FBI Explores AI System That Could Predict Future Criminal Threats

A federal procurement document contains a phrase that sounds harmless until you read what sits around it: “Predictive Modeling Using Enhanced Data with Traceable Lineage.”

The FBI wants technology that can compare people against existing government records, identify patterns and similarities, and help analysts find additional information. No one is being arrested by an algorithm based on this document. No contract has been signed.

But the proposal raises a striking question: how far can predictive surveillance go before a pattern becomes a reason to investigate someone?

The FBI Is Looking for More Than Information

The request for information was posted by the FBI’s procurement division in March 2026. According to the source documents, predictive modeling is one of six requirements the bureau is considering.

The language describes software that would work with existing government datasets containing “enriched or enhanced data elements with documented source attribution.”

The system would then analyze new information for “similarity, pattern alignment, and attribute correlation” against existing records.

In plain English, the technology would look for connections between people and information already held by the government.

That distinction matters. The request does not explicitly describe software that identifies someone who is about to commit a particular crime. Instead, it describes a system designed to find patterns that could lead analysts toward additional information.

The Key Phrase Is ‘Pattern Alignment’

Imagine a government database contains records about people who have previously attracted investigative attention.

A new person’s information enters another system. The software compares attributes, associations, records, or other data points against what is already stored.

If the person resembles existing records closely enough, the system could flag the connection for further attention.

That is where predictive surveillance becomes different from ordinary database searching. The machine is not simply retrieving information someone already knows to look for. It is helping decide which connections deserve attention in the first place.

The Watchlist Has Already Expanded Beyond Its Original Name

The FBI office behind the request also has a significant history.

The Terrorist Screening Center was established in 2003. In March 2025, the FBI renamed it the Threat Screening Center.

The change came as the federal government increasingly focused on a broader range of domestic threats and political violence.

The

maintained by the center is approaching 2 million names. According to the source material, fewer than 10,000 of those names belonged to Americans at the beginning of President Donald Trump’s second term.

That means the overwhelming majority of people on the list are foreign.

The concern raised by critics is what happens if technologies originally associated with terrorism screening become increasingly useful for identifying domestic patterns.

The White House Has Given Agencies a List of Behaviors to Watch

The March procurement request landed alongside a broader policy shift.

National Security Presidential Memorandum-7, signed on September 25, 2025, directs federal agencies to address domestic terrorism and organized political violence.

The memorandum identifies several ideological themes it says can animate violent conduct, including “anti-Americanism, anti-capitalism, and anti-Christianity,” as well as extremism involving migration, race, and gender.

The document is explicitly focused on violent conduct.

But another passage asks agencies to identify “behaviors, fact patterns, recurrent motivations, or other indicia” shared by organizations and entities coordinating violent acts.

Those categories are broad enough to be translated into data points.

That does not mean political beliefs automatically become crimes. It does mean the boundaries between political activity, associations, behavior, and investigative interest deserve close scrutiny when automated systems begin connecting them.

One Hypothetical Shows How Quickly the Data Can Accumulate

Patrick Eddington of the Cato Institute examined the potential implications in an analysis published March 31, 2026.

He described a hypothetical woman who attends immigration enforcement protests, donates to a bail fund, posts criticism of ICE, and works with human rights attorneys overseas.

None of those activities, by themselves, establishes criminal conduct.

Eddington’s concern was that AI could connect individually innocuous pieces of information into a larger surveillance profile.

He described the possibility as “the use of AI to assemble individually innocuous queries into an aggregate surveillance dossier, invisible to any single oversight node.”

The example is hypothetical. The underlying authority for FBI assessments is not.

Under the 2008 Attorney General’s Guidelines, an FBI Assessment requires an “authorized purpose” but does not require particular factual predication.

That means an investigation can begin without evidence that a specific crime has already occurred.

The Oversight Picture Has Also Changed

The source material points to two developments that critics say make the issue more significant.

President Trump fired the three Democratic members of the Privacy and Civil Liberties Oversight Board in January 2025, leaving the board without a quorum.

Several months later, FBI Director Kash Patel disbanded the bureau’s Office of Internal Auditing.

The office had been created in 2020 following revelations involving FBI surveillance practices and FISA abuses.

The result is a debate about both the technology and the mechanisms intended to catch mistakes or misuse.

The Government Already Knows Watchlists Can Get People Wrong

There is another reason predictive systems deserve scrutiny: existing government databases are not perfect.

A Government Accountability Office review released in August 2025 found that some U.S. persons had been misidentified as watchlisted individuals or remained on a watchlist after they were no longer warranted.

Between December 2021 and September 2023, Americans submitted roughly 20,000 redress inquiries to the Department of Homeland Security.

Only 289 of those inquiries concerned watchlist issues.

Around one-third of those cases ended with the individual being removed from the watchlist.

That produces a troubling possibility when combined with automated pattern matching. A mistake in one database can potentially become an input into another system.

The Problem Gets Harder When Machines Look for Similarity

Predictive software does not need to understand a person in the same way a human does.

It can identify statistical relationships between records.

That can be useful when investigators are overwhelmed by enormous quantities of information. It can also create problems when correlation is treated as evidence.

The source material points to facial recognition as an example of the broader problem.

NIST tested 189 facial recognition algorithms from 99 developers in 2019 and found significant differences in false-positive rates between demographic groups. In some cases, the differences ranged from 10 to 100 times depending on the algorithm.

NIST has since found much smaller demographic differences in the most accurate systems.

Facial recognition is also different from the predictive technology described in the FBI request.

Still, the example demonstrates a basic weakness of automated identification: accuracy can vary, and a confident machine-generated result can influence the human being reviewing it.

The FBI Says Humans Will Remain Responsible

There is another side to the story.

FBI Director Kash Patel has argued that artificial intelligence can help agents process enormous amounts of information that humans could not realistically examine manually.

He has credited AI-assisted tip triage with helping stop alleged school attack plots in North Carolina and New York.

Patel said, “if we had just humans look at it, we would never sift through them all.”

Those accounts rest on Patel’s description of the cases, according to the source material.

The FBI has also published its own principle for artificial intelligence use: “A human being is ultimately accountable for the actions taken, not an AI.”

That distinction is central.

An algorithm can recommend where investigators look. A human is still supposed to decide what happens next.

Three Questions Will Decide Whether Predictive Surveillance Goes Too Far

The procurement document does not establish that the FBI has built a system capable of predicting who will commit a crime.

It establishes that the bureau is exploring technology capable of finding patterns across government data.

That leaves several practical questions:

• What data goes into the system? The accuracy of any prediction depends heavily on the records used to create it.
• What happens after a person is flagged? A recommendation becomes consequential if it triggers surveillance, questioning, searches, or other investigative steps.
• How can someone challenge an error? Existing watchlist redress problems show why correcting false matches can become a serious issue.

Those questions become even more important when the system is searching for people who resemble records rather than people already connected to a specific crime.

The Eight Words Point to a Much Bigger Debate

Nothing in the March request says the FBI can predict future criminals with certainty.

No contract has been signed based on the material described here, and there is no indication that Americans are currently being arrested because an AI system predicted their behavior.

What exists is a procurement request, a changing domestic security policy, and a government agency interested in technology that can uncover patterns across enormous datasets.

The danger is not necessarily a machine deciding who is guilty.

It is the possibility that a machine quietly decides who looks interesting enough for a human being to investigate.

That is a much smaller step on paper. It could become a very different one in practice.