New Jersey Just Built a Data-Broker License Regime with $1.5M Fees. Your AI Training-Data Privacy Model Needs to Account for It.

On June 30, 2026, New Jersey signed into law the most expensive data-broker registration regime in the United States. If you source, license, or fine-tune on third-party datasets that contain personal information, the privacy calculus of your supply chain just changed. The fee schedule alone tops out at $1.5 million per year. The per-record penalty for unlawfully selling sensitive data is $50,000. And the statute's definition of who qualifies as a covered entity is broad enough to sweep in companies that have never thought of themselves as data brokers. This post walks through the specifics, the ambiguities, and what they mean for engineering and data teams building on top of purchased or scraped corpora.

Key Takeaways

• New Jersey's Assembly Bill 5328 creates a graduated annual licensing fee for data brokers that reaches $1.5 million for entities handling data on more than 4.5 million consumers, dwarfing California's $6,000 fee and Vermont's $100.
• The statute's scope extends beyond traditional data brokers to businesses with direct consumer relationships that sell or license personal information onward, potentially catching AI training-data vendors and "enriched dataset" providers who never registered anywhere.
• Sensitive-data violations carry a $50,000-per-record civil penalty, meaning a single noncompliant dataset transaction can generate liability in the tens of millions of dollars.
• Critical ambiguities (no defined reference period for consumer counts, no annual registration deadline) compound risk for buyers who rely on vendor representations about dataset scope.
• The safest engineering response is to treat dataset provenance as a first-class audit trail and to evaluate privacy-preserving alternatives (synthetic data, differential privacy, on-device fine-tuning) that sidestep the fee-and-penalty calculus entirely.

What Did New Jersey Actually Pass?

Assembly Bill 5328 (P.L.2026, c.25) establishes a registration regime for data brokers and a set of restrictions on the sale and licensing of sensitive personal data. Governor Mikie Sherrill signed it on June 30, 2026. The bill had been introduced on June 28 and approved by the Assembly, Senate, and Governor two days later, passing alongside the annual appropriations act with minimal public input.

The speed matters. Statutes that move this fast tend to carry drafting issues. This one does.

How Do the Fees Work?

The fee schedule is graduated by consumer count. At the bottom: $5,000 annually for entities handling data on 100,000 consumers or fewer. At the top: $1.5 million annually for more than 4.5 million consumers. For context, California's current annual registration fee is $6,000, Vermont's is $100, and Connecticut's is $2,500. New Jersey's regime is not incrementally more expensive. It is categorically different.

The penalty structure amplifies this. A data broker, controller, or data collector that unlawfully sells, offers for sale, or licenses sensitive data faces a civil penalty of $50,000 per record. Run that number against even a modest dataset. Ten thousand records of sensitive data sold without proper registration or consent produces $500 million in potential exposure. The statute means it.

Why Does the Scope Matter for AI Data Pipelines?

The scope matters because the statute reaches further than any comparable law. Unlike similar laws in California, Vermont, and Connecticut, New Jersey's law covers not only traditional data brokers but also many businesses that have a direct relationship with consumers and subsequently transfer personal information to third parties. The statute introduces the concept of a "data collector," a category that captures companies you might not associate with brokerage at all.

Think about what this means for the AI training-data supply chain. A company that collects consumer data through its own product, then sells or licenses a curated, labeled, or enriched version of that data to an AI vendor for model training, may now fall squarely within the statute's registration and fee requirements. They are not a "broker" in the colloquial sense. They have a direct relationship with their users. But the act of selling or licensing that data onward is what triggers coverage.

If you are buying datasets from these vendors, your vendor's compliance posture is now your problem. A vendor that should have registered and paid $1.5 million annually but did not is a vendor whose data may be unlawfully sold. That fact taints your dataset.

What Ambiguities Should Engineering Teams Worry About?

Several, and they are not minor.

First: the law sets no annual registration deadline and does not specify the reference period for counting the consumers whose data a business sells or licenses. That consumer count determines which fee tier applies. If you license a dataset containing records on 4.4 million consumers, you are in one tier. At 4.6 million, you owe $1.5 million per year. But "consumers" here means New Jersey consumers, and the statute does not spell out how to determine residency within a blended dataset. Your vendor's good-faith estimate could be second-guessed by an enforcement action. Penalties for failing to register accrue at $2,500 per day.

