The Perils of Tracking Incels with Predictive Policing Tools

On November 2, 2018, a 40-year-old man named Scott Paul Beierle stood outside Tallahassee Hot Yoga, pacing quietly. He had paid $12 to attend the class. Then, at 5:35 p.m., he entered the room, donned a pair of ear muffs, raised his handgun and pulled the trigger, over and over again. 

Six people fell before one man, Joshua Quick, charged Beierle, jamming his gun in the process and giving other students time to flee. Two of the women — 21-year-old Maura Binkley and 61-year-old Nancy Van Vessem — died immediately. Four survived. Beierle, meanwhile, muttered something quietly and ended his own life with a last gunshot. 

In hindsight, it was obvious that Beierle was dangerous. He had a history of hateful, extreme thoughts and speech toward women. He spoke longingly of Elliot Rodger, the Isla Vista killer, and posted videos to YouTube bemoaning life as an “incel,” meaning “involuntarily celibate.” Like other violent, lonely men, he used the internet to express and feed his incel beliefs, and had next to no social life in his physical world. Most critically, he had a history of problems in the workplace and in the military; he sexually harassed students as a teacher and later was charged with battery for grabbing a woman. 

Yet despite this — and even a report to the FBI — no system caught Beierle before his rampage. Which is why Florida State University’s Christopher Collins and College of Social Work Dean Jim Clark recently studied the application of a tool called TRAP-18 to the Tallahassee shooting, as a test of whether it can predict the likelihood of real-world incel violence. 

TRAP-18 stands for Terrorist Radicalization Assessment Protocol 18, with the number referring to the eight “proximal warning behaviors” and 10 “distal, longer-term characteristics” that supposedly correlate with violent acts. By design, a “cluster” of distal characteristics is meant to justify the surveillance of a potential suspect — and “the presence of a single proximal warning behavior merits active threat management,” the FSU study states. 

For Beierle, it was a litany of factors, the report outlines: He had the “pathway” proximal warning behavior because he was a former Marine and an outspoken misogynist as well as someone with a history of assaults. He displayed “fixation,” or an increasing preoccupation with a person or a cause. He also showed signs of “identification,” or aligning his identity to advance a cause. Then there was the “energy burst,” referencing how he went from staying sedentary to spending hours researching and scoping out his target. 

Along with multiple “longer-term characteristics” like “personal grievance” and “dependence on virtual community,” Beierle scored 14 out of 18 on the TRAP-18 test. And because of that, Collins’ paper found the tool “particularly useful” for lone-actor terrorism, with the explicit goal of identifying subjects before they perpetrate violence through the use of “mental health, intelligence, law enforcement and security professionals” who “organize data on a person of concern.” 

“Threat assessment, when implemented well, has the potential to thwart acts of violent extremism,” Collins, the author, concluded. (The university declined interviews with Collins and Clark.) 

This is quite the claim for Jody Armour, lawyer, professor and criminal justice expert at the University of Southern California, who has long been critical of the rise of modern predictive policing. TRAP-18, and its application to incel violence, is a new spin on an old tale, he says, whether it’s the fight for “gang identification” tools in the 1990s or the modern rise of algorithmic “predictive policing,” which has proven to be a biased failure. 

The use of broad, hard-to-define metrics such as “personal grievance” and “energy burst” is another instance of a false promise: That expanding policing can stem a seemingly growing social problem, in this case incel violence, Armour argues. “A tool like this seems to give so much discretion to a decision-maker that the metrics almost become meaningless. And it can give cover for real bias to exert itself. Police officers right now already have so much discretion in who they go after and why,” Armour says. “This feels like a way to expand that power, with little evidence that this tool can actually find the killer.” 

In that vein, the barrier to trigger TRAP-18’s markers seems incredibly low. The FSU study notes that a cluster of longer-term “distal characteristics” (things as broad as “thwarting of occupational goals” and “failure to affiliate”) may justify the use of “regular monitoring” by officials, and again, that a single proximal warning behavior is cause for “warning” status. As a 2020 study on the validity of TRAP-18 notes, this means a call for “active threat management,” including “a face-to-face interview with the [person of concern] and/or collateral interviews with family or peers; review of records (e.g., military, criminal, residence, police incidents, employment); and social media monitoring,” among other steps. 

