When Securus Technologies president Kevin Elder announced his company's new AI system trained on years of prison phone calls, he framed it as a breakthrough in crime prevention. The system, now piloting in multiple correctional facilities, scans inmate calls, texts, and emails in real-time, flagging conversations that might indicate planned criminal activity. On the surface, it sounds like a technological solution to an age-old problem. But the truth is more complicated—and more troubling.

The Promise Versus The Reality

Securus, which provides telecommunications services to over 3,500 correctional facilities across the United States, has access to an unprecedented dataset: millions of hours of recorded conversations between inmates and their contacts outside prison walls. The company began building its AI tools in 2022, training models to recognize patterns in language, tone, and conversation structure that might indicate criminal planning.

"Our goal is to prevent crimes before they happen," Elder told MIT Technology Review. "If we can identify someone planning a retaliation shooting or a drug deal while they're still incarcerated, we can alert authorities and potentially save lives."

But here's where the narrative breaks down. The system operates on a fundamental assumption: that patterns from past criminal communications can reliably predict future criminal behavior. This assumption ignores critical realities about prison communications and human behavior.

The Feedback Loop Problem

The AI was trained exclusively on prison communications—conversations that already exist within a surveillance context. Inmates know their calls are monitored, which fundamentally changes how they communicate. They use coded language, speak in metaphors, or avoid sensitive topics altogether. The training data is inherently distorted.

"You're training an AI on data that's already been filtered through the awareness of surveillance," explains Dr. Sarah Chen, a computational linguist who studies prison communications. "The model learns patterns from people who are trying to hide their intentions. When it's deployed, it looks for those same hiding patterns. You create a feedback loop where the system becomes increasingly sensitive to evasion tactics rather than actual criminal intent."

This creates a dangerous escalation. As the AI gets better at detecting coded language, inmates develop more sophisticated codes. The system then trains on those new patterns, becoming more sensitive. The result isn't better crime prediction—it's an arms race of communication obfuscation.

The Constitutional Gray Zone

Prison communications exist in a legal gray area. While inmates have reduced privacy rights, they still retain some constitutional protections. The Fourth Amendment prohibition against unreasonable searches applies, though courts have granted corrections officials broad latitude.

Securus's AI system operates continuously, scanning every communication without individualized suspicion—a practice that would likely be unconstitutional if applied to the general public. The company argues this is justified by the unique security needs of correctional facilities.

But legal experts are concerned about the precedent being set. "What happens when this technology inevitably 'graduates' from prisons to other contexts?" asks constitutional lawyer Michael Rodriguez. "We're already seeing similar systems proposed for schools, public housing, and probation programs. The prison becomes the testing ground for surveillance technologies that eventually migrate to broader society."

The Accuracy Myth

Perhaps the most dangerous misconception is the assumption of accuracy. Elder claims the system has "high precision" in identifying criminal planning, but provides no verifiable metrics. Independent researchers who have studied similar systems report false positive rates between 15-40% for emotion detection in speech—and emotion detection is far simpler than predicting criminal intent.

Consider the implications of a false positive. An inmate's conversation about a fictional crime in a movie script, a metaphorical discussion about "taking care of business," or even a heated argument with a family member could trigger an alert. That alert goes to corrections officers, potentially resulting in disciplinary action, loss of privileges, or extended incarceration.

"The system creates its own reality," says Chen. "Every false positive reinforces the belief that criminal planning is everywhere. Corrections staff start seeing threats in ordinary conversations. The prison environment becomes more paranoid, more punitive, and ultimately less rehabilitative."

The Rehabilitation Paradox

This brings us to the central contradiction of prison surveillance AI: systems designed to increase security may actually undermine rehabilitation—the very thing that reduces recidivism.

Successful reintegration requires maintaining healthy relationships with family and community. But when every conversation is scanned by an AI looking for criminal patterns, inmates self-censor. They avoid discussing employment opportunities that might sound suspicious ("I know a guy who needs some work done"). They hesitate to talk about neighborhood dynamics that could be misinterpreted as gang activity. They steer clear of emotional conversations that might register as "agitated" or "planning."

"What we're seeing is the chilling effect on communication," says Rodriguez. "Inmates stop using the phone system for anything but the most banal conversations. They lose connection with their support networks. And we know from decades of research that strong family and community ties are among the strongest predictors of successful reentry."

The system may prevent some crimes, but it likely creates more future criminals by isolating inmates from the very relationships that help them stay out of prison.

What Comes Next

Securus is currently piloting the system in several states, with plans for nationwide deployment. The company is also exploring applications beyond prisons, including for individuals on parole or probation.

But the real story isn't about this specific implementation—it's about the pattern it represents. We're entering an era where AI surveillance is becoming normalized in contexts where oversight is minimal and the subjects have limited ability to protest.

Several states have introduced legislation requiring transparency about algorithmic systems used in corrections, but none have passed comprehensive regulations. The Federal Communications Commission has jurisdiction over prison phone systems but hasn't addressed AI surveillance specifically.

The Path Forward

If we must use AI in corrections, several safeguards are essential:

• Independent validation: Accuracy claims must be verified by third-party researchers with access to the system and data.
• Transparency: Inmates should know exactly what the system monitors and how alerts are generated.
• Appeal mechanisms: There must be clear processes for challenging false positives.
• Sunset provisions: Systems should require regular reauthorization based on demonstrated effectiveness.
• Rehabilitation metrics: Systems should be evaluated not just on crimes prevented but on their impact on recidivism rates.

The uncomfortable truth is that AI crime prediction in prisons creates the illusion of control while potentially making the underlying problems worse. It treats symptoms while ignoring causes, focuses on containment rather than rehabilitation, and prioritizes technological solutions over human ones.

As this technology spreads, we must ask difficult questions: Are we building systems that make prisons safer, or systems that make prisons more efficient at containing people? Are we preventing crime or simply relocating it? And most importantly, are we creating a future where surveillance becomes the default solution to social problems—starting with those who have the least power to resist?

The reality is that no AI can solve the complex human problems that lead to crime. What it can do—and what Securus's system appears to be doing—is create new problems while convincing us we've found a solution. That's a dangerous misconception, and one we can't afford to believe.