FBI Says AI Cuts Threat Response to Hours, While Also Seeking Predictive AI for the Terror Watchlist

The FBI is feeding tips through artificial intelligence to figure out which ones matter most, and Assistant Director Christopher Raia says it's working. He told CBS News's Anna Schecter this week that AI has cut the time between a credible tip and field action from weeks down to hours, sometimes minutes.

"We've sprinkled in AI and now these tips rise to the top and we can get them out to the field to action," Raia said. He's a 23-year FBI veteran who currently serves as an Assistant Director within the bureau's Counterterrorism Division.

The timing lines up with a steep drop in violent crime. The FBI's annual crime report, released in August, showed the murder rate hit a 20-year low in 2025 and violent crime overall posted its largest one-year decline on record. Aggravated assaults fell 7.2%, rapes dropped 7.6%, and robbery fell 18.5%, according to the bureau's own data.

Raia doesn't credit AI alone. He pointed to a "combination of things": state and local law enforcement, the border "being sealed up," and what he called a renewed focus on violent crime.

How the triage actually works

The FBI's National Threat Operations Center fields hundreds or thousands of tips a day from around the world. Before AI got involved, Raia said, the volume outran what human analysts could handle. "The FBI and other agencies were sitting on a bunch of risk," he said. "We didn't have the manpower to analyze it."

Now algorithms flag the most urgent tips so agents can move fast. Raia cited the case of Zachary Charles Newell of Newport, North Carolina, who posted online threats to "shoot up a black preschool" and harm children. AI bumped the tip to the front of the queue, and Newell was arrested quickly. He was sentenced in March to two years in federal prison. Without AI, Raia said, the case could have sat for days or weeks.

Raia also defended Flock Safety's automated license plate readers, which the company supplies to roughly 6,000 law enforcement agencies and which share data with federal authorities. "Those are vital in our investigations as far as leads and as far as finding folks," he said. Privacy advocates have raised concerns for years about how that plate-reader network gets used and who gets access to the data, though Raia didn't address those concerns directly in his comments to CBS.

Asked about civil liberties, Raia said the bureau has been transparent with Congress and oversight boards, pointing to nearly 50,000 pages of documents turned over for review in the past year and a half. "Somebody has to police the police, right?" he said. "We are not taking away the privacies or the civil liberties of Americans, whether it's AI or any of the other responsibilities that we have as an agency."

A bigger, separate AI push: the terrorism watchlist

While Raia was talking triage tools for tips, the FBI has also been shopping for something more ambitious. A procurement request posted on SAM.gov on March 27, 2026, and reported by Military.com, shows the bureau's Threat Screening Center, the office that runs the government's consolidated terrorism watchlist, wants AI capable of predictive modeling across federal databases.

The request is only a request for information, not a purchase or a deployed system. But its specifications are notable. The bureau wants software that can search records across separate government systems, summarize them in plain language, and, most significantly, use "predictive modeling using enhanced data with traceable lineage" to flag where investigators might find more relevant information once new data comes in.

Military.com's reporting is careful to note the document doesn't say an algorithm would independently declare someone a terrorist or place them on the watchlist. It's described as a tool to help human analysts find connections, not a machine that hands down verdicts.

That distinction matters. Critics of predictive policing tools argue that once an algorithm flags a statistical correlation, investigators may end up treating that correlation as evidence of dangerousness, even without a stated rule change to that effect. And people flagged on the watchlist typically aren't told a match occurred when it happens during something like a traffic stop, according to Military.com, which makes the system hard to challenge or audit from the outside. That's a fair concern to raise about a system that already operates with limited public transparency, regardless of whether AI makes it faster or slower.

The procurement request also cites Homeland Security Presidential Directive 6 and NSPM-7, a 2017 directive on sharing threat identifiers. Military.com reported that the Trump administration issued a new memo under the same NSPM-7 label in 2025, directing agencies to investigate domestic terrorism and organized political violence tied to anti-capitalism, anti-Christianity, and positions on migration, race and gender. The procurement document itself doesn't specify which version of NSPM-7 it's invoking or reference those categories directly, according to Military.com's own reporting, so the link between the predictive-AI request and that broader 2025 directive is not fully established by the document.

Agencies have also been watching anti-AI and anti-data center activism for signs of what officials call "anti-tech violent extremism," per Military.com, another category with no fixed legal definition yet.

The bigger adoption trend

None of this is happening in isolation. A market-research estimate cited by Police1 puts the U.S. "AI in Law Enforcement" market at roughly $3.5 billion in 2024, with 7% annual growth projected. A 2025 Police1 survey found 90% of law enforcement professionals support AI adoption, and nearly two-thirds think it'll make their work more efficient.

That support is real, and the crime numbers Raia cited are real too. What's still unresolved is whether the predictive watchlist tool the FBI is scoping out ever gets built, what guardrails would govern it if it does, and whether Congress gets a real look at it before it's deployed rather than after.