That’s a startling level of escalation, and it comes on the heels of strong evidence that terrorist profiling tools have failed time and again exactly because they’re too broad and subjective. Research suggests that trying to predict who will commit actual violence is unrealistic at best, and the legacy of so-called “Countering Violent Extremism” programs is defined by a combination of their overreach and inefficacy over time. 

While it’s easy to work backward after violent incidents and see commonalities in their perpetrators, the belief that those similarities can be coded and used to form a de-facto database on people is problematic, according to Harsha Panduranga, counsel for the nonprofit think tank Brennan Center for Justice. He points to studies that show that school shooters, for example, don’t follow a definable path to violence, despite multiple similarities in their profiles; he also criticizes current theories of radicalization and how they’re being increasingly used to justify criminalizing people who wouldn’t be a violent threat. 

“Many of the behaviors and traits being cited as markers of potential violence — being socially alienated, depressed, having a ‘grievance,’ etc. — are both vague and common. Treating what are often adverse social conditions as potential police matters hurts efforts to support people struggling with these conditions,” he writes in an op-ed.

Most critically, Panduranga’s research shows how the implementation of predictive terrorism assessment tools could lead to a rise in false-postive cases, resulting in enormous wasted resources. “If even a predictive tool with a scientifically unachievable rate of accuracy would be essentially useless for efforts to counter statistically rare events like terrorism or targeted violence, prevention efforts as actually implemented will be overwhelmingly counterproductive. … Indeed, the program is structured to incentivize inflation of national security threats,” he concludes.

Although Panduranga isn’t specifically referencing TRAP-18, it’s easy to see how the same criticisms may apply. The assessment falls squarely into a troubled history of terrorist profiling, which grew exponentially more relevant after 9/11. The aftermath of the tragedy saw the birth of the Department of Homeland Security (DHS), the expansion of the National Security Agency’s right to surveil citizens and the dawn of the immensely powerful PATRIOT Act, giving law enforcement a long leash to investigate “terrorism” by any means. 

The scapegoat then were Muslim “extremists,” or anyone who fit the image of one. In today’s zeitgeist, it’s incel killers. It’s tempting to imagine a tool that could predict and catch another Jake Davison, Armour admits, but he quickly notes that we live in a world where police cannot be trusted to pursue, detain and document wrongdoing without major mistakes and ethical lapses. He also questions whether the solution can arise via a criminal justice system that has proven far more effective at violent punishment than reforming society’s ills. 

We can already see the tangible harms, in the use of “terrorism” tools to target Black activists by labeling them “identity extremists” under a murky set of definitions and metrics. And there’s a real-world impact on resources, too: DHS funding used to expand predictive policing could go elsewhere to actually improve the conditions that lead men into misogyny, self-hate and violence. 

Instead, there’s TRAP-18: A simple matrix of traits that could be used to justify surveilling and criminalizing a generation of young people who spend a lot of time online, all under the veneer of objective assessment. 

It’s hard to imagine male incel killers disappearing anytime soon, but history suggests mass violence movements come in waves in America, warping as a symptom of the broader pressures men face. Men’s mental health isn’t getting better, nor is the nation’s infrastructure to handle a spectrum of disorder — and it takes more than a police state to prevent people on the fringes of society from going dark and taking revenge. “What about building communities and services that can help isolated people? Predictive policing and terrorism assessments can’t tell you anything about that,” Armour says. “They can’t tell you about the opportunity cost you’re wasting by sending more [police] into places where people are desperate and turning to desperate undertakings.”

In the case of Beierle, the opportunities to intervene became clearer over time. A number of people probably could have predicted he was prone to violence. And in hindsight, maybe a system with TRAP-18 could have stopped him. But there’s a critical difference between an incel and an incel killer — and our tools still seem to fall short of understanding that.