Second: the statute took effect immediately upon signing, but actual enforcement has been deferred. The New Jersey Attorney General's Division of Consumer Affairs announced that businesses do not need to comply with the registry or fee requirements until June 2027. The initial registration window will run from April 1 through June 30, 2027, with additional guidance expected before then.

Third: there is a political wrinkle on top of the legal ambiguity. A senior administration official for Governor Sherrill told a New Jersey news outlet that the administration "would not enforce the law until the legislature fixes certain defects that have come to light." So the law is on the books, effective immediately in theory, deferred until mid-2027 in practice, and the governor's own office has flagged defects. This is not stability. This is a statute you cannot safely ignore but also cannot safely rely on in its current form.

How Does This Interact with Other 2026 Data-Transparency Rules?

New Jersey is not operating in isolation. It became the seventh state to enact a data broker law, and the second this year, following Connecticut. The patchwork is growing, and each state's definitions, fee structures, and exemptions differ.

Meanwhile, California's Generative AI Training Data Transparency Act, which entered into force in January 2026, requires providers of generative AI systems to publish high-level summaries of training datasets, including sources and whether datasets include copyrighted or licensed material where known. That disclosure requirement does not impose fees. But when you combine California's transparency mandate with New Jersey's fee-and-penalty regime, the result is a compliance environment where you must both disclose what data you trained on and ensure that every link in the supply chain was lawfully registered and licensed.

If your training data provenance is opaque, you now face risk on two axes simultaneously.

What Is "Provenance Debt" and Why Does It Matter Here?

Provenance debt is the accumulation of untracked or poorly documented data lineage in your training pipeline. It functions like technical debt, except the interest compounds in legal exposure rather than engineering hours.

The core problem is straightforward and common. Pouring a scraped dump and a broker-purchased licensed dataset into the same training bucket, with no labels, is still common practice. Once the two sources are blended, they can no longer be pulled apart. That fact alone can be grounds for failing a 2026 data audit.

Under New Jersey's regime, provenance debt becomes a balance-sheet risk. The fee tiers and the $50,000-per-record sensitive-data penalty hinge on how many New Jersey consumers' records are embedded in a dataset. "We don't know" is no longer a viable answer. If you cannot demonstrate that a dataset does not contain New Jersey consumer data, you cannot demonstrate that it falls outside the statute. Your vendor might be registered. Your vendor might not be. If you never asked, and the data is already blended into your model weights, you have no way to remediate.

Engineering teams need to treat dataset lineage the way finance teams treat audit trails: mandatory, structured, and queryable. Not because it is best practice (it has been best practice for years, and most teams still skip it) but because the cost of not doing it now has a dollar figure attached.

Are There Exemptions That Might Shield AI Research or Training?

There are exemptions, but they are narrow. Exemptions include protected health information under HIPAA, financial institutions covered by federal law, and data collected, processed, or disclosed as part of research meeting certain safeguards under federal law.

The research exemption is the one most likely to be cited by AI companies, and it is the one most likely to disappoint. "Research meeting certain safeguards under federal law" is a reference to frameworks like the Common Rule, which governs federally funded human-subjects research. Commercial AI model training does not typically fall within these frameworks. A company fine-tuning a model on licensed consumer data for a commercial product is not conducting federally regulated research. The exemption is unlikely to apply.

HIPAA-covered data and financial-institution data have their own carve-outs, but these reflect existing federal preemption patterns, not a broad intent to exclude AI use cases. If your training data contains consumer behavioral data, location data, browsing history, purchase records, or demographic information sourced from or about New Jersey residents, you should assume you are within scope until guidance says otherwise.

What Should Data and Engineering Leads Do Right Now?

The enforcement window does not open until mid-2027. The guidance has not been issued. The statute may be amended. None of that is a reason to wait. Here is what is actionable today.

Audit your vendor contracts for registration obligations

Go through every third-party data agreement your team has signed in the past 24 months. Identify which vendors would qualify as data brokers or data collectors under the New Jersey statute. Ask them directly whether they plan to register. If they cannot answer, that is signal. You want contractual representations about compliance with applicable data-broker registration laws, and you want indemnification provisions that are specific to this exposure. Generic "comply with all applicable laws" clauses are not sufficient when the per-record penalty is $50,000.

