Microsoft bans US police from using Azure OpenAI for facial recognition

Equipping police body cameras with cutting-edge AI sounds like a sci-fi thriller, but big tech is already drawing the boundaries in real life.

Microsoft just updated the terms of service for its enterprise Azure OpenAI Service, formally barring U.S. police departments from using the platform's generative AI models for facial recognition. The move tightens restrictions around OpenAI's tech just as law enforcement vendors race to integrate generative models into frontline policing.

The updated policy explicitly prohibits integrations used by or for U.S. police departments for facial recognition, covering both current and future image-analyzing models. Crucially, a separate clause applies to any law enforcement agency globally, explicitly banning real-time facial recognition technology on mobile cameras like bodycams and dashcams in uncontrolled, in-the-wild environments.

Walking back the broad ban

The update caused immediate confusion. Initial wording suggested an outright ban on police departments using Azure OpenAI entirely. Microsoft quickly clarified after publication that the original text contained an error: the prohibition applies strictly to facial recognition, not routine police administrative work.

The timing is hard to ignore. Just a week earlier, Axon, the dominant maker of police tasers and body cameras, launched a tool powered by GPT-4 designed to summarize audio recorded from officers' body cameras. Civil liberties advocates immediately pushed back, warning that model hallucinations and historical racial biases in training data could easily taint official police reports.

While it remains unclear whether Axon relies on Azure OpenAI Service or if Microsoft's move was a direct reaction, the cloud giant clearly felt the need to set explicit guardrails.

Defense doors stay wide open

Microsoft's new restrictions leave plenty of room for lucrative government business. The blanket ban on facial recognition applies only to U.S. police, and the global ban targets mobile cameras in the wild, leaving back-office facial recognition on stationary cameras untouched.

More broadly, neither Microsoft nor OpenAI is walking away from defense contracts. Earlier this year, reports revealed OpenAI began working with the Pentagon on cybersecurity capabilities, relaxing its historic ban on military applications. The Intercept also reported that Microsoft pitched OpenAI's image generator DALL-E to the Department of Defense for military operations.

Back in February, Microsoft added Azure OpenAI Service to Azure Government, its specialized sovereign cloud. Candice Ling, senior vice president of Microsoft Federal, confirmed at the time that the service would be submitted for additional defense authorizations.

Why it matters

We think Microsoft is threading a very specific needle here. By explicitly banning real-time facial recognition on mobile police cameras, the company avoids the worst-case PR scenario: an officer arresting an innocent citizen because an AI model misidentified them on a jittery bodycam feed in poor lighting. Drawing the line at uncontrolled real-time identification is both sensible and defensible.

At the same time, this shows where enterprise AI vendors see the actual regulatory heat. Generative text summaries of police audio are deemed acceptable commercial territory, even with lingering hallucination risks. Surveillance and biometric identification, however, remain toxic for public-facing cloud contracts.

The real issue is enforcement. Microsoft can cut off access to Azure OpenAI, but law enforcement hardware makers do not rely solely on proprietary cloud APIs. As open-source multimodal models become lighter and faster on edge hardware, police tech suppliers could simply bypass cloud gatekeepers altogether.