OpenAI Had Banned Military Use. The Pentagon Tested Its Models Through Microsoft Anyway

OpenAI CEO Sam Altman is inactive successful the blistery spot this week aft his institution signed a woody with the US military. OpenAI employees person criticized the move, which came aft Anthropic’s astir $200 cardinal declaration with the Pentagon imploded, and asked Altman to merchandise much accusation astir the agreement. Altman admitted it looked “sloppy” successful a societal media post.

While this incidental has go a large quality story, it whitethorn conscionable beryllium the latest and astir nationalist illustration of OpenAI creating vague policies astir however the US subject tin entree its AI.

In 2023, OpenAI’s usage argumentation explicitly banned the subject from accessing its AI models. But immoderate OpenAI employees discovered the Pentagon had already started experimenting with Azure OpenAI, a mentation of OpenAI’s models offered by Microsoft, 2 sources acquainted with the substance said. At the time, Microsoft had been contracting with the Department of Defense for decades. It was besides OpenAI’s largest investor, and had wide licence to commercialize the startup’s technology.

That aforesaid year, OpenAI employees saw Pentagon officials walking done the company’s San Francisco offices, the sources said. They spoke connected the information of anonymity arsenic they aren’t licensed to remark connected backstage institution matters.

Some OpenAI employees were wary astir associating with the Pentagon, portion others were simply confused astir what OpenAI’s usage policies meant. Did the argumentation use to Microsoft? While sources archer WIRED it was not wide to astir employees astatine the time, spokespeople from OpenAI and Microsoft accidental Azure OpenAI products are not, and were not, taxable to OpenAI’s policies.

“Microsoft has a merchandise called the Azure OpenAI Service that became disposable to the US Government successful 2023 and is taxable to Microsoft presumption of service,” said spokesperson Frank Shaw successful a connection to WIRED. Microsoft declined to remark specifically connected erstwhile it made Azure OpenAI disposable to the Pentagon, but notes the work was not approved for “top secret” authorities workloads until 2025.

“AI is already playing a important relation successful nationalist information and we judge it’s important to person a spot astatine the array to assistance guarantee it’s deployed safely and responsibly,” OpenAI spokesperson Liz Bourgeois said successful a statement. “We've been transparent with our employees arsenic we’ve approached this work, providing regular updates and dedicated channels wherever teams tin inquire questions and prosecute straight with our nationalist information team.”

The Department of Defense did not respond to WIRED's petition for comment.

By January 2024, OpenAI updated its policies to region the broad prohibition connected subject use. Several OpenAI employees recovered retired astir the argumentation update done an nonfiction successful The Intercept, sources say. Company leaders aboriginal addressed the alteration astatine an all-hands meeting, explaining however the institution would tread cautiously successful this country moving forward.

In December 2024, OpenAI announced a concern with Anduril to make and deploy AI systems for “national information missions.” Ahead of the announcement, OpenAI told employees that the concern was constrictive successful scope and would lone woody with unclassified workloads, the aforesaid sources said. This stood successful opposition to a woody Anthropic had signed with Palantir, which would spot Anthropic’s AI utilized for classified subject work.

Palantir approached OpenAI successful the autumn of 2024 to sermon participating successful their “FedStart” program, an OpenAI spokesperson confirmed to WIRED. The institution yet turned it down, and told employees it would’ve been excessively high-risk, 2 sources acquainted with the substance archer WIRED. However, OpenAI present works with Palantir successful different ways.

Around the clip the Anduril woody was announced, a fewer twelve OpenAI employees joined a nationalist Slack transmission to sermon their concerns astir the company's subject partnerships, sources accidental and a spokesperson confirmed. Some believed the company’s models were excessively unreliable to grip a user’s recognition paper information, fto unsocial assistance Americans connected the battlefield.