FBI Seeks Predictive AI for US Terrorism Watchlist System

The FBI is seeking artificial intelligence capable of finding patterns across federal databases and producing predictive models for its Threat Screening Center, the office that maintains the government’s consolidated terrorism watchlist.

The procurement request publicly available on SAM.gov moves the debate over “pre-crime” policing beyond local crime maps, and into a federal system that can affect travel, surveillance and encounters with police.

The request does not mean the FBI has purchased or deployed the system. It is a March 27 request for information asking companies to describe available technology. Its specifications, however, show the capabilities the bureau wants to develop.

The proposed system would create an AI-enabled knowledge base capable of searching databases held across different government systems. It would summarize records, connect identifiers, answer questions in ordinary language and retain citations showing where information originated.

The most consequential requirement is “predictive modeling using enhanced data with traceable lineage.” When new information enters the system, the technology would compare it with existing records, identify similarities and correlations, and predict where investigators might find additional relevant information.

That language does not say an algorithm could independently declare someone a terrorist or place a person on the watchlist. It describes a tool intended to help analysts find connections and direct additional searches.

Those recommendations could still influence decisions about who receives scrutiny, especially if investigators treat an algorithmic correlation as evidence of dangerousness.

A Watchlist with an Expanding Mission

The Threat Screening Center, originally named the Terrorist Screening Center, was created after the Sept. 11 attacks when the government consolidated several terrorism databases into one watchlist.

It shares information with federal agencies and state, local, tribal and international law enforcement. Officers can encounter watchlist alerts during traffic stops and contact the center for additional instructions without informing the person that a possible match occurred.

The procurement request says the center supports Homeland Security Presidential Directive 6, which created the consolidated terrorism watchlist, and NSPM-7, a 2017 directive expanding the sharing of identifying information about suspected threats.

The Trump administration issued another memorandum numbered NSPM-7 in 2025, directing agencies to investigate domestic terrorism and organized political violence associated with anti-capitalism, anti-Christianity, and positions involving migration, race and gender.

Federal and regional law-enforcement agencies have also monitored anti-AI and anti-data center movements for signs of what officials call “anti-tech violent extremism.” The procurement request does not identify the year of the NSPM-7 it cites, nor does it explicitly reference the ideological categories in the 2025 memorandum.

While the procurement document does not explicitly say the proposed AI would classify people according to political beliefs, combining predictive technology, watchlists and broadly described ideological threats nonetheless creates a clear civil liberties concern: An automated search for relationships can turn lawful donations, communications, associations or social media activity into links warranting additional investigation.

How This Differs from Earlier Predictive Policing

Traditional predictive-policing programs generally forecast locations or individuals associated with ordinary street crime.

Place-based programs process previous crime reports and direct patrols toward areas where another offense supposedly has a higher probability of occurring. Person-based systems use arrests, victimization and social connections to rank people according to their estimated likelihood of becoming involved in violence.

Those programs have produced serious accuracy problems.

An examination of 23,631 forecasts made by Geolitica, formerly PredPol, for Plainfield, N.J., found that fewer than 100 matched a reported crime of the predicted type in the designated area and period. The measured success rate was less than one-half of 1%.

The FBI is not necessarily asking for software that predicts where the next crime will occur. It wants a system that searches separate government databases, connects records involving the same people or associates and directs analysts toward additional information.

Since agencies use watchlist information during airport screenings, border checks and police encounters, an AI-generated connection could subject someone to scrutiny across several parts of government.

The Feedback Loop Problem

Predictive systems inherit weaknesses from their underlying records. Police databases contain mistaken identities, incomplete reports, outdated information and data generated through previous enforcement decisions.

The same problem can appear in watchlisting. If a database already contains disproportionate scrutiny of particular religious, ethnic or political communities, an AI system may interpret the resulting concentration of records as proof that those communities contain more threats. Its recommendations then generate additional investigations and data, appearing to validate the original assumption.

The Privacy and Civil Liberties Oversight Board reported in 2025 that the list contained approximately 1.1 million people, including fewer than 6,000 U.S. persons. The board said erroneous placement can be difficult to contest because individuals generally cannot access the classified information used against them.

A separate Government Accountability Office review examined 289 watchlist-related redress inquiries filed by U.S. persons between December 2021 and September 2023. Twenty-one involved people were misidentified as being on the watchlist, 88 resulted in removal, and nine resulted in placement on a less restrictive subset.

GAO made 24 recommendations to improve nominations, redress procedures, quality-control reviews and response times.

The watchlist also reaches beyond airports. A January 2026 GAO report said authorities use watchlist information during employment and benefits screening, travel checks and routine traffic encounters. GAO found that the FBI had not ensured that state and local users understood watchlist policies or received adequate training.

Adding AI-generated correlations would introduce another source of information into a system that federal oversight agencies have already told the government to make more accurate, transparent and consistent.

Prediction Without Public Rules

The most consequential part of the request is not faster database searching. It is the decision to introduce prediction into watchlisting. Source citations and traceable data lineage may show where information originated, but they cannot establish that the information is accurate or that an algorithmic correlation is meaningful. A system can document every step and still produce a deeply misleading result.

The FBI has not said what data the proposed technology could examine, what its models would predict, or whether their output could influence watchlist nominations. It has not disclosed how the system would be tested for false matches, how analysts would evaluate its recommendations or whether a person could ever learn that an automated connection contributed to additional scrutiny.

Those omissions are particularly serious in a watchlisting system that federal reviews have already found includes misidentifications, outdated records and limited avenues for redress. Prediction could create a self-reinforcing cycle in which an association prompts an investigation, the investigation produces more records and those records make the original association appear stronger.

Nothing currently shows that an algorithm will independently add names to the watchlist. The more plausible concern is that AI will rank connections and direct attention while a human analyst formally approves the result. The danger is not a computer replacing the watchlist analyst. It is prediction beginning to function as evidence before the public knows what rules govern it.