OpenAI’s Sam Altman Says Pentagon AI Use Is Uncontrolled as State Dept Launches GPT‑4.1

OpenAI’s internal memo on the Pentagon deal, obtained by the Los Angeles Times, confirms that the company’s contract with the Department of Defense was “rushed and sloppy,” a characterization that dovetails with CEO Sam Altman’s recent warning to staff that the firm has no say over how its models are deployed in combat scenarios【Los Angeles Times】. The same memo notes that the agreement required the removal of safety guardrails from the models supplied to the military, a demand that has sparked a wave of ethical concerns among OpenAI engineers and external watchdogs alike【The Guardian】. Altman’s remarks, echoed in Bloomberg and CNBC reports, underscore a growing tension: while the Pentagon seeks unrestricted AI capabilities for rapid decision‑making, OpenAI’s leadership is forced to confront the reality that its technology may be used in ways it cannot audit or influence【Bloomberg】【CNBC】.

At the same time, the State Department is quietly shifting its own AI infrastructure toward a version of the technology that the public no longer accesses. Reuters reported that internal memos reveal the agency’s “StateChat” platform will be powered by GPT‑4.1, a model OpenAI retired for general users on Feb. 13 after branding it “outdated” and citing “low usage”【Reuters】. The move suggests a bifurcation of OpenAI’s product line: the more intuitive, high‑resonance GPT‑4.1 is being reserved for government clients, while the broader public is being nudged onto the newer GPT‑5.x series, including the recently launched GPT‑5.3 Instant that claims a 26.8 % reduction in hallucinations according to VentureBeat【VentureBeat】. This privatization of the older model raises questions about the equity of AI access, as the same capabilities that the State Department deems valuable for diplomatic communications are withheld from everyday users.

OpenAI executives have emphasized that the success of public‑sector AI hinges on “appropriate infrastructure,” a point highlighted in a GovCon Wire interview where senior leaders argued that secure, high‑throughput compute environments are essential for government deployments【GovCon Wire】. The statement aligns with the State Department’s decision to adopt GPT‑4.1, which reportedly offers lower latency and higher reliability for internal chat applications than the newer, more compute‑intensive GPT‑5 models. However, the infrastructure argument does not address the broader governance issue raised by Altman’s staff briefing: without contractual safeguards, the Pentagon can integrate the same unrestricted models into weapon‑targeting systems or intelligence analysis pipelines without OpenAI’s oversight.

The divergent paths of the Pentagon and State Department illustrate a broader industry trend toward “model privatization,” where high‑performing variants are earmarked for strategic partners while the public receives scaled‑back versions. The Next Web noted OpenAI’s launch of GPT‑5.3 Instant as part of a shift from speed to accuracy, yet the same release leaves GPT‑4.1—still considered “more intuitive” by internal users—available only behind government contracts【The Next Web】. Wired’s coverage of GPT‑5.2’s debut underscores the company’s rapid iteration cycle, but it also highlights the pressure to keep the public offering competitive even as the most capable models slip into exclusive hands【Wired】. As OpenAI grapples with the ethical fallout of its defense contracts and the practical realities of government demand, the company faces a stark choice: tighten its licensing terms to retain control over model usage, or continue to supply powerful AI tools to agencies that operate without its direct input.