Skip to content

Rights and contracts

California AB 2013 explained for companies that supply training data

By SourceX Editorial · Reviewed by Noah Loul ·

Short answer

California AB 2013 requires developers of generative AI systems offered to the public in California to post a high-level summary of the datasets used to train them, not the data itself. Suppliers have no direct posting duty, but that summary can describe a supplier's records, rights and contents, so the license should settle how they are described.

Key takeaways

  • AB 2013 puts the posting duty on developers of generative AI systems offered to Californians, not on companies that license them records.
  • The summary covers topics such as dataset sources or owners, whether data was licensed, and whether it includes personal information or IP.
  • The law reaches systems released before it took effect, and fine-tuning or retraining can count as a substantial modification.
  • AB 2013 does not require developers to publish datasets, prices or contract terms.
  • Map each disclosure topic to an agreed sentence in the license before signing, not after a summary is posted.

What is California AB 2013?#

California AB 2013, often called the Generative AI Training Data Transparency Act, is a state law that requires developers of generative AI systems or services made available to Californians to post documentation about their training data on their websites. It is a transparency law: it does not ban any type of training data, and it does not require anyone to hand datasets over.

The law matters to companies that license records because a developer's public documentation is now a place where licensed datasets may be described. A support history, a set of engineering records or an archive of job tickets you license could be summarized on a page that customers, competitors and journalists can read.

The requirement is already in force. It reaches systems released before the law took effect as well as new ones, and documentation is due again before each new covered release or substantial modification. The statute text sets the exact dates and how far back it reaches, so read those provisions with counsel rather than relying on secondary summaries.

Who must comply, and which systems are covered?#

AB 2013 places its duty on the developer, which the law describes broadly as a person or entity that designs, codes, produces or substantially modifies a generative AI system or service for use by members of the public. A company that only licenses records to a developer is usually not a developer for this purpose and has nothing to post.

Two boundary points matter to suppliers. First, the law treats retraining or fine-tuning that materially changes a system as a substantial modification, so a buyer that fine-tunes a public system on your records may still have to post documentation. Second, the duty attaches to systems offered for public use; a model used only inside a buyer's business may fall outside it, though plans can change. The statute also exempts a few narrow categories, such as systems whose sole purpose is security and integrity.

The indirect effect on suppliers is real. To write an accurate summary, a developer needs facts that only you hold, so expect diligence questionnaires and cooperation duties in the license.

What do AB 2013 disclosures cover, and what could each reveal?#

AB 2013 disclosures cover a list of high-level topics about each dataset used in training, and nearly all of them can touch a supplier. The table paraphrases the topics as the statute describes them; counsel should check the exact wording before relying on it.

The supplier's risk is less about trade secrets leaking and more about identification. A narrow description, such as dispatch records from a regional HVAC contractor, can point to one company even without a name, and several topics combined can narrow it further.

What do AB 2013 disclosures cover, and what could each reveal?
Disclosure topicWhat it could reveal about a supplierQuestion to settle in the license
Sources or owners of the datasetsYour company name, or a description that points to youNamed, described by industry, or grouped with other sources?
How the datasets serve the system's intended purposeThe task your records helped teach, such as resolving support casesIs a functional description enough, with no customer or product names?
Number of data points, which may be given as general rangesThe scale of your archiveWill only broad ranges be used?
Types of data pointsWhich record families you supplied, such as tickets or dispatch notesWhich wording for the record types is approved?
Whether datasets include copyrighted, trademarked or patented material, or are entirely public domainWhether your records carry IPWho characterizes IP status, and on what evidence?
Whether datasets were purchased or licensed by the developerThat a commercial data deal existsCan the deal type be confirmed without naming you?
Whether datasets include personal information or aggregate consumer informationWhether customer or employee details were in the source recordsDoes the statement match your de-identification accurately?
Cleaning, processing or other modification, and its purposeHow your records were preparedCan preparation be described without exposing your methods?
Collection period, and whether collection is ongoingHow far back your archive goes and whether a feed continuesWill dates be given in whole years, and does a refresh count as ongoing?
Dates the datasets were first used in developmentRoughly when the deal beganIs the year enough?
Whether the system used synthetic data generationWhether synthetic records were produced, possibly from yoursDoes the license allow synthetic generation at all?

