In August 2026, the Delete Request and Opt-out Platform (DROP), run by the California Privacy Protection Agency (CPPA), entered its mandatory phase. A California resident files one request on one website, and every registered data broker in the state must delete that person’s data within 90 days, statutory exemptions aside. As of August 1, brokers must log into the platform and process new requests at least once every 45 days (CPPA’s DROP guidance for data brokers). On paper, nothing else in the United States goes this far: state-by-state comparisons rate California’s privacy framework as the most far-reaching in the country (Security.org’s comparison).

At exactly this moment, researchers from Stanford RegLab, Stanford HAI, and the University of Washington published a paper whose title does not mince words: “Privacy Without Remedy” (arXiv:2605.21376, accepted at FAccT 2026). They examined every one of the 522 data brokers then registered in California. The share that fully met the statutory transparency obligations: 9%.

Terms first. A data broker (defined in SB 362) is a company you have never dealt with directly that collects and resells your personal data anyway: purchase histories, location traces, property records, health signals (the CPPA’s own list includes search history and Social Security numbers), packaged for advertisers, insurers, and debt collectors. The CCPA, California’s 2018 consumer privacy law, gave residents the rights to delete, to know, and to opt out of sale. The 2023 Delete Act went further: it moved broker registration under the CPPA and requires brokers to register annually and publicly report five metrics on consumer rights requests (how many were received, complied with, and denied, plus mean and median processing time), posted in their privacy policy by July 1 each year. The first deadline was July 1, 2025 (SB 362 text).

Where the 9% comes from

The team did two things. First, they manually checked the privacy policies of all 522 registered brokers against the five required metrics. 9.2% reported all of them. The breakdown is no better: 53% reported deletion-request metrics, 53% reported opt-out-of-sale metrics, and only 36% reported metrics for requests to know what data is sold or shared. About 45% reported no rights-request numbers at all, and the paper makes a point of noting that nobody can tell whether that means no consumer ever asked or nobody bothered to count. Those are two very different signals for a regulator, and right now they are indistinguishable.

The second exercise says more. The team sampled 250 brokers and walked through the request process as a consumer would. At 43% of brokers, it was impossible to exercise all the rights the law grants. 64% of the flows contained at least one design feature that adds friction, what the literature calls dark patterns: interface choices that quietly keep you from finishing what you came to do. The paper lists the tricks. 43% make you resubmit the same information across multiple forms; 37% impose identity verification on opt-out-of-sale requests, which the CCPA explicitly forbids (Hunton’s analysis of the Ford case); 21% insert CAPTCHAs, in the worst case eight in a row for a single opt-out; 10% have submission links or contact addresses that simply don’t work. Some consumers who did get a request through received the confirmation email from a company they had never heard of, with no way to know whose system the request actually landed in.

The fines are aimed at the wrong thing

Why is compliance still in single digits two years in? The paper’s diagnosis lands on the penalty structure. The Delete Act’s fines attach to failing to register ($200 per day) and, under DROP, to failing to process deletion requests. There is no fine for registering and then skipping the transparency report. Registering costs money; blowing off the reporting obligation costs nothing. Consumers can’t sue over it either: there is no private right of action here (the legal term for an individual’s standing to take a violator to court directly; the CCPA grants one only for certain data breaches, and transparency violations don’t qualify). The only recourse is reporting the violation to the CPPA and waiting in line behind its limited enforcement resources.

Jennifer King, who co-led the study, put it plainly: “Without clear reporting requirements and consistent financial consequences, businesses simply won’t do it.” (Stanford HAI)

Enforcement isn’t absent. In March 2026 the CPPA fined Ford $375,703 for exactly this kind of friction: Ford required consumers to click an email confirmation link before it would process opt-out requests, and treated unclicked requests as never made (CPPA announcement). That action ran under the CCPA rather than the broker-specific Delete Act, but it shows the regulator now treats procedural friction itself as a violation. The problem is coverage. The CPPA fined two unregistered brokers in January 2026 and says more than ten similar actions are underway (CPPA enforcement announcement), but enforcement comes one case at a time, against more than 600 registered brokers plus an unknown number of unregistered ones. The paper stops at noting the agency’s limited resources; the arithmetic that follows is mine: spread across that population, the odds of any single broker being audited look too low to deter much.

The pipe leaks upstream; AI collects downstream

The detail in this study most relevant to AI practitioners sits in the registry itself. SB 361, signed in October 2025, requires brokers to disclose at registration, starting in 2026, whether they sold or shared personal data with developers of generative AI systems in the past year. More than 30 have already said yes (Stanford HAI).

Put the two findings side by side: the deletion pipeline fails at scale, and the same companies supply AI training pipelines. The paper stops at the disclosure; what follows is my own extrapolation. A deletion right runs against a broker’s database: you ask, and it deletes the records it holds. The Delete Act extends that duty to the broker’s service providers and contractors, but says nothing about third parties who already bought the data (SB 362, §1798.99.86). So if your data was sold to an AI company before you filed the request, the request doesn’t follow it there. Once training finishes, the data’s influence is baked into the weights. Removing it afterwards is not impossible: retraining the model from scratch without the data works, and machine unlearning (removing a specific data point’s influence from an already-trained model) is an active research field. But retraining at that scale is rarely practical, and unlearning is still an open research problem: surveys in 2026 are still debating how to verify that anything was actually forgotten and whether forgotten content gets relearned (Unlearning in LLMs). As far as I could find, no major model provider ships it as a product feature. So in practice, a deletion right works best when exercised before the data enters a training set, and what this study shows is that in precisely that window, the right mostly cannot be exercised.

For people building models or applications, this lands as a supply-chain due-diligence problem. When you buy training data from a broker or a data vendor, one question is worth adding: does this dataset contain records that should have been deleted and survived only because the vendor ignores requests? The California registry and this paper give you a checkable starting point. Whether the vendor is on the registry, whether it reported its transparency metrics, and what its received-versus-complied numbers look like are now all public information. Does an upstream compliance defect translate into legal liability at the model layer? I haven’t found a direct precedent. But the regulator singling out “sold to generative AI developers” as a disclosure item tells you which direction it is looking.

For California residents, my read: DROP is worth using. One submission covers more than 600 registered brokers (official DROP page), far more practical than filing forms broker by broker. Keep expectations flat, though. If only 9% of brokers clear the transparency obligation, the easiest one to check, actual deletion performance can only be harder to trust. And for anyone tracking the industry, the paper’s method is reusable: don’t ask companies whether they comply, count the numbers they post and walk through their flows. California passed what state-by-state comparisons rate as the most far-reaching privacy law in the country (Security.org), then measured in the field that a strict law is not strict enforcement. What’s worth watching next is not another bill. It’s whether the CPPA starts writing Ford-sized fines to data brokers.

References