In a move that blurs the line between security and pre-crime, a major U.S. prison telecom provider has begun using artificial intelligence to analyze the private communications of incarcerated individuals. Securus Technologies, which handles phone, video, and messaging services for over 3,400 correctional facilities nationwide, has developed and is now piloting an AI model designed to predict and prevent crimes by scanning calls, texts, and emails. The system, trained on years of historical inmate communications, represents a fundamental shift from reactive monitoring to proactive algorithmic intervention within the justice system.

From Recording to Predicting: The Securus AI Surveillance System

According to MIT Technology Review, Securus President Kevin Elder confirmed the company began building its AI tools in earnest several years ago. The initiative leverages the company's vast archive of recorded communications—a dataset accumulated over decades from inmates who have little choice but to use Securus's expensive services to maintain contact with the outside world. This archive, comprising millions of hours of audio and video, provided the training ground for machine learning models to identify patterns, linguistic cues, and conversational contexts associated with criminal planning.

The pilot program now actively applies this trained model to live and recorded communications. The AI scans for specific indicators that human monitors might miss, flagging conversations for further review by security personnel. While Securus has not publicly detailed the exact "risk factors" the model targets, such systems typically analyze vocabulary, tone, speech patterns, discussed locations, names, and temporal references. The stated goal is straightforward: to stop crimes—including violence, drug operations, or witness intimidation—before they occur by intercepting the planning phase.

The Technical Promise and the Ethical Abyss

On its surface, the technology presents a compelling solution to a persistent problem. Correctional facilities are porous environments where criminal activity often continues via smuggled contraband phones or through monitored systems using coded language. Human monitors cannot possibly review all communications in real-time. An AI that can triage this flood of data, highlighting the tiny fraction of conversations that warrant human attention, seems like an efficiency breakthrough. It promises enhanced safety for correctional officers, the public, and even other incarcerated individuals.

However, this promise rests on a foundation of profound ethical and technical quicksand. The core problem is the training data itself. An AI trained exclusively on the communications of a prison population is learning from a dataset inherently skewed by systemic biases. The U.S. incarcerates people of color at disproportionately high rates; poverty is a overwhelming predictor of incarceration. An AI trained on this data may learn to associate certain dialects, vernacular, or cultural references—rather than genuine criminal intent—with "suspicious" activity. This risks automating and amplifying the very biases the justice system is already accused of perpetuating.

"You are essentially building a racial and socioeconomic profiling engine and calling it crime prediction," says Dr. Alisha Williams, a data ethicist at the Center for Democracy & Technology who studies carceral tech. "The model isn't finding 'crime'; it's finding patterns that look like the crime it was shown, which was itself a product of biased policing and sentencing."

Due Process in the Algorithmic Age

The deployment of this AI triggers serious due process concerns. What happens when a conversation is flagged? Does it lead to extended solitary confinement, loss of privileges, or new criminal charges? Inmates have severely limited rights, but they are not entirely extinguished. The opacity of the AI's decision-making—a common issue with complex machine learning models—makes it nearly impossible for an incarcerated person or their counsel to challenge a "risk score" or understand why their communication was flagged. This creates a black-box disciplinary system.

Furthermore, the system chills protected speech. The First Amendment still applies within prisons, albeit in a limited form. Knowing that an opaque AI is analyzing every word for nefarious intent will inevitably deter individuals from speaking freely with family, discussing legal strategies with attorneys, or even using metaphorical or hyperbolic language. It imposes a pervasive, invisible censor.

The Slippery Slope of Function Creep

A critical fear among advocates is function creep. A tool sold for preventing violence and drug trafficking inside prisons could easily be adapted for broader surveillance goals. Could it be used to identify gang affiliation based on speech patterns? Could it flag discussions of political activism or protest organization as "risky"? The infrastructure for total communicative surveillance is now being normalized in an environment with the fewest privacy protections, setting a dangerous precedent.

Securus's move is not happening in a vacuum. It is part of a wider trend of "correctional tech" that includes biometric monitoring, gait analysis, and network analysis of communication graphs. The carceral system is becoming a laboratory for the most intrusive forms of surveillance, which often later migrate to the general public. The technologies tested on incarcerated populations—a captive market with minimal ability to refuse—frequently pave the way for their use on society at large.

What Comes Next: Regulation, Transparency, or Escalation?

The pilot by Securus forces a urgent societal question: How do we govern predictive policing within walls? Currently, there is almost no regulatory framework specifically governing AI in correctional facilities. The development and deployment of these tools is left to vendors and prison administrators, with little independent oversight, public auditing, or accountability.

Moving forward, several steps are critical to prevent harm:

• Auditability Mandates: Any AI used for disciplinary or security purposes must be subject to rigorous, independent third-party audits for bias and accuracy. The algorithms and their key performance metrics should not be protected as corporate trade secrets when they directly impact human liberty.
• Transparency for the Monitored: Incarcerated individuals and their contacts must be clearly notified that AI analysis is being used, and they must have a meaningful process to appeal flags or decisions derived from it.
• Narrow Scope Limitation: The use of such tools must be strictly legally circumscribed to prevent violent or serious criminal acts, not expanded to monitor general behavior or associations.
• Public Oversight: Legislatures and oversight boards must urgently examine this technology, creating guardrails before it becomes an entrenched, unexamined norm.

The Securus AI pilot represents a technological solution to a complex human problem. But in its current form, it risks solving a problem of inefficiency by creating far greater problems of injustice, opacity, and automated discrimination. The real test won't be whether the AI can predict a crime, but whether our society can predict and prevent the grave harms such powerful, unregulated tools will inevitably cause if left unchecked. The surveillance future is being beta-tested in our prisons today, and its lessons will soon be applied to us all.