How AI Data Services Misrepresent Real People

In September 2026, a federal judge in Chicago allowed author and former FBI operative James Keene to proceed with defamation claims against Google over AI Overviews, the generated summaries that appear above search results. The summaries had stated that Keene was serving a life sentence without parole, two claims that the court found sufficient to proceed past a motion to dismiss, according to the court’s September 14 opinion.

Keene had helped the FBI gather evidence against a serial killer. As Above the Law described the case, the summaries confused aspects of that story with Keene’s own history. Each cited his Wikipedia page, which did not contain the false assertions, according to the court’s opinion.

The case illustrates a failure that can extend beyond one product. Automated systems can take loose associations between a person and a fact, such as a shared name, a nearby topic, or an adjacent company, and publish them as settled statements.

Much of the burden of finding and contesting those statements falls on the people they describe.

Key points

A federal court let James Keene pursue defamation claims over two Google AI Overviews that falsely portrayed him as imprisoned while citing a page that made no such claim.
Assigning one person’s records to another predates generative AI: a 2024 study found that 60% and 50% of participants had at least one false-positive discrepancy in two types of private background checks.
An analysis commissioned by The New York Times found that more than half of correct AI Overview answers were not fully supported by their cited sources; a separate study found that users rarely click those sources.
B2B data vendors deliver professional profiles described as verified directly into AI assistants, where a guess and a confirmed fact can arrive in the same format.
US courts have allowed some AI defamation claims to proceed, a German court issued a provisional injunction, Texas privacy law excludes commercial and employment contexts, and a federal data broker proposal was withdrawn in 2025.
Plausible professional errors may be less likely to be litigated than severe accusations, leaving their prevalence and downstream consequences poorly measured.

An Old Matching Problem at a New Scale

Attaching one person's records to another predates generative AI. A 2024 study in Criminology compared official state criminal records for 101 people with reports from private background check companies. Sixty percent had at least one false-positive discrepancy in checks from a regulated provider, and 50 percent in reports from an unregulated aggregator: charges appeared in commercial reports that were absent from the corresponding official state reports.

The findings included mismatched information and incomplete case histories. According to the University of Maryland, the providers matched people using names and birth dates rather than unique identifiers such as fingerprints. Such matching can conflate records belonging to different people.

The errors ran in both directions, and they appeared in checks from regulated and unregulated providers alike. About nine in ten participants also had at least one false-negative error, meaning a charge on their official record was missing from the commercial report.

Interviews with participants described how faulty reports limited their access to jobs, housing, and education. Lageson told the university that people have little legal recourse when facing these errors.

Commercial data brokers add a second source of error through inference, the practice of predicting traits that a company has not directly observed. The Record reported in 2024 that brokers routinely sell such predictions to advertisers, insurers, and other buyers. The top privacy official at Acxiom, one of the largest brokers, acknowledged to the publication that the inferences the company sells are not always correct.

According to the same reporting, inaccurate predictions can affect matters ranging from a person's web browsing to a major financial transaction. Errors compound when companies combine inferred data to draw conclusions about customers.

These findings show that unreliable matching and inaccurate inference were established problems before AI search existed. The change is in the cost of turning an uncertain match into finished prose.

A generative system can convert a database entry into a fluent biography or answer and deliver it to any reader who searches a name, at relatively low marginal cost per person described.

Citations That Do Not Support the Claim

Generative search adds the appearance of evidence to these statements. In an analysis commissioned by The New York Times, the startup Oumi tested Google AI Overviews against a standard set of factual questions across thousands of searches, according to Oumi’s published findings. With Google's newer model, the summaries reached the correct answer roughly nine times in ten, yet 56 percent of those correct answers were not fully supported by the sources the summaries cited.

The analysis also found that answers from the newer model were more likely to be unsupported by their sources than answers from the older one, even as overall accuracy improved. Google rejected the analysis, with a company spokesperson telling the Times that the study had “serious holes.” The finding measures a gap between assertions and their cited support, though the benchmark did not specifically test professional biographies.

A separate study by researchers at the Pew Research Center, posted to arXiv, tracked the browsing of a panel of US adults. Visits to search pages that showed an AI Overview led to a click on one of the cited sources about 1 percent of the time.

The same panel data, collected in 2025, showed that users who saw an AI Overview were less likely to click any other search result. They were also more likely to end their browsing session on that page, at roughly one in four visits compared with about one in six without a summary.

Taken together, the studies suggest that citations may give an impression of substantiation even when readers seldom inspect them and the cited support is incomplete.

Keene's case follows this pattern, since the linked page did not contain the claim it appeared to support. Canadian musician Ashley MacIsaac alleged in an Ontario lawsuit that an AI Overview falsely identified him as a convicted sex offender and that organizers cancelled a concert after seeing it, according to Search Engine Journal.

His claim argued that Google was liable for the summary's defective design and should have known the feature could return untrue information. MacIsaac discontinued the suit in July 2026 without a ruling on the merits, CTV News reported.

The harm alleged in his case reached him through a third party's decision, made on the basis of a summary he had no part in producing.

