NATO personnel may soon be able to rely in part on the AI models behind ChatGPT. OpenAI is currently considering providing the military organization’s public or “unclassified” networks with AI.
A source from the Wall Street Journal confirmed the talks, which could of course turn out without a deal being signed. Initially, it seemed to involve all of NATO’s IT environments, but CEO Sam Altman was misspoken according to a spokesperson for OpenAI. Nevertheless, the company’s thinking remains evident: it is eyeing large, critical organizations in the public sector on multiple continents. After Anthropic decided that the AI requirements from the US Department of Defense did not provide enough “guardrails” for the deployment of the technology, OpenAI soon emerged as an alternative.
Opportunistic and sloppy
Altman quickly admitted that the deal between his OpenAI and the Pentagon came across as “opportunistic and sloppy.” He also said that the safety mechanisms in models such as GPT-5.3 would remain intact in military use, as had been the case for Anthropic’s Claude models.
Negotiations between Anthropic and the Pentagon broke down over the use of AI for mass surveillance and fully autonomous weapons systems. The US government subsequently labeled the company a “supply chain risk.”
OpenAI takes over
In an updated statement on March 2, following the closing of the deal on February 27, OpenAI stated that its AI systems “will not be intentionally deployed for domestic surveillance of U.S. citizens and subjects.” The Pentagon confirmed that AI services will also not be used by intelligence agencies such as the NSA.
OpenAI drew three explicit red lines: no mass domestic surveillance, no fully autonomous weapons systems, and no automated decisions with high consequences. Models are deployed exclusively via cloud infrastructure, with security-cleared employees involved.
Altman acknowledged the necessary reputational damage internally yesterday. “I think this was an example of a complex, but right decision with extremely difficult brand consequences and very negative PR for us in the short term,” he said during a company meeting.
During that same meeting, OpenAI employees were reportedly told that they were not allowed to make operational decisions. Altman stated that the Pentagon wants to utilize the AI company’s expertise for the deployment of AI, but that OpenAI itself should not express an opinion on whether decisions in this regard are good or bad.
NATO exploration ongoing
Whether the NATO deal will ultimately go ahead is still unclear. For now, it is an exploration of contractual possibilities. NATO has not responded to a request for comment. However, a potential deal is a sign that OpenAI is becoming less dependent on its consumer arm. Whereas subscribers have uninstalled ChatGPT on a large scale and switched to Claude, collaborations with the public sector will be much more secure. OpenAI for Government, introduced in June 2025, was initially focused solely on the US, but with NATO it also appears to be gaining traction in Europe.
This will not easily be followed up by national governments in Europe. Sovereignty is incompatible with the unrestrained use of OpenAI’s tooling; significant compromises will be needed around data protection, running on European servers, and a possible exit strategy. These are situations in which OpenAI, like Anthropic, Google, and Microsoft, has the advantage that there is no mature European player operating at the same level. Mistral from France is the only real AI model provider on offer residing in Europe, but in no benchmark (or in most subjective experiences) does it appear to be equivalent to Claude, Gemini, or the most recent GPT models.
Read also: Digital sovereignty feels good, but is it really?