What Flock’s AI Search Tool Actually Does
Flock’s platform lives inside a web‑based console that officers open on a laptop or tablet. The interface presents three main search modes:
1. FreeForm (plain‑language) search – an officer types a description such as “person wearing a red hoodie and a tattoo of a dragon on the left forearm.” The text is vectorised and compared against numeric embeddings of every frame captured by the city’s cameras. The system returns a ranked list of video snippets that best match the description. 2. Smart Sort – after the initial list appears, the officer can vote on a few results (thumbs‑up or thumbs‑down). Each vote re‑weights the underlying similarity scores, pushing footage that looks more like the approved clips to the top. This iterative loop lets a single officer refine a massive result set in seconds. 3. Watch‑list / area‑wide search – an officer draws a polygon on a city map, selects a description, and tells the system to run continuously across all cameras inside that zone. The backend streams live frames to the model and flags any matches in real time.
The UI also shows case codes (a short alphanumeric tag the officer must enter) and a confidence score (0‑100) for each result. A tooltip warns, “Results may be inaccurate; verify with other evidence before acting.”
> *Example walkthrough*: An officer in Chicago wants to locate a suspect described only by clothing. They select FreeForm, type “person in a blue denim jacket, black backpack,” set the case code “CJ‑2026‑001,” and hit Search. Within seconds the panel lists ten clips, each with a confidence rating. The officer clicks the first clip, sees the suspect on a street corner, and uses Smart Sort to up‑vote it, instantly reshuffling the list so similar footage rises to the top.
The underlying model is a proprietary vision‑language transformer trained on millions of public‑domain images; the exact architecture is not disclosed by Flock. For a technical deep‑dive see the WIRED reverse‑engineering report Source.
The Promise of Guardrails: Policies vs. Practice
Flock’s public documentation (updated August 2026) lists three core safeguards:
* Data‑retention limits – raw video and search results are automatically deleted after 30 days unless a case code is attached, in which case they are kept for up to 90 days for investigative use. * Mandatory case codes – every search must be tagged with a pre‑approved case identifier. The UI disables the Search button until a code is entered. * Automated audit logs – each query, the officer’s description, the model’s confidence score, and any subsequent Smart Sort votes are logged. An internal dashboard flags searches that exceed a confidence‑threshold of 80 % without a case code, generating an email alert to the department’s compliance officer.
What the code actually does
The JavaScript bundle captured by WIRED shows that the audit‑log call (/api/audit) is optional until the end‑of‑year rollout. Until then, departments can toggle the “enforce‑audit” flag off, allowing searches to proceed without any server‑side record. Moreover, the retention timer is hard‑coded to 90 days for *any* case code, regardless of local policy, and there is no cryptographic proof that deleted footage is unrecoverable.
Real‑world implementation gaps
* Several police departments (e.g., Austin PD) reported that their IT staff could disable the case‑code requirement via a hidden config file, effectively reverting the UI to a “free search” mode. * A Freedom of Information Act request in Illinois revealed that audit logs for 2025 contained null entries for 12 % of all searches, indicating that the logging service was down for weeks. * The promised “automated audits” are limited to flagging high‑confidence searches; they do not block the officer from proceeding, nor do they trigger any external oversight.