FOLLOW US

WASHINGTON — Clearview AI is testing an assistant designed to expand facial-recognition leads into personal profiles, gathering possible addresses, online accounts and associates. The prototype, InquiryIQ, surfaced in publicly accessible website code, WIRED reported Thursday.

Testing included a model from xAI, the maker of Grok. The interface also accepted age, gender and race, although their effect on results remains unclear. Clearview AI told WIRED that police have never used InquiryIQ and that it has no plans to release the current version. WIRED’s investigation disclosed the project Sept. 10.

For someone wrongly identified, an expanded dossier could make the error harder to untangle. Addresses, photographs and social connections might appear to support one another even when investigators started with the wrong person. Safeguards would have to follow that information from the first search through any resulting prosecution.

Clearview AI’s existing policies require human checks, and state disclosure laws can give defendants access to how police found them. Federal oversight records show that agencies have addressed some weaknesses while leaving other privacy work unfinished.

Clearview AI’s safeguards depend on police

Clearview AI says its existing database contains more than 70 billion publicly available images. The company markets facial recognition as a way to identify suspects, find victims and accelerate investigations. Its public website describes a service already operating at an enormous scale.

Under the company’s published standards, investigators must identify a lawful reason for a search. Trained personnel must examine possible matches, and another reviewer must check a proposed identification. Clearview AI also says it records the submitted image, search purpose and searching user’s identity. Its principles describe those requirements as protections against mistakes and abuse.

Those commitments provide specific standards against which customers’ conduct can be measured. They also leave practical questions for any broader research system: who verifies each new claim, what evidence supports that verification and who reviews a disputed finding?

Checking a face and checking a biography involve different judgments. An officer might reasonably conclude that two photographs show the same person while still attaching someone else’s address or criminal record. Each additional connection needs its own support before investigators treat it as established information.

A convincing profile can still be wrong

Generative AI introduces another source of error. The National Institute of Standards and Technology warns that these systems can confidently produce false information, a problem often called hallucination. Its 2024 generative AI risk guidance also identifies the danger of people placing excessive trust in automated outputs.

NIST recommends checking sources and citations in generated outputs during testing and ongoing monitoring. It also calls for recording errors and near misses, giving organizations a way to see whether corrective measures work.

Consider a hypothetical search involving two people with similar names. A system could connect one person’s photograph to the other person’s employment history. If several websites repeat the same mistaken information, a reviewer could encounter multiple apparent confirmations that all trace back to one error.

The risk also extends to accurate information stripped of context. An old address may be genuine but no longer current. A photograph may show two people at the same event without establishing a personal relationship. Those distinctions can affect whom investigators contact and what conclusions they draw.

Clearview AI advertises facial-recognition accuracy above 99% on specified tests. That claim concerns image matching; it cannot establish the reliability of an AI-generated biography. A system needs separate evaluation for the different claims it produces.

NIST has explained that its facial-recognition evaluations test algorithms, rather than complete commercial products. Its research on demographic differences also emphasizes that performance varies with the algorithm, task and data. Those findings do not provide an error rate for InquiryIQ.

Evaluating a police research assistant would therefore require testing the entire process. Relevant measures would include mistaken connections, unsupported statements and whether reviewers catch errors before accepting them. A strong result on one component cannot answer all three questions. Reviewers would also need to distinguish information found on a webpage from conclusions the system inferred about a person.

Wrongful arrests expose failures in verification

Existing investigations demonstrate the consequences of accepting a computer-generated lead too readily. A Washington Post investigation described Christopher Gatlin’s wrongful identification and 16-month detention before prosecutors dropped his charges. The newspaper also documented departments that proceeded without independent corroboration after facial-recognition searches. Its investigation of police practices examined the gap between written safeguards and officers’ decisions.

The report also described Detroit’s $300,000 settlement with Robert Williams after a wrongful arrest. The city agreed to require independent evidence about AI-identified suspects before officers sought an arrest warrant.

A separate Post report published in April described an Oklahoma woman, Kimberlee Williams, jailed for six months before prosecutors dropped bank-fraud charges. A bank investigator had identified her through facial recognition. Montgomery County police did not disclose that origin when seeking charges, according to records reviewed by the newspaper. The reporting on Williams’ arrest illustrates how a flawed identification can move between investigators without its technological origin remaining visible.

Neither report establishes any involvement by InquiryIQ. Both show how an initial match can influence later decisions and why defendants need access to the investigation’s starting point.

More material in a file would not necessarily solve that problem. If investigators attach additional information to the wrong person, the file could become longer while the central identification remains unsupported.

Defendants need the investigation’s starting point

The government already has duties to disclose certain information favorable to an accused person. Under the Supreme Court’s Brady doctrine, those duties cover material evidence concerning guilt or punishment. They can also include information that undermines prosecution evidence or witnesses.

The Justice Department’s discovery policy directs federal prosecutors to seek relevant favorable information from the prosecution team, including participating police officers. The policy calls for reviewing investigative files and substantive communications. It also says the department’s broader disclosure standards do not create additional rights or remedies for defendants.