Implement source-level tagging in your data pipeline

Every dataset that enters your training or fine-tuning pipeline should carry metadata about its source, the vendor or method of acquisition, the date of acquisition, and any geographic or demographic scope information the vendor provided. This is not complicated engineering. It is a few additional fields in your data catalog. The reason most teams skip it is that it feels like overhead when nothing is on fire. The New Jersey statute is the fire.

If you have already blended sources without labels, document what you know and when the blending occurred. You cannot undo it, but you can demonstrate good-faith efforts to establish provenance going forward.

Evaluate privacy-preserving alternatives to personal-data ingestion

The cleanest way to avoid the entire fee-tier and penalty calculus is to not ingest personal data in the first place. Synthetic data generation, differential privacy techniques, and on-device fine-tuning are not hypothetical. They are production-ready for many use cases. They are not free, and they introduce their own tradeoffs in model quality and engineering complexity. But they eliminate an entire category of regulatory risk, which has a quantifiable value now that New Jersey has put a price on it.

We are not claiming these techniques are universally applicable. For some tasks, you need real consumer data, and no synthetic substitute will do. But for the tasks where synthetic or privacy-preserving approaches are viable, the cost-benefit analysis just shifted hard in their favor.

Do not assume the statute will be repealed or gutted

The governor's office has acknowledged defects. The legislature may amend the law. But data-broker registration regimes, once enacted, do not typically disappear. They get refined. The fee schedule might change. The definitions might narrow. But the structural obligation to register, to pay, and to face per-record penalties for noncompliance is here to stay in some form. Build your compliance posture for the floor, not for the best-case scenario.

How Does This Change the Cost of Third-Party Data for Model Training?

It raises it, directly and indirectly. Directly: vendors who must now register and pay six- or seven-figure annual fees will pass those costs through to buyers. If your data vendor's fee exposure jumps from $6,000 (California) to $1.5 million (New Jersey), expect that delta to show up in your next contract renewal.

Indirectly: the compliance overhead of vendor due diligence, provenance tracking, and geographic filtering adds engineering and legal cost to every dataset acquisition. These are not one-time costs. They recur with every new vendor, every new dataset, and every statutory amendment in every state.

The cumulative effect is that the marginal cost of adding another third-party consumer dataset to your training pipeline is rising faster than most data teams have budgeted for. If your 2027 data acquisition budget was built on 2025 assumptions about licensing costs, revisit it.

What Does This Mean for the Broader AI Data Supply Chain?

The New Jersey statute is a pricing signal. It tells you what one state legislature thinks the externalities of unregulated consumer data trading are worth: somewhere between $5,000 and $1.5 million per year per entity, with $50,000 per record as the penalty for getting it wrong on sensitive data.

Other states will calibrate differently. Some will be lower. Some will follow New Jersey's lead. The direction is clear: the era of low-cost, low-friction personal data acquisition for AI training is contracting. Not ending, but contracting. Every new state registration regime, every new transparency requirement, every new per-record penalty adds friction to the supply chain.

If you are building a company that depends on access to large volumes of third-party consumer data, this is a structural cost trend you need to model. It is not a one-time compliance event. It is a recurring expense line that will grow as the patchwork expands.

For teams that have already invested in privacy-preserving data practices, this is validation, not disruption. For teams that have not, the window to get ahead of it is the next twelve months, before the New Jersey registration window opens and before the next three or four states pass their own versions.

We build Selina around a simple premise: your data should stay yours. Memory is encrypted at rest (it is NOT end-to-end encrypted, because a slice of each request reaches a frontier provider at inference). Files and transfers via SelinaSEND are end-to-end encrypted. We keep non-content operational metadata for a short retention window. These are design choices rooted in the same principle this statute is trying to enforce from the regulatory side: personal data has a cost, and the default should be to minimize how much of it you hold and move. If you want to see what that looks like in practice, start a free 7-day trial, no card required.

Michael C.

Founder & Principal Engineer, Selina Labs

Michael builds Selina, a privacy-first AI that remembers you across conversations. He ships security-sensitive AI in production — real attacks, real fixes, measured in minutes and dollars — and writes about privacy, security, and LLMs from that seat. Top Rated Plus and expert-verified on Upwork.

• www.upwork.com
• x.com
• medium.com
• selinalabs.io