Flock Tests Police AI That Turns Driving Patterns Into Names and Addresses

01Flock Tests Police AI That Can Identify Drivers by Driving Patterns and Link Them to Names and Addresses

Flock Safety is testing an artificial intelligence tool that lets police start with a place, time period and pattern of driving, then find people whose vehicles match. That reverses the company’s traditional license-plate search: instead of looking up a known vehicle tied to an investigation, officers can ask the system to generate possible targets or witnesses.

Flock’s cameras record vehicle movements in more than 6,000 US communities. The company has long said its technology identifies plates rather than drivers, but code obtained by WIRED shows how the new product, OS Investigate, can connect movement records to names, home addresses and other personal information.

WIRED found the code among more than 450 files publicly loaded by Flock’s website login pages. It describes 69 prewritten prompts that officers can select, edit and submit, along with 45 tools capable of accessing plate scans, camera metadata, arrest and case records, 911 dispatch logs, ballistics results and commercial databases containing information including Social Security numbers, birth dates, phone numbers, email addresses and relatives.

Nineteen prompts search for behavioral patterns rather than a particular record. Fourteen require no plate number, name or physical description: an officer supplies only a location, time window and behavior.

One prompt seeks vehicles seen most frequently in a neighborhood during specified hours over the previous 14 days, presenting them as potential witnesses. The initial result is a list of license plates; other OS Investigate tools can convert those plates into names and home addresses.

Other prompts search for vehicles that visited at least three retail locations in three days, multiple banks in a week or several gas stations between midnight and 5 am. Searches can also identify repeat trips between cities or vehicles passing through several areas in sequence. A filter excludes buses, semitrailers, work vans and trailers from some results.

Another prompt begins with people arrested more than twice within two years for any non-drug offense. It directs the system to map their homes, retrieve calls for service at those addresses and choose three people for a “workup”—Flock’s term for a background check assembling vehicles, prior listings as a suspect, relatives, phone numbers and online accounts.

Flock did not dispute the reported capabilities but said OS Investigate is separate from its license-plate-reader product and helps investigators use information their agencies can already access. The company said only a small group of law-enforcement partners is testing it and that its features may change substantially before broader release. Flock has not identified those agencies, disclosed how many searches have occurred or confirmed which capabilities will reach the final product.

02Google Discover Will Let Users Retune Its Recommendation Feed With Chat Instructions

Google is adding a conversational way to customize Discover, the feed of recommended articles based on a person’s activity across Google Search and the company’s apps. Instead of relying only on those inferred interests, Discover will let users explicitly describe what they want to see.

Google says the feature will roll out to the Google app in the coming days. It will appear in the three-dot menu within the Discover feed and open a chatbot-style interface where users can enter their preferences in natural language.

After receiving an instruction, the chatbot will confirm the user’s choices and list the types of content that Discover plans to prioritize. If its interpretation is incomplete or inaccurate, the user can provide more information or correct it through the same conversation.

The new recommendations will not appear immediately after the preference is entered. Users must select “Refresh your feed” before the changes take effect. Google says the system will then remember those preferences for future visits, giving explicit instructions a continuing role in shaping a feed that has previously depended on activity signals.

The source does not specify where the feature will launch first, which languages it will support, or how long saved preferences will remain in effect. It also does not say whether users will have a central place to review or delete everything the system has remembered.

Google announced two related personalization changes. In the Android version of Google News, users will be able to personalize daily audio briefings. The company is also updating Preferred Sources, a feature that surfaces more content from selected publishers in Search’s “Top stories” section, AI Overviews, and AI Mode.

Publishers will now be able to place an interactive Preferred Sources button on their own websites. Readers can use it to add an outlet to their preferred list without leaving that publisher’s page.

03In Two Six-Day Tests, AI Agents Completed Research Engineering but Produced No Top-Conference-Caliber Results

AI agents completed much of the practical work needed for two artificial-intelligence research projects but failed to produce results worthy of a top machine-learning conference, according to a new study. The finding tests one requirement for “recursive self-improvement,” the prospect of AI systems helping improve themselves with little human supervision.

To separate original research ability from memorization and retrieval, the researchers used a “shadow evaluation”: an agent was asked to solve research questions taken from two high-quality papers that had not yet been published. Because the papers were unpublished, the agent could neither have memorized their answers during training nor find them online. The questions came from papers submitted to NeurIPS 2026.

The team tested Anthropic’s Claude Opus 4.8 through OpenClaw, an open-source agent framework. For each task, the agent received six days, $3,000 in Anthropic API credits, a GPU budget, its own virtual computer and access to the open web. Its assignment was to produce a paper meeting the standards of a top AI conference.

The agents handled the engineering workload. They reviewed existing literature, ran hundreds of experiments and compiled their findings. The original papers’ authors then assessed the AI-generated papers as conference reviewers would—and rejected both.

The failures centered on research judgment rather than execution. The agents sometimes tested hypotheses on unusually small synthetic datasets, struggled to explain their work clearly and made no novel contribution. They explored too few alternatives, committed prematurely to weak approaches and abandoned promising hypotheses using limited evidence.

When an approach failed, the agents could make small adjustments but could not fully retreat and begin again with a fundamentally different direction. Feedback did not solve that problem: after receiving criticism from helper agents and external AI review tools, they generally narrowed their claims and added caveats instead of changing their methods. They also used time, computing resources and tokens ineffectively.

The results constrain claims about near-term automated, open-ended AI research, but only narrowly. The study covered two tasks; the graders knew the papers were AI-generated; and the researchers exercised substantial judgment in designing the evaluation. It also remains unknown whether the failures arose mainly from Claude Opus 4.8, the OpenClaw workflow or the allocated resources. The team is testing another model, but no results have been reported.