The AI Cop: Inside the Texas Police Department Using Bilingual AI to De-Escalate Spanish-Speaking 911 Calls

Recently, Monica Sanchez, a 911 dispatcher in Marfa, Texas, a small desert town where the closest trauma center is more than an hour away and the air smells like creosote, received a call from a frantic Spanish-speaking resident who described a stomach ailment. The caller used a regional term that Sanchez had never heard before; it was derived from a dialect that was older and thicker than the border Spanish she had grown up speaking in New Mexico. That would have required calling a third-party translation line, entering a passcode, waiting on hold, and hoping the caller didn’t hang up a year ago. This time, in a matter of seconds, an AI-powered system recognized the word, decoded it, and scrolled the translation across her screen.

It’s a brief moment that can be easily missed. However, it is at the heart of something bigger and messier than most people realize: the collision between machine learning, language diversity, and emergency response that is currently taking place in Texas patrol cars and dispatch centers. Carbyne’s live audio translation system, which instantly translates Spanish-language 911 audio into English text and provides dispatchers with a running transcript of both sides of the conversation, was first implemented by the Presidio County sheriff’s office. Similar tools have since been introduced by North Texas 911 centers, which boast language identification in less than two seconds. AI-equipped body cameras that can instantly translate more than sixty languages are now worn by Bexar County deputies. The debate over whether the technology truly performs as promised by its vendors is not keeping up with the rapid spread of the technology.

In these departments, there’s a sense of relief mixed with cautious optimism that something truly helpful has finally arrived. Sanchez uses straightforward, pragmatic language to explain the distinction in situations where seconds count and misunderstandings can be harmful. In Spanish, she never learned numbers. Addresses used to be difficult. She continues to converse with the caller while reading them off the screen. She told GovTech, “It makes things a lot quicker than going to the language line.” Anyone who has listened to a 911 recording can attest to how quickly those calls can spiral out of control; eliminating even thirty seconds of confusion alters the calculation of who receives assistance and how quickly.

However, it’s important to consider what these systems fail to address. A few years ago, Austin posted a Reddit thread that tells a tale that would not have been resolved by real-time transcription. When a family of mostly women and young children had to deal with a broken security alarm, they encountered an officer who became more hostile when they were unable to provide identification or respond to inquiries in English. There was no call for a translator. There was no use of a phone-based service. Only when English-speaking neighbors came outside and confirmed the family’s identity did the situation defuse. There was technology. The policeman did not grab for it. There is still a stubbornly large gap between what is available and what is actually used in the field.

Bilingual pay incentives are offered by police departments all over Texas; in Austin, they are typically small, at about $175 per month. It has been challenging to find Spanish-speaking police officers, especially in larger cities where the work entails greater risk and less community support. One jail employee in East Texas talked about translating for officers from various agencies and occasionally being called in off-duty because no one else could. The city made financial savings. The fundamental shortage did not change. Although AI translation doesn’t address the recruitment issue, it does mask its most obvious symptoms, which may be why administrators find it so appealing.

Whether these tools alter interactions’ emotional texture rather than just their informational content is what’s truly intriguing—and still unknown. Jennifer Eberhardt’s research at Stanford has looked at how police language during stops affects results, which raises concerns about whether AI could eventually train officers to use more composed language. AI is already being used by some departments to look for de-escalation patterns in body camera footage. Perhaps the true goal is something more akin to behavioral nudging, conveyed via an earpiece or a scrolling transcript, and translation is merely the first step.

The technology is currently doing something more straightforward and tangible. Callers in Presidio County are continuing to stay on the line after previously hanging up in frustration. Transcripts are being read smoothly by dispatchers who previously struggled with dialect differences. It’s not a revolution. It’s a patch for a system that most likely required reorganization decades ago. However, patches that are applied correctly can occasionally last longer than anyone anticipates.