32 Data Brokers Told California They Sell to AI (shape family: `number-outcome`, 47 chars)

Introduction

California's data broker registration form has a yes-or-no box on it that reads: shared or sold consumers' data to a developer of a GenAI system or model in the past year. Thirty-two of the state's 603 registered brokers checked yes. One of them, PitchBook, answers the obvious follow-up on its own website: "Which AI tools can I use PitchBook data in? The connectors run today in Claude, ChatGPT, Microsoft 365 Copilot, Perplexity, Hebbia, and more."

Also on that list of 32: Equifax Information Services, Experian Marketing Solutions, TransUnion Risk and Alternative Data Solutions, and Dun and Bradstreet. You have a file with at least one of them. You didn't open it, you can't read it, and the company holding it has told the State of California in writing that it sold or shared consumer data with an AI developer last year.

Nothing here leaked. The disclosure is a government filing the broker submitted itself, the confirmation is a marketing page the broker published to sell more of it, and both are online right now, searchable by company name. The receipt is a form, filed on a deadline, with a fine attached.

One Box on a California Registration Form

The box exists because of Senate Bill 361, the Defending Californians' Data Act, which Sen. Josh Becker wrote and Gov. Gavin Newsom signed on October 8, 2025. Effective January 1, it amended the 2023 Delete Act to add five recipient questions to the annual registration every California data broker files by January 31: did you sell or share consumer data with a foreign actor (China, Russia, North Korea, or Iran), the federal government, other state governments, law enforcement outside a subpoena, or "a developer of a GenAI system or model." That last one is subsection (S) of Civil Code § 1798.99.82(b)(2).

Miss the registration and it's $200 a day, and CalPrivacy has held that the same fine covers brokers that don't disclose their data-sharing practices.

What comes out the other end is a spreadsheet. The agency publishes the whole registry as a downloadable CSV: 603 rows, one per registered broker, GenAI question in column 25. No FOIA request and no login, just a spreadsheet program.

The law behind it had a loud month. On August 11 CalPrivacy, the California Privacy Protection Agency, announced its first enforcement action combining the CCPA and the Delete Act: a $116,490 order against LocateSmarter, an Iowa skip-tracing broker — $30,600 of it for late registration under the Delete Act, $79,890 for a separate CCPA opt-out and data-minimization violation. Two days later it fined Cybba, a Boston ad-tech broker, $52,400 for missing the registration deadline. Neither case touched the AI box.

PitchBook Checked Five of Those Boxes

PitchBook Data Inc. is a Seattle company-intelligence firm owned by Morningstar, and its registry row says yes to all five of SB 361's recipient questions: GenAI developer, foreign actor, federal government, other state governments, law enforcement. The same filing discloses that it collects consumers' gender identity or expression data.

The buyers are named on PitchBook's own property, not inferred from the filing. Past the connectors FAQ, a March 12, 2026 company announcement lists an "expanding network of AI partnerships, which include Anthropic, Farsight AI, Finster, Hebbia, Model ML, OpenAI, Rogo, and Writer."

That pairing proves less than it looks like. PitchBook's connectors are licensed enterprise integrations; the Premium Connector requires a PitchBook license, an enterprise license with the AI partner, and an SSO-enabled connection. That is query-time retrieval, someone at a client firm asking Claude for a company's funding history and Claude fetching the live PitchBook record. It is not proof that PitchBook data trained either model, and the record doesn't support saying so. PitchBook told California it sold or shared consumer data with a GenAI developer, and PitchBook says its data runs inside Claude and ChatGPT today.

PitchBook is also the wrong company to picture if you're worried about your own file. Its product is business and professional profiles, and it answered no on minors' data, government ID numbers, citizenship status, biometrics, and precise geolocation. It logged 949 deletion requests in 2024 and complied with all 949 in whole, none denied. On consumer-rights handling that beats most of the registry. What makes the 32 personal runs through the other names on the list, the ones whose files are assembled out of people who never signed up for anything.

"Fully Licensed for Commercial AI Training"

The sentence is Similarweb's own, from its Data for AI page: "Yes, our data is fully licensed for commercial AI training and deployment. Unlike scraped data, ours comes with clear usage rights." Where PitchBook's connectors are a retrieval integration, Similarweb is selling the training use case itself. The same page sources the data to "millions of opted-in users," describes "all user behavior data aggregated and anonymized," and offers custom Claude assistants.

