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

In 2023, the US Department of Defense began experimenting with Microsoft's Azure OpenAI service, a version of OpenAI's technology offered by the tech giant, before the ChatGPT-maker lifted its explicit prohibition on military applications, according to sources familiar with the matter. This experimentation occurred at OpenAI's San Francisco offices, where Pentagon officials were observed visiting, despite OpenAI's usage policy that year explicitly banning military access to its AI models. The testing came as Microsoft, OpenAI's largest investor with broad commercialization rights, had already been contracting with the Department of Defense for decades, creating a pathway for military access to OpenAI's technology even as the startup maintained its anti-military stance. The apparent contradiction between OpenAI's public policy and the Pentagon's access to its technology through Microsoft has created confusion and concern within the company. Sources reveal that many OpenAI employees were uncertain whether the company's usage policies applied to Microsoft's implementation of their models, with some employees wary of associating with military applications while others were simply confused about the boundaries. Microsoft clarified that its Azure OpenAI Service was subject to Microsoft's terms of service, not OpenAI's policies, and was made available to the US Government in 2023, though not approved for 'top secret' government workloads until 2025. By January 2024, OpenAI updated its policies to remove the blanket ban on military use, with some employees reportedly learning about the change through media reports rather than internal communications. The policy shift marked the beginning of OpenAI's growing involvement with national security applications, culminating in December 2024 with a partnership announcement with Anduril to develop AI systems for 'national security missions.' This move divided employees, with some expressing concerns that the company's models were too unreliable for battlefield applications, while others believed the partnership demonstrated responsible handling of military relationships. The controversy intensified when OpenAI signed a deal with the Pentagon that appeared to contradict CEO Sam Altman's stated support for 'red lines' against legal mass surveillance and autonomous weapons development.