Verification as a Product Label

Business-to-business data vendors sell records about professionals to sales, recruiting, and marketing teams, and they increasingly deliver those records directly into AI assistants. ZoomInfo, one of the largest vendors, offers an app inside ChatGPT and describes its professional profiles as continuously verified so that contacts are reachable, according to the company's product page.

The company describes a database of 500 million contacts and 100 million companies, reachable through connectors for several major AI assistants. A profile retrieved this way appears inside the assistant's answer, where the user sees the vendor's data without the vendor's own page around it.

The company's description of its enrichment service shows how such products work. A user can ask an AI assistant to complete a partially filled sales lead, and the service returns details such as an email address, a direct phone number, and a current job title. The product is designed to fill gaps in incomplete business records.

A field can be empty for several reasons. The vendor may lack information, the person may have chosen not to disclose it, or the vendor may hold only a candidate value produced by matching or inference.

Unless a service records why a field was empty and how it was filled, a guess and a confirmed fact arrive in a customer's system in the same format.

The records then move beyond the vendor. A profile delivered into one company's customer database may be copied into other tools and other firms' systems, outside the view of the person it describes. Repetition across those systems can look like confirmation even when every copy traces back to a single entry.

Research on people search websites, a related category of data broker, shows how one source can appear as many. A 2024 study published by the Privacy Enhancing Technologies Symposium found that four corporate groups were behind 14 of the sites examined.

A profile that appears on several sites may therefore reflect one underlying record rather than several independent findings.

Accountability Arrives After the Damage

Courts have begun allowing claims over generated statements to proceed, though the US cases discussed here have not produced final judgments on the merits. In July 2026, the Delaware Superior Court denied Google's motion to dismiss a defamation suit by activist Robby Starbuck over statements produced by Google's AI tools.

The court decided the motion under established defamation law and the plaintiff-friendly standard that applies at the pleading stage, leaving the novel questions about AI for later. The opinion notes that Starbuck alleged formal notices from his lawyers had put Google on further notice of the statements.

In Keene's case, the court held that two of four challenged summaries plausibly supported a claim of defamation per se, a category of statements, such as false accusations of crime, that the law presumes harmful. The court also found that Keene had adequately alleged actual malice, the standard of knowing or reckless falsehood, at the pleading stage, according to the court’s September 14 opinion.

Google had argued that readers would not treat an AI-generated summary as a factual statement. The court disagreed, reasoning that an ordinary user searching Keene's name was “seeking factual information,” as quoted by Above the Law. A business that buys a contact profile from a data vendor is likewise seeking factual information.

In Germany, a Munich regional court issued a preliminary ruling in 2026 holding Google responsible for false statements in AI Overviews about two businesses. The court described the summaries as “independent, new, and substantive statements” and ordered Google to remove them or pay a fine, according to Lawfare. Google said it would appeal.

The German publishers had sent Google a cease-and-desist letter, and its failure to promptly stop the claims proved important to the ruling, Malwarebytes reported.

Notice and the response to complaints have figured prominently in these cases, though the US rulings did not establish final liability. The person described still often faces the first burden of discovering the statement, identifying where it appeared, and reporting it.

Correction Rights That Stop at Professional Records

Correction outside the courts is also difficult. A 2022 study by many of the same researchers, published in the Proceedings on Privacy Enhancing Technologies under a title comparing the process to whack-a-mole, examined people's experiences removing their data from people search websites. The 2024 follow-up found that data access requests were mostly unsuccessful and removal remained opaque, while identifying common ownership across multiple sites.

Many general consumer-privacy correction rights exclude professional records. The Texas Data Privacy and Security Act defines a consumer as a resident acting in an individual or household context and excludes anyone acting in a commercial or employment context.

Most state privacy laws contain the same exclusion, with California the main exception, according to the law firm Fisher Phillips.

A federal proposal would have brought more data-broker activity within existing consumer-reporting obligations. In December 2024, the Consumer Financial Protection Bureau proposed treating certain data brokers as consumer reporting agencies under the Fair Credit Reporting Act, which carries accuracy and dispute requirements.

The bureau withdrew the proposal in May 2025, stating that rulemaking on the subject was not necessary or appropriate at the time, according to the law firm the Federal Register.

Professional records sold between businesses therefore often fall outside general consumer-privacy correction rights, although other laws can apply depending on how records are used. In practice, challenging an error often begins with the subject discovering it.

Errors That Never Reach Court

The cases discussed here involve accusations of murder and sexual offenses, which can warrant costly litigation. A false job title, a wrong employer, or an incorrect middle name may never produce a lawsuit, even if it influences hiring, investment, or business decisions. Plausible errors may also be harder for readers to notice than obviously absurd ones.

Public research has documented these failures in criminal records, consumer profiles, and AI search answers. Published measurements of how often business data vendors assign people to employers they never worked for, and how long a corrected error persists in downstream systems, remain difficult to find.

The background check study offers one method for producing such figures, since it compared commercial reports against official records for the same group of participants.

A record that stated each claim's origin, its supporting evidence, and whether it was observed, inferred, or copied would allow a reader to distinguish a confirmed fact from a guess. Keene's summaries carried a citation, and the cited page did not contain the claim.

Until that distinction is preserved, the burden of establishing the truth will continue to fall on the people being described, rather than the services publishing claims about them.