Similarweb's California filing describes the same company in different terms. It discloses collecting consumers' government-issued identification numbers, citizenship or immigration status, and gender identity. One document was written for AI buyers, the other for a regulator. Similarweb also reported receiving zero consumer rights requests of any kind in 2024.

Who Benefits

Two forces are squeezing the traditional broker business at once. Since August 1, every registered broker has to check California's deletion platform at least once every 45 days and clear the requests it finds within 45 days, and more than 300,000 Californians had enrolled as of June 2. Scraped training data got expensive at the same time: Anthropic's $1.5 billion settlement with authors won final court approval on July 20 in Bartz v. Anthropic, and the New York Times' case against OpenAI and Microsoft is still open.

Selling to AI developers answers both problems with the same inventory. The data itself never changes, which is the part that reframed this for me. What changes is the paperwork wrapped around it, and brokers are now selling that paperwork as hard as the data itself. Similarweb puts it in the pitch: unlike scraped data, ours comes with clear usage rights.

For the buyer, the benefit is cover as much as capability. A lab that licenses from a broker has moved the provenance question one contract away from itself, and if the people inside that dataset never agreed to be sold into an AI product, that's now an argument about the broker's consent chain. OpenAI has signed roughly two dozen content and data deals, per LLMPulse's July 2026 map of the market; the largest, with News Corp, could be worth more than $250 million over five years, the Wall Street Journal reported in May 2024, citing people familiar with the terms. Broker data is the cheap version of the same insurance.

CalPrivacy benefits too. The record only exists because it built one. Registration runs $6,000 a broker, so 603 filings clear $3.6 million a year before late fees, funding the deletion platform and the agency's Data Broker Enforcement Strike Force.

Nobody Verifies the 32

Every field in that registry was typed in by the company it describes, and nobody verifies it before publication. EPIC's Justin Sherman ran into that in March, when he published a list of 33 brokers whose filings disclosed sales to foreign actors and seven of them contacted him afterward to say their own entries were wrong. Two of the seven, Clay Labs and Warmly, are also on the GenAI list, and both have since flipped their foreign-actor answer to no in the live file while leaving the GenAI answer at yes. Sherman's conclusion: "CalPrivacy and other regulators should not treat information self-reported by brokers as ground truth."

The errors can run in either direction, and the Stanford RegLab team behind the August 11 policy brief thinks the bigger gap is on the other side. That brief cites this same registry for the same 32, so it's one dataset read twice, not two sources agreeing. The team reviewed all 522 brokers registered for the 2025 reporting year and found 9% fully complied with the Delete Act's transparency-reporting rules. Forty-five percent filed no consumer rights-request metrics at all. The researchers' own read on the AI disclosures is that they are far from complete.

There's a live disincentive to fix a wrong answer, too. In June, CalPrivacy fined WINR Data, Clarivate, and Decision Resources between $10,800 and $12,200 each, per Privacy Daily, for reporting foreign-actor sales they hadn't made. All three registered on time and caught the mistake themselves. Bloomberg Law quoted Executive Director Tom Kemp in July: "I want to be very clear that data brokers have a legal obligation to provide only true and correct information."

Moody's Corporation is the row that sticks with me. It checks all five recipient boxes, the same as PitchBook, and in the same filing it reports denying 1,516 of 1,533 deletion requests, 98.9%, plus all 87 requests it got to know what it holds. Denials aren't automatically violations; the law lets a broker refuse a request it can't verify, and the form never asks why. The registry made Moody's disclose all five buyers without making it tell any one person what's in their own file.

The Bottom Line

The registry sits at cppa.ca.gov/data_broker_registry, searchable by company name and filterable by who a broker sells to, GenAI developers included. California residents can go further: DROP sends one deletion demand to all 603 registered brokers at once. It won't tell you which of them holds your file, since no such lookup exists, and it doesn't need to. Cory Doctorow called the sign-up an obstacle course in July; Techlore's Henry Fisher answered a month later that two of the fields Doctorow called required are actually optional. Doctorow: "I strongly believe that you should go through the tedious, performatively difficult DROP process anyway." Fisher's closing words are "Sign up."

About 1% of California adults have. No other state asks its data brokers the AI-developer question, and none has a tool like DROP. The only reason anyone knows that Equifax's data unit, Experian's marketing arm, and the Morningstar subsidiary that says its data runs inside Claude and ChatGPT all answered the same question the same way is that California made them answer it. Next January the same box gets filled in again, by the same companies, about themselves. The 32 who said yes have been counted and quoted in a Stanford policy brief, but nobody has checked the 571 who said no.