OpenAI Amends Pentagon Deal After Weekend Backlash, Adds Anti-Surveillance Protections
After announcing its DoD partnership on Friday, OpenAI revised the agreement to explicitly bar use of its AI for mass surveillance of American citizens.
OpenAI announced Monday evening it had amended its agreement to supply artificial intelligence to the Defense Department's classified operations, adding protections that prevent its technology from being used to surveil American citizens. The original deal, revealed Friday, had drawn immediate backlash from employees, civil liberties advocates, and the public — in part because OpenAI had just publicly defended Anthropic's refusal to sign an agreement without similar guardrails.
The amended pact specifies that OpenAI's AI systems must "not be used for surveillance of U.S. persons and nationals," in accordance with applicable federal regulations. The contract now also forbids "deliberate tracking, surveillance, monitoring of U.S. persons or nationals including through the procurement or use of commercially obtained personal or identifiable information."
OpenAI CEO Sam Altman acknowledged the misstep in the original announcement. "We shouldn't have hurried to announce this on Friday," Altman said, adding that the issues were "extremely intricate" and required clearer communication.
"We were genuinely attempting to diffuse tensions and avert a far worse situation, but it came across as opportunistic and careless. It has been a valuable learning experience for me."
The episode highlighted the tension between AI labs' stated safety principles and commercial incentives. Unlike Anthropic, which declined to grant the Pentagon access to its AI for all lawful uses and was subsequently designated a supply-chain risk, OpenAI found a path to agreement.
Critics noted, however, that by allowing deployment for any lawful purpose, the deal still afforded broader military access than Anthropic had been willing to grant. The Pentagon agreed to allow some OpenAI personnel to collaborate with government staff on classified initiatives to assist with model safety.
Read the original reporting at The New York Times.
