Leaked files show Flock hires Philippine contractors to tag U.S. camera footage, sparking privacy panic and leaving residents asking exactly…

An accidental data leak revealed that Flock, the company behind automatic license plate readers and AI-powered cameras, relies on contractors in the Philippines to review and label the video and audio used to train its machine-learning systems, material reviewed by 404 Media shows. The exposed files included step-by-step training guides, sample frames from camera feeds, and a dashboard that lays out queues of annotation work. Reviewers were instructed how to tag images and audio that show people and vehicles recorded inside the United States.

The disclosure raises questions about who can access footage collected by Flock’s cameras and where the people doing the reviewing are based. Flock cameras are installed in thousands of U.S. communities and are used by law enforcement on a daily basis for investigations such as carjackings and stolen-vehicle searches. Local police agencies have performed numerous lookups inside the system on behalf of Immigration and Customs Enforcement.

Companies that build AI systems commonly hire overseas annotators because labor costs tend to be lower than domestic alternatives; the practice is widespread across the tech industry. The work tied to Flock differs in one key way: it involves tagging material captured by a surveillance network that continuously monitors the movements of U.S. residents, a factor that increases the privacy sensitivity of the dataset.

Flock’s cameras automatically scan license plates and capture vehicle color, brand, and model for every passing car. Authorized users can search the company’s network to trace where a vehicle has traveled across multiple camera locations. Authorities regularly query that database without obtaining a warrant, a practice cited in litigation filed by the American Civil Liberties Union and the Electronic Frontier Foundation against a city reported to be covered by nearly 500 Flock cameras.

The company applies machine learning to detect license plates, vehicles, and people, including what clothes they are wearing, inside recorded footage. A Flock patent also refers to cameras detecting "race."

Multiple tipsters pointed 404 Media to an exposed online panel that displayed metrics tied to Flock’s annotation effort. The dashboard included counters labeled "annotations completed" and "annotator tasks remaining in queue," with annotations defined as the notes workers add to reviewed footage to train models. Tasks listed on the panel included categorizing vehicle makes, colors, and types, transcribing license plates, and performing "audio tasks." The panel indicated that some workers completed thousands upon thousands of annotations over two-day periods. The material also noted that Flock recently started advertising a feature that will detect "screaming."

The exposed interface included a roster of people assigned to annotation work. Using those names, 404 Media located several profiles that showed annotators living in the Philippines, based on LinkedIn entries and other public pages. The documents suggested many annotators were contracted through Upwork, a gig and freelance marketplace where companies hire designers, writers, developers and what the platform lists as "AI services."

Tipsters directed reporters to several publicly available Flock slide decks that explained in greater detail how workers should classify footage. It remains unclear which camera feeds annotators were accessing, but screenshots embedded in the guides show frames containing vehicles with U.S. license plates from New York, Michigan, Florida, New Jersey and California. Road signs and other contextual cues appear in sample images, and one frame includes an advertisement for a law firm in Atlanta, reinforcing that the material came from inside the country.

One audio-focused slide instructed annotators to "listen to the audio all the way through," then select from a dropdown menu that included labels such as "car wreck," "gunshot," and "reckless driving." Another slide associated tire screeching with someone "doing donuts." Since it can be difficult to distinguish an adult scream from a child scream, the guides told reviewers to use a second dropdown to record their confidence level, with options like "certain" and "uncertain."

A separate set of slides explained how to handle people captured in frames: annotators were directed not to label people seated inside cars but to tag those riding motorcycles or walking on public ways.

After 404 Media contacted Flock about the exposed panel, the online dashboard was taken down. Flock declined to comment on the leaked materials.