U.S. cops are letting GPT-4 write police reports

Police body cameras capture everything from heated arguments to police dogs barking. Turning hours of that raw audio into written incident reports usually eats up half a cop's shift.

Now, some US departments are outsourcing that paperwork to OpenAI. In Oklahoma City, officers are feeding bodycam audio into an AI tool called Draft One, which spits out completed incident reports in just eight seconds.

Dialing back ChatGPT's imagination

Draft One was launched earlier this year by Axon, the dominant maker of police body cameras and Tasers. Under the hood, it runs on OpenAI's GPT-4.

To counter ChatGPT's infamous habit of making things up, Axon senior product manager Noah Spitzer-Williams told the Associated Press that the company dialed down the model's creativity settings. The goal is to keep the output strictly factual without embellishments.

In Oklahoma City, prosecutors advised the department to limit Draft One to minor incidents that don't involve felonies, violent crimes, or arrests.

Other jurisdictions are jumping in with both feet. Police departments in Fort Collins, Colorado, and Lafayette, Indiana, are already using Draft One across all kinds of cases, with one police chief telling the AP that officers love it.

Axon claims it tested 382 sample reports across consistency, completeness, and word choice severity, finding no statistically significant racial bias compared to raw transcripts.

The rubber-stamp problem

Critics aren't sold. Legal scholar Andrew Ferguson warned that the sheer speed and ease of automated writing could make officers far less careful when reviewing paperwork.

There is also the issue of systemic bias. Linguistic research shows frontier models like GPT-4 often carry covert prejudices against dialects like African American English, potentially misinterpreting casual slang as hostile or suspicious.

Axon insists that every AI-generated report must be reviewed and signed by a human officer, keeping legal accountability with the department.

Notably, Axon experimented with having AI analyze the video footage directly, but CEO Rick Smith pump the brakes on computer vision, acknowledging that race and identity issues in policing make video analysis too sensitive for now.

Why it matters

Draft One highlights the classic AI efficiency trap. Freeing up officers from desk duty sounds great on paper, but criminal reports form the bedrock of the entire justice system.

We think relying on human officers as the sole safeguard is a flawed design. When a tired cop is handed an eight-second summary that reads like professional legal prose, the temptation to rubber-stamp it is immense. Subtle AI hallucinations or biased tone shifts will inevitably slip through.

It also sets up a dangerous game of accountability hot-potato. If an innocent person is wrongfully charged due to a warped summary, Axon will point to the officer's signature, while the department blames the algorithm.

Body cameras were rolled out to keep police honest through objective footage. Turning that evidence over to an LLM risks laundering algorithmic bias into sworn testimony before anyone has agreed on ground rules.