records of enemy surveillance

cara esten hurtle

artist's statement

When I first heard drones buzzing around the air in Oakland in the late 2010s, I confess I just found them annoying. In the early 2020s, they felt threatening, a sign of someone watching me in ways I couldn't control. But when, during Pride, SFPD used drones to follow and arrest people in my community for having the audacity to exist in public, I wanted to exert control, I wanted to become the one who surveils instead of the subject of surveillance. So, I bought a used DJI drone off of Facebook Marketplace. And my drone's target was the embodiment of commonplace surveillance in America today, Flock cameras. These little devices send data back to police departments, which then use them for everything from arresting people for loitering, to tracking people who have had abortions.

More than anything, what stood out to me about the process of photographing these devices was geography: they exist at the boundaries of rich neighborhoods like Piedmont to track who enters; they exist embedded into poor neighborhoods to track the people who live there. Place exposes their purpose.

Flock devices use computer vision machine learning algorithms to detect a license plate, isolate it from the rest of the frame (via a process called segmentation) and then read the numbers and letters on it. It then ties the license plate to a particular location and time and records it in its system. That system is able to be queried by police departments across the country.

The video on the homepage of this site also uses computer vision, but to provide anonymity, rather than take it away. It uses the same types of detection algorithms that systems like Flock do, in this case You Only Look Once (YOLO v26) fine tuned with drone specific images. In enforcing anonymity, it is imperfect and imprecise. This, then, is the problem with anonymization: to provide total anonymity, a system must be able to block identification for every individual image. To remove that anonymity may only require a single frame.

In Records of Enemy Surveillance, I hope to point a finger at the culture of state surveillance as it exists in America, it's shape, and it's geography. These systems are mostly talked about in the abstract, but they exist, on real street corners, and are brought into being with actual hardware and algorithms that we can examine and dissect.

What are these photos of?

The red squares are Flock cameras. They are networked devices with cellular radios and solar panels which read license plates via machine learning algorithms. Police have access to the full network of cameras, and have used it to stalk women, detain immigrants, and enforce anti-abortion laws.

How were these images captured?

They were captured using two drones: a DJI Mini 2 and a DJI Air 3S. In some images, the photographer is visible; she is also under surveillance. Future models of drones from this manufacturer have been banned in the United States under unsubstantiated accusations of Chinese surveillance.

All images were captured under FAA Part 107 rules for commercial UAS flights.

Generative AI disclosure

No outputs from generative AI were used anywhere in this project. This website is entirely hand-written, in the classic style. The video piece on the homepage makes use of the YOLOv26 classification and detection neural network, fine-tuned on VizDrone. Because I am not a computer vision expert, I loaded up a local LLM (Qwen3.8-27B) on my MacBook Pro to provide feedback on the code I wrote to anonymize the video, without incorporating any LLM-written code into the project. It provided a number different approaches for cleaning up the code, only one or two of which were actually useful and had an impact on the code I wrote. No cloud-based AI was used here.

Thanks to