OpenAI has revised its agreement with the United States government over the use of its artificial intelligence in classified military operations. The move follows criticism that the company rushed the original deal and failed to explain how its systems would be used.
On Monday, Sam Altman, the chief executive of OpenAI, said the company would add new restrictions to the contract with the United States Department of Defense. The changes make clear that OpenAI’s systems cannot be used to spy on Americans.
Altman also said intelligence agencies, including the National Security Agency, cannot use OpenAI’s models without further changes to the agreement. Any additional use would require a new modification to the contract.
Altman admitted the company moved too quickly when it announced the deal last week.
“The issues are complex and demand clear communication,” he wrote in a post on the social platform X (social media platform). “We were trying to avoid a worse outcome, but it looked opportunistic and sloppy.”
The controversy began after a dispute between Anthropic and the Pentagon. Anthropic refused to remove a corporate rule that blocks the use of its AI model, Claude AI, for mass surveillance or fully autonomous weapons. The disagreement led the administration of Donald Trump to restrict Anthropic’s technology across federal agencies.
Soon after, OpenAI announced its own agreement with the Defense Department to allow its models to operate on classified government networks. The company said the contract included more safety rules than earlier AI deals with the government.
OpenAI, Anthropic, and the Pentagon, A High-Stakes Shift in the AI Landscape
Still, the news triggered a strong reaction from some users of ChatGPT. According to market intelligence firm Sensor Tower, the number of daily app uninstalls rose by about 200 percent after the partnership became public.
At the same time, Claude climbed to the top of the Apple App Store rankings, suggesting some users switched platforms.
Despite the policy dispute, reports indicate that Claude still appears in some military analysis tools linked to the conflict between the United States, Israel, and Iran. The Pentagon has not commented on the status of those systems.
The debate reflects a wider question: how much power should governments and private companies hold over AI used in warfare?
Modern militaries already rely on AI to manage large flows of information. Systems can scan satellite images, track troop movement, and help analysts sort intelligence data. These tools do not make battlefield decisions on their own, but they help humans work faster.
One company that plays a major role in this area is Palantir Technologies. Its data analysis platforms support intelligence and military operations across the United States and its allies.
Palantir’s defence platform, known as Project Maven, combines satellite imagery, reports, and sensor data into a single system. Analysts can then run commercial AI models through the platform to identify potential threats or targets.
Balancing Combat Speed Against AI Safety Risks
According to Louis Mosley, who leads Palantir’s operations in the United Kingdom, such systems help military teams make decisions faster. In some cases, the tools highlight enemy vehicles or assets on satellite images.
Even so, experts warn that large language models can make mistakes. These systems sometimes generate false information, a problem known in the AI field as “hallucination.”
Military officials say they try to limit that risk by keeping people in charge of the final decision. NATO officers working with Project Maven say a human always reviews AI output before action takes place.
Amanda Gustave, the chief data officer for NATO’s Task Force Maven, has said that AI will not replace human judgment. “There is always a human in the loop,” she explained.
Still, some researchers worry about the direction of the technology. Mariarosaria Taddeo, a professor at University of Oxford, argues that the Pentagon’s split with Anthropic removes one of the strongest voices for strict AI limits.
With fewer companies pushing for firm rules, she says the risk grows that AI tools could move closer to autonomous weapons.
For now, OpenAI says its updated contract aims to set clear boundaries. Whether those safeguards hold as military AI expands remains an open question.
