Clearview AI, the company that made its name in 2020 by scraping more than 3 billion photographs to turn faces into names for police, has built and privately tested an AI tool that picks up where the face match ends. WIRED found the tool, an experimental "analyst assistant" called InquiryIQ, in files that Clearview's login page sends to any visitor's browser before authentication. It is designed to take a detail an investigator surfaced through a Clearview search and automatically widen the search across the open web — opening pages, analysing images, assembling what it finds into a profile that may include suspected employers, aliases, acquaintances and physical characteristics. Clearview says it is a prototype that was never offered or delivered to any customer, and that no release is planned in its present form.
Source: wired.com
According to the code WIRED examined, InquiryIQ can search web pages and images, browse pages, run face recognition against photos it finds, assemble a "candidate graph" of possible identities and connections out of the investigator's starting data, and enrich a profile with suspected addresses, phone numbers, employers, social media accounts, arrest history and aliases. The interface text states the tool should "automatically find and enrich personal data from web sources."
The inputs are as notable as the outputs. InquiryIQ is meant to run after an investigator has already done a face search and flagged details worth pursuing; those details can be added to a profile along with age, gender, race, hair colour and eye colour. The interface says this information helps the AI make "smarter decisions when executing search queries." One of the models Clearview tested for those decisions was SpaceXAI's — Elon Musk's company, maker of Grok, formed when SpaceX and xAI merged in February. Musk has marketed Grok as an alternative to supposedly "woke" AI systems, and the chatbot has repeatedly drawn criticism for racist, extremist and provocative output. How SpaceXAI would have used the demographic fields, and what difference they would have made to results, is unclear. Clearview declined to speculate.
WIRED found the tool using the same technique it had used a month earlier to uncover OS Investigate, an AI search tool under development at Flock Safety, whose interface it reconstructed from publicly served files. The Clearview files contain code and thousands of lines of interface text — instructions, warnings, feature descriptions — showing how the company designed the tool and how it described what the tool could do. They do not show what runs on Clearview's servers, how well any of it works, or who used it.
Clearview's position is that a detailed interface no longer implies an imminent product. Modern AI tools, the company says, make it far faster to build elaborate prototypes, which is why its engineers are encouraged to push such work forward quickly while that remains possible. Chief executive Amos Kyler told WIRED that SpaceXAI and the other models visible in the interface were there to compare outputs from different models during testing, not so that police could pick a model for a given investigation. No law enforcement officer, he says, has used InquiryIQ. The company also rejects the "automated investigator" framing, describing the tool as a limited way to automate web searching that detectives already do by hand.
That framing matters because it is a retreat from the line Clearview held for years. The company's consistent claim was that its responsibility ended at surfacing possible leads through face recognition; in a 2022 post, founder Hoan Ton-That wrote that the investigator had to click through the links and do the follow-up research themselves. InquiryIQ takes over part of that follow-up.
The company's history explains why the distinction was worth defending. Clearview was founded in 2017 and spent its early years largely out of public view; that same year Peter Thiel put $200,000 into it, and it soon began signing up police departments with free trials across the country. A 2020 HuffPost investigation documented Ton-That's ties to the far right, including a 2016 dinner at the Republican convention attended alongside white nationalist Richard Spencer and membership in a private online community that included far-right activists and extremists; he later apologised for his earlier posts. Also in 2020, The New York Times reported that Clearview had collected more than 3 billion images from Facebook, YouTube, Venmo and millions of other sites, and the company became overnight one of the most contested surveillance firms in the United States.
What followed was demands from tech companies to stop scraping their users, then lawsuits, then regulatory investigations. Clearview kept scraping. By its own account the database has grown from more than 3 billion images in 2020 to more than 70 billion today, and the company says over 2,000 law enforcement agencies use its technology. Ton-That stepped down as chief executive in December 2024 and left the board the following spring. Kyler, who became CEO in October last year and joined Clearview as an engineer in 2019, presents the company in a flatter, more technical register: control, auditing, oversight. The mission, he says, has not changed; the current phase is refinement, and making sure the product meets the standard of good faith customers expect.
