Pentagon AI Deals Sidestep Mass Surveillance Restrictions

The Pentagon has signed agreements with eight major AI companies—including SpaceX, OpenAI, Google, NVIDIA, Microsoft, Amazon Web Services, and Oracle—to deploy advanced AI capabilities on classified military networks. These May 2026 deals came after a public clash with Anthropic, which refused to remove contractual restrictions prohibiting its Claude language model from being used for mass domestic surveillance and fully autonomous weapons.

When Anthropic declined to permit "any lawful use" of its technology, the Defense Department moved to blacklist the company from future contracts, prompting a lawsuit. Other AI firms accepted the Pentagon's terms, with some asserting their agreements already ban mass domestic surveillance.

Yet Pentagon officials maintain no additional restrictions are necessary because existing laws and policies already prohibit mass surveillance of Americans. This position rests on a critical assumption: that commercially purchased data and foreign intelligence collection don't count as mass surveillance—even when they capture vast quantities of information about U.S. citizens.

Why it matters

The distinction matters because the Defense Department routinely purchases detailed records of Americans' movements, web browsing, and associations from commercial data brokers without warrants. Large language models can now analyze this information at unprecedented scale, automatically assembling scattered data into comprehensive profiles of individuals' lives, beliefs, and associations. Under the current administration's expanded interpretation of foreign intelligence authorities—including designating domestic civil society groups as having foreign ties—these AI-powered surveillance capabilities could be redirected toward political opponents.

The commercial data loophole

Following Edward Snowden's 2013 revelations about bulk phone metadata collection, Congress reformed and eventually shut down the NSA's Section 215 program. The Supreme Court's 2018 Carpenter decision established that the government needs a probable cause warrant to compel phone companies to disclose location data that creates a "detailed chronicle of a person's physical presence."

The Defense Department has circumvented these protections by purchasing similar information on the open market. Disclosed documents show U.S. Special Operations Command bought location data from a Muslim prayer app with 98 million downloads. The Defense Intelligence Agency receives commercially available geolocation metadata from smartphones, including U.S. devices. Multiple military branches have purchased access to databases containing over 100 billion daily records of global internet traffic, sometimes including browsing history and communication contents.

The Pentagon's position is that purchasing this data differs from compelling its production, so warrant requirements don't apply. Congressional efforts to close this gap, including the Fourth Amendment Is Not For Sale Act, have stalled despite bipartisan House support.

How LLMs amplify the risks

Pre-AI analytical platforms like Palantir's Gotham could link datasets by matching identifiers such as names or phone numbers. Large language models go further: they can identify individuals from writing style and content even without explicit identifiers. Recent studies show LLMs can match pseudonymous accounts to public profiles and re-identify participants in redacted interview datasets.

These models also enable mass profiling at scale. Rather than manually categorizing information, LLMs can directly process unstructured material and extract facts or infer attributes. An analyst can query the system with characteristics—political ideology, likelihood to protest—and identify everyone in a dataset matching that profile. Studies confirm LLMs can accurately infer political views and demographic attributes from text that doesn't explicitly disclose them.

The technology's confidence can be misplaced. OpenAI's own reports warn its models "can be confidently wrong," with research showing LLMs overestimate their accuracy by 20-60%. Yet automation bias may lead analysts to trust system outputs, particularly when institutional incentives favor acting on potential threats over questioning false positives.

Inadequate safeguards

Existing rules don't address these risks. Intelligence activities operate under Executive Order 12333, which authorizes broad foreign intelligence collection. The Trump administration has expanded what qualifies as foreign-linked activity, issuing National Security Presidential Memorandum 7 to investigate domestic groups for purported foreign ties and designating organizations as terrorist entities to justify surveillance of U.S. groups with any connection to them.

The 2024 Policy Framework for Commercially Available Information requires heightened protections for "sensitive" data but leaves each agency to define sensitivity, including contested questions about what constitutes "substantial" volumes of Americans' information or "pattern of life" data.

These details were first reported by Faiza Patel at Just Security.