For an investigation involving AI, potentially significant material could include a competing identification, conflicting dates or evidence that an apparent associate was someone else. Whether particular records must be disclosed depends on the applicable rules and circumstances. There is no automatic promise that a defendant will receive every search or a vendor’s source code.

Preservation comes before that legal dispute. To challenge a claim effectively, the defense may need the original image, the source behind a reported connection and the officer’s record of verification. If police retain only a final narrative, reconstructing how they reached it becomes harder.

That creates a practical distinction between allowing an officer to reject an AI suggestion and maintaining a record someone else can examine. An accountability system must make the relevant decisions reviewable after the investigation has moved on.

Maryland puts disclosure and corroboration into law

Maryland provides a concrete example of legal requirements directed at those problems. Its facial-recognition evidence statute prohibits using a result as the sole basis for probable cause or positive identification. The result must have support from additional, independently obtained evidence. The law also restricts when facial-recognition results can be introduced in criminal proceedings.

A separate disclosure provision requires the state to reveal relevant facial-recognition use under Maryland’s discovery rules. That includes naming the systems and databases searched and disclosing results that led to further investigative action.

Maryland also restricts investigative uses, requires trained independent verification and bars identifying someone solely because of political beliefs or lawful activities. The same provision limits covered criminal investigations to listed offenses and bars live identification. It also generally restricts analyzing people engaged in constitutionally protected activity unless police have reasonable suspicion of criminal conduct. The statute provides for excluding certain improperly obtained results and evidence derived from them, subject to its exceptions.

These requirements give defense lawyers provisions they can cite in court. They also illustrate why describing a product as an investigative aid cannot answer every legal question about its use.

The Maryland provisions concern facial recognition. Applying them to a system that combines image searches, web research and generated conclusions would require examining its actual operation. A product’s name or marketing description would not resolve which parts of an investigation fall within the statute.

Federal privacy work remains unfinished

Federal oversight records show both progress and unresolved problems. In 2023, the Government Accountability Office found that seven federal law-enforcement agencies had initially used outside facial-recognition services without requiring training. Six agencies with available data had conducted roughly 60,000 searches while lacking those requirements.

Agencies subsequently made changes. GAO records show the FBI adopted a training requirement in December 2023. Homeland Security Investigations established a process in June 2024 to verify training before granting Clearview AI access. The Justice Department also issued an interim facial-recognition policy in December 2023.

However, GAO’s recommendation tracker still listed two Justice Department privacy recommendations as open as of January 2026. A Homeland Security recommendation addressing outstanding privacy requirements remained open following a July 2026 update. These findings concern existing services, not an assessment of InquiryIQ.

The record shows why training, privacy compliance and oversight require separate scrutiny. An agency can improve access controls while still having outstanding work on how it documents and protects personal information. Adding a new research capability would raise further questions about what the existing reviews actually cover.

Public information raises a separate privacy question

Accuracy alone cannot settle how much information government should collect about someone. Even a correct identification can lead to scrutiny of relatives, professional contacts or people who merely appear in the same photograph. The justification for investigating the original subject does not explain every possible inquiry into other people.

The Supreme Court has recognized constitutional limits on some forms of digital collection. In Carpenter v. United States, decided in 2018, it held that obtaining historical cell-site location records constituted a Fourth Amendment search. The court emphasized how such records could expose an extensive account of a person’s movements.

On June 29, the court addressed Google location data in Chatrie v. United States. It held that police conducted a Fourth Amendment search when they obtained the defendant’s location information through a geofence warrant. The justices returned the case for further consideration of the warrant’s compliance with constitutional requirements.

Both decisions concerned location records held by companies. Neither established a blanket warrant requirement for reading public websites or decided the legality of InquiryIQ. Their application to other technologies requires attention to the information collected and the method of obtaining it.

For a future police profiling system, that means separate questions about access, scope and use. Whether a piece of information is publicly visible would be one part of the analysis. The legal basis for additional investigative steps would still need examination.

Correcting a database does not automatically correct a case

Clearview AI offers another route for some people concerned about their information. Its privacy request page lists access, correction, deletion and other options for residents of specified states. It also provides an Illinois opt-out process for appearing in search results. Available options differ by jurisdiction.

The page’s Illinois policy distinguishes blocking future searches from deleting retained data. Clearview AI says it retains certain Illinois information to comply with legal preservation obligations, while blocking it from future results.

Those requests address information held by the company. They do not establish that an agency has corrected an investigative report or that a prosecutor has disclosed a disputed identification. Someone affected by an error may have separate questions about the commercial database and the government’s case file.

For any future deployment, useful safeguards would include a way to record corrections, identify affected reports and alert investigators who relied on the original information. A correction that never reaches the decision-maker could leave the practical consequences of an error in place.

The same principle applies before an arrest. Agencies evaluating an AI research assistant would need to establish who can authorize a search, how claims are verified and what records remain available for review. Those decisions would shape whether a person could effectively contest the system’s conclusions.

A defendant should be able to examine how their name entered an investigation and what evidence kept it there. As police technology assembles more information around an identification, access to that history becomes essential to finding where an investigation went wrong.