Andrew Guthrie Ferguson, a George Washington University law professor who studies AI and policing, calls this kind of automated investigation "digital prowling" — a system that builds a picture of a person out of scattered traces that used to sit in separate places and were hard to connect. Woodrow Hartzog, a privacy researcher at Boston University, makes the sharper structural point: the sheer labour of investigation was itself a limit on surveillance. Existing privacy protections were written for a world in which the state had to spend real effort to learn about someone, and some rules were never written at all because watching everyone was not practical. Tools like InquiryIQ remove that constraint. Experts told WIRED the same shift cuts both ways: work that took days or weeks collapses into minutes, which solves crimes faster and also makes it cheap enough to investigate people no one would otherwise have bothered with.
Clearview's answer to all of this is the human in the loop. After a search, InquiryIQ is meant to present the identities, connections and other details it found, and the officer accepts or rejects each one before it enters the profile. The company warns that automatically generated demographic information, social media data and arrest history "may or may not be accurate," and requires the officer to confirm independent verification before accepting a result. Kyler says human review is the foundation of the design: the system finds possible leads, not truth, and it is the analyst's job to separate fact from noise.
Hartzog's response is that the human in the loop offers limited comfort. Over time, investigators come to rely on the automated system, and the person checking the machine becomes a box to tick. United States v. Sant in Minnesota, a case involving undercover Homeland Security Investigations agents and the surveillance of political activists, shows what that looks like on paper. Defence lawyers obtained a Clearview report matching photographs from protests spanning roughly 15 years. Every result was marked "accepted by Guy Gino" — including one Clearview itself had flagged as "less likely," with a footnote noting it could only be exported after user acceptance. The defence says the report contained photographs of an entirely different man, his pregnant wife and his young daughter. There is no indication InquiryIQ was used in that case.
Reliability is a second problem, and in the prototype it was configurable. The interface includes a control for selecting which model does the research, listing xAI and Amazon Bedrock, the platform for accessing other companies' models. In 2025, xAI said an unauthorised change to Grok's system prompt caused the chatbot to inject claims about a "white genocide" in South Africa into unrelated conversations; less than a month later, after another change to its instructions, Grok began posting antisemitic messages and praising Adolf Hitler. SpaceXAI did not respond to a request for comment. Michael Price, litigation director of the Fourth Amendment Center at the National Association of Criminal Defense Lawyers, says a hallucination-prone chatbot would not be trusted as an informant in any other setting — and ties that directly to the evidence police rely on when arguing probable cause in court. Price, who worked on Chatrie v. United States, the major Supreme Court case on geofence warrants, says the reliability question is not specific to Grok: models are trained on internet material that includes incompetent posts, conspiracy theories and the prejudices of the online environment.
The prototype defence is the weakest part of Clearview's account, and not because it is untrue. Thousands of lines of interface copy, per-model selection, accuracy disclaimers and a mandatory verification checkbox are not the artefacts of a weekend experiment; they are the artefacts of a product being designed for a customer who exists. Kyler is almost certainly right that nobody in law enforcement ran it. He is also describing a company that encourages its engineers to push this kind of work forward "while that remains possible" — a phrase that concedes the constraint Clearview is racing is regulatory, not technical.
The more interesting question is what the demographic inputs were for. Every other design choice here is defensible on efficiency grounds: automating clicks an investigator would make anyway is an argument you can have. Feeding race, gender and age into the system as decision aids is a different kind of choice, one that shapes which leads the machine pursues rather than how fast it pursues them. Clearview's response was to decline to speculate about its own interface. That is the one answer a prototype cannot excuse.
Ferguson allows one possible upside. AI-assisted investigation could leave a clearer record of how police arrived at a suspect: today officers routinely run database searches, discard leads and follow hunches without documenting the steps, whereas a system that stores prompts, queries, chosen models and investigative directions would be far easier to audit. His optimism is bounded by 16 years of watching police technology get deployed before the agencies using it decide what the rules are — new technology, same script.
The company's own numbers make the point better than its critics do. Between 2020 and today, through lawsuits, regulatory investigations and demands from every major platform it scraped, Clearview's database grew from 3 billion images to more than 70 billion. Nothing in that period slowed the collection down. There is no reason to expect the next phase — the part where the collection starts reasoning about what it holds — to run into a different kind of obstacle.