Subscribe to the free future newsletter
free future home
Police technology company Axon has received a lot of attention for its “Draft One” product, which uses body camera recordings to create a first draft of a police report for officers after an incident. However, the company has another AI product called Policy Chat that is less discussed.
The product uses a combination of large-scale language models (LLM) and an AI technology called search augmented generation (RAG) to provide frontline personnel with answers about the department’s official procedures and policies. Examples of questions that Axon may answer include: “What are the steps to respond to a domestic violence call?” “What is the department’s off-duty employment policy?” In theory, a department could have a phone bank of attorneys that employees could call with questions. The product can best be thought of as attempting to provide a scaled-up, automated form of that service.
In February I attended a workshop on AI and law enforcement. At that time, the commander present expressed a desire to use the LLM to search and analyze his department’s seemingly vast policy library. The problem was that her department’s policies weren’t publicly available, so she couldn’t upload to Anthropic, OpenAI, Google, etc. Perhaps Axon has heard similar laments and is trying to respond to their demands.
One question is why aren’t all police departments’ policies available to the public? But apart from that, this product also raises other questions.
Corporate access to police materials
If we accept for the sake of argument that there are some internal documents that are legitimately withheld from the public, one question would be whether Axon has access to internal police documents that are not accessible in other parts of the world. In theory, organizations can keep documents private and perform RAG processing purely locally. The demand for such processing is high and it is becoming increasingly feasible. However, for Axon, using Policy Chat requires uploading departmental documents to the company. Axon’s access to nonpublic documents raises a number of questions, including whether a company can use that access to gain an advantage over competitors, vendors, employees, union efforts, and more. Why should Axon be different from Anthropic, Open AI, or Google? Are there provisions in the contract between the department and Axon to address these issues? (Such questions, of course, pertain to many police AI products; more on this in a future post.)
Impact on officer training
I also wonder how this product will affect executives. One former police command staff member I spoke to expressed two concerns. First, “police officers aren’t going to do the training and they’re going to go, ‘Just ask the AI in the field.'” And two, they’re going to be distracted by the AI when they’re in the field. “When we’re trying to talk to a victim, we don’t want them to be half-listening to what the policy is,” he said, “and the last thing we want an officer who is potentially making a hot call to worry about is what the policy is.” Essentially, he said, this is a tool that was never intended to be used, and if it was, “it would be a huge failure in the training department.”
Indeed, police policy manuals can be large, ranging from a few hundred pages to over 2,000 pages in the case of the New York City Police Department. But a former police officer said of his department, “If you boil down to operational matters, we have about 35 policies for patrol. We have about 20 investigations.” Police officers shouldn’t need AI to understand every detail of these policies.
unavoidable error
Another obvious question is, if products of this kind are used, how police departments and vendors like Axon intend to address the inevitable illusions, errors, and biases that all of today’s language models create. Who is responsible if an officer violates a department policy (perhaps an important one) due to an AI error? Axon boasts that its product “provides instant answers and decisions you can confidently make,” but it also warns officers to “always check official policy sources.” Axon is speaking with both mouths here. We like to brag about how much time AI saves us, but we also shift all responsibility for shortcuts onto the shoulders of individual executives. It seems like a trap. (More details in a follow-up post.)
Ultimately, I don’t think this product will raise the same level of concern as an AI-generated police report, but it’s worth noting as another example of the premature introduction of AI into the criminal justice system. This usage requires close supervision. Like many companies selling to law enforcement and other markets, Axon is eager to embrace and market the AI shortcuts the technology can offer. But there are times when you shouldn’t take shortcuts.
