A patrol officer gets an automatic license plate reader hit at 2 a.m. and makes a stop based on the alert. Two months later, in a suppression hearing, defense counsel asks a narrower question than anyone expected: not whether the plate matched, but how the system generated the alert, what its documented false-positive rate is, and who is able to explain that in court.

That question, not the stop itself, is where the real pressure on AI-assisted policing sits right now.

What’s Already Deployed

Automatic license plate readers are already running in over 18 US states, feeding real-time alerts to patrol units and investigators. Predictive analytics tools that flag higher-probability locations or persons of interest are similarly embedded in day-to-day operations at many agencies. None of this is new information to anyone running an investigations unit. What’s less often examined is what the adoption curve left behind.

Tools got deployed largely on the strength of their detection performance: faster hits, more matches, quicker triage. Governance, the documentation trail that lets an agency explain why a specific alert fired, who reviewed it, and what the model’s error rate actually is, has generally not kept pace with deployment speed. That gap doesn’t show up on day one. It shows up the first time someone with standing to ask, a defense attorney, an oversight board, a journalist, asks for it.

The Deeper Issue: Bias Doesn’t Disappear, It Relocates

The concern about predictive policing reinforcing historical bias is well documented. A NAACP issue brief on the topic makes a specific point worth sitting with: predictive tools trained on historical enforcement data tend to reproduce the patterns embedded in that data, including patterns that reflect where officers were previously deployed rather than where crime actually occurred at equivalent rates.

That’s a different claim than “the algorithm is racist.” It’s a claim about where the human judgment moved. Instead of an officer deciding, in the moment, where to patrol, a model trained on years of prior deployment decisions now makes a version of that same decision at scale, upstream, and with a veneer of statistical neutrality that a single officer’s judgment never carried. The bias didn’t get removed by the software. It got relocated to a point in the pipeline that’s harder to see, audit, or challenge in real time.

This matters for a concrete reason: an agency that adopts predictive tools without auditing the training data is not eliminating the discretion problem policing has struggled with for decades. It’s inheriting that problem in a form that’s more opaque to the officer using the tool and, often, to the agency’s own leadership.

The Harder Question: Is This an AI Problem or an Audit Problem

It’s tempting to frame this as a technology risk that better AI will eventually resolve. That framing undersells what’s actually missing in most agencies: a standing audit function that reviews model outputs against outcomes on a recurring basis, independent of the vendor that built the tool.

Some of what looks like an AI problem is really a governance gap that predates AI entirely, the same gap that made it hard to audit human-driven patrol allocation decisions for bias before predictive tools existed. Better models help. They don’t substitute for an agency deciding, deliberately, who reviews flagged decisions, how often, and what happens when a pattern of disproportionate impact shows up in that review.

Europol’s Innovation Lab is one of the more serious attempts at building that function at scale, treating AI adoption in cross-border investigations as a research and governance question simultaneously, rather than adopting first and building oversight later. The lab’s structure, a standing space for testing tools against both investigative value and rights protections before wide deployment, is closer to what individual agencies will eventually need internally, even at a fraction of that scale.

What Smaller Agencies Are Actually Up Against

Most investigations units don’t have Europol’s resources or a dedicated AI governance team. That’s the honest constraint, and it shapes what’s realistic to recommend. A small unit adopting AI-assisted lead prioritization needs the audit trail as much as, arguably more than, a large one, precisely because it has fewer people available to catch a problem informally before it becomes a legal or reputational one.

What that looks like in practice is less about acquiring a governance department and more about building the habit into the workflow itself: every AI-assisted flag carries a record of why it fired and what data informed it, every case that used an algorithmic assist is tagged as such, and every few months someone actually reviews a sample of those flags against what happened afterward. This is the kind of structure platforms like Hubstream are built to support, keeping the provenance of a lead, not just the lead itself, attached to the case as it moves through review, so the explanation exists before anyone has to reconstruct it under pressure.

None of this requires a large team. It requires deciding, before the first flag fires, that the record will exist.

Practical Questions Worth Asking Now

Before the next suppression hearing or public records request arrives, it’s worth testing a few things directly. If asked in court, could your agency explain how a specific ALPR or predictive flag was generated, and produce documentation of its error rate? Who, specifically, is responsible for reviewing flagged decisions for disproportionate impact, and how often does that review actually happen? Is your team distinguishing between an AI tool that failed and a training-data problem the tool merely inherited? And if an outside reviewer asked for the audit trail behind last quarter’s flagged cases, would one exist?

The Question That Was Always Coming

Adoption of AI-assisted policing tools has outpaced the infrastructure to explain them, and that gap was always going to surface in a courtroom, a hearing, or a headline before it surfaced in an internal review. The agencies in a stronger position aren’t necessarily the ones with the most advanced models. They’re the ones that can already answer the question a skeptical outsider is going to ask: not “does the tool work,” but “can you show your work.”