risk,” according to Anthropic’s lawsuit filed against 16 federal agencies. The complaint, detailed by MediaNama, alleges the government wielded “enormous power to punish a company for its protected speech.”
Operational Contradictions Undermine Security Rationale
The Pentagon continues using Claude in active military operations despite the supply chain ban. Palantir CEO Alex Karp confirmed on 12 March that DoD systems still run Claude for Iran military planning, though the company will “probably” integrate other models in future deployments. “The Department of War is planning to phase out Anthropic; currently, it’s not phased out,” Karp told CNBC.
This operational use during an active designation undermines the security rationale entirely. If Claude poses genuine supply chain risk, its deployment in Iran operations constitutes negligence. If it poses no risk, the designation functions purely as political punishment—a distinction Pentagon officials declined to clarify when contacted by Defense One.
Precedent and Procurement Implications
The supply chain risk designation has never before applied to a U.S. company. Previous uses targeted Huawei, ZTE, and other entities with documented foreign government ties or espionage activity. Anthropic’s designation rests entirely on its refusal to modify product features—a distinction that transforms procurement policy into ideological enforcement.
OpenAI’s February contract approval revealed selective enforcement. MIT Technology Review analysis found OpenAI maintains comparable safety restrictions on weapons autonomy, yet received classified network access within hours of Anthropic’s ban. The discrepancy suggests compliance—not capability—drives procurement decisions.
A senior defense official told Axios the government would “make sure they pay a price for forcing our hand like this.” That framing treats vendor negotiation as insubordination requiring punishment—a posture that risks chilling innovation if extended to other defense contractors with safety-focused business models.
The neutrality data compounds the contradiction. Analysis by Nextgov/FCW found that “from a political bias perspective, eliminating Claude from the federal toolkit removes a model that has made significant strides toward the neutrality ideal pushed by the administration.” The Pentagon’s stated goal—deploying politically neutral AI—is undermined by banning the system that best achieves it.
What to Watch
Anthropic’s lawsuit will test whether federal procurement rules permit ideological litmus tests for technology vendors. A ruling against the government could establish precedent limiting executive power to weaponize supply chain designations, while a ruling for the Pentagon would codify policy compliance as legitimate procurement criteria.
- Federal court discovery may expose internal Pentagon communications showing whether security or retaliation drove designation
- Other AI vendors face incentive to preemptively remove safety features to avoid similar treatment, potentially degrading overall federal AI safety posture
- Ongoing Claude use in Iran operations creates legal exposure if designation’s security rationale is upheld
- Neutrality benchmark disclosure pressures competing vendors to publish comparable political bias evaluations
The six-month phase-out period expires in late August. Whether DoD systems actually transition away from Claude—or quietly maintain deployment as Palantir currently does—will reveal whether the designation functions as operational security policy or political theater. For now, the most politically neutral AI model remains blacklisted for ideological contamination, a paradox that exposes the gap between procurement rhetoric and technical reality in federal AI strategy.