Internal FBI and Pentagon documents detail early research to improve AI to aid with identifying or tracking Americans, later folded into the Horus search tool
Drew Harwell / Washington Post : Mastodon: @drewharwell@mastodon.social . Tweets: @aclu and @oknox Mastodon: Drew Harwell / @drewharwell@mastodon.social : New: The FBI actively helped research facial recognition software that could be used to identify people from street cameras and flying drones at distances of more than half a mile away. … Tweets: @aclu : This sweeping government surveillance software is a nightmare for our privacy rights. Lawmakers need to close the door on government abuse of this technology now, before it's too late. https://www.washingtonpost.com/ ... Olivier Knox / @oknox : I am grateful for reporters like @drewharwell who cover the deployment of facial recognition technology in America. https://twitter.com/...
Context & Ripple Effects
The Washington Post's document drop lands mid-fight: the ACLU has been in federal court since 2019 over its claim that the FBI, DOJ, and DEA run a de facto nationwide facial recognition surveillance system, and critics have spent years cataloguing how US law enforcement actually deploys this tech with little transparency. What is new here is the FBI's role shifting from customer to co-developer — internal FBI and Pentagon records show active research into identifying people from street cameras and drones at long range, later folded into the Horus search tool.
The arc matters because the deployment side is no longer hypothetical. A 2025 investigation found police departments in a dozen states making arrests on facial recognition matches alone, with no corroborating evidence arresting suspects identified through facial recognition. The documents connect the lab work to that street-level reality.
First-order effects
- The FBI moves from defendant to evidence problem: the ACLU's existing suit against DOJ, FBI, and DEA gains primary-source documentation that the bureau actively funded identification-and-tracking research rather than merely licensing finished tools.
- Agencies tied to Horus now face disclosure pressure over capabilities most Americans never knew existed — identification from flying drones at distances beyond half a mile.
Second-order effects
- Congressional oversight follows the ACLU's call to act before the technology hardens into policy, putting future procurement of biometric search tools under legislative scrutiny for the first time in a serious way.
- Vendors building similar identification-from-camera systems face a reputational and contracting split: agencies may quiet their programs while cities and states weigh restrictions, fragmenting the buyer base.
Third-order effects
- If the pattern holds — classified research maturing into deployed tools like Horus, then out to local departments — government biometric surveillance becomes an infrastructure layer rather than a series of discrete cases, forcing either comprehensive federal regulation or a patchwork of state bans.
- The documents deepen the accountability gap the ACLU flagged in 2019: when development happens inside classified Pentagon and FBI channels, courts and legislatures learn about capabilities only after deployment, inverting normal oversight sequencing.
The trend: Government AI identification tools are moving from classified R&D to embedded policing infrastructure faster than legal frameworks can constrain them.