What AB 2013 does not require developers to disclose#

AB 2013 does not require developers to disclose the commercial or record-level detail that suppliers usually worry about most. The documentation is a summary of datasets, and its topics stop well short of the deal itself.

Those limits are useful in negotiation. If a developer's draft says the law requires it to publish deal terms or customer details, it is asking for more than the statute's topics, and the license can confine any public statement to what the law actually requires.

  • The datasets themselves: nothing in the summary requires publishing or sharing records.
  • Prices, payment terms or other commercial terms; the topic is only whether data was purchased or licensed.
  • Names of the individual customers, employees or other people who appear in the records.
  • A supplier's consent before the developer posts; any consent right has to come from the contract.
  • Any posting by a company that only licensed records and did not develop or modify the system.

How to draft the required-by-law exception#

A confidentiality clause in a data license does not override a developer's legal duty to disclose, so the drafting task is to narrow the required-by-law exception without blocking compliance. Most licenses allow disclosure required by law, and a developer's counsel may decide that the sources-or-owners topic calls for identifying you.

General counsel usually tighten that exception in three ways. Limit any disclosure to the minimum the law requires. Tie it to an approved description of the dataset agreed at signing. And make the developer responsible for the accuracy of its own documentation, with the supplier answering only for the facts it provided. A plain sentence often does the work: any statutory training data summary will describe the licensed records using the approved description, and will not identify the licensor by name unless the licensee's counsel concludes in writing that the law requires it.

Make the clause survive termination and cover later versions. Because documentation is due again for each substantial modification, a model retrained on your records long after delivery can bring the description back into view, and the license should still govern it then.

Facts to have ready when a developer asks#

The facts to have ready are the ones each AB 2013 disclosure topic asks about, kept for each licensed dataset, because a supplier that answers quickly has the most influence over what the developer writes. Vague answers invite a developer to fill gaps with its own assumptions, and those assumptions end up in a public document.

Keep a short factual sheet for each licensed dataset, organized by the disclosure topics, and have counsel review it once. The same sheet answers diligence questionnaires, supports the approved description and gives you a reference point if a published summary later misstates something.

  • Sources or owners: the legal entity that holds the records and the originating systems, such as Zendesk and Jira for a software company.
  • Volume: an approximate record count you are comfortable seeing published as a broad range.
  • Types: record families included and excluded, in plain words.
  • Personal information: which personal and confidential details preparation removed, and how results were checked.
  • IP status: known third-party material that was carved out, such as attachments created by other companies.
  • Collection period: the date range of the records and whether refreshes are planned.
  • Deal type: licensed, not sold, with the company keeping ownership.

Illustrative: a consulting firm agrees how its records will be described#

Illustrative: a fictional operations consulting firm licenses its proposal library, project review notes and internal playbooks to a developer building a planning assistant that will be offered to the public. The developer's draft documentation schedule names the firm as a dataset owner, gives the exact date range of the records and says the data includes personal information.

The firm's general counsel works through the AB 2013 topics one by one. The parties agree to describe the source as project delivery records from a US management consulting firm, to give the collection period in whole years, and to state that client names and staff details were removed before delivery, so the personal information line reflects the prepared records rather than the raw archive. The developer confirms the data was licensed without stating terms, keeps a right to name the firm only if its counsel concludes in writing that the law requires it, and agrees to give advance notice in that case. The partners approve the license with the schedule attached.

How SourceX prepares suppliers for disclosure questions#

SourceX handles AB 2013 questions in the Rights and Approval steps of the SourceX five-step transaction. The provenance and privacy record in the SourceX Evidence Packet hold most of the facts a developer needs for its summary, such as where records came from, the period they cover and what preparation removed, so the developer starts from documented facts rather than assumptions.

The supplier approves every step, so the description of its records is settled alongside the license rather than discovered later on a developer's website.

Frequently asked questions

Does AB 2013 apply to developers based outside California?

AB 2013 is aimed at developers whose generative AI systems are made available to people in California, so the developer's headquarters is not the main test. For suppliers the question is different: wherever you are based, a covered developer may describe your dataset in its documentation. Counsel should confirm scope for any specific system.

Does fine-tuning a model on our records trigger AB 2013?

It can. The law treats retraining or fine-tuning that materially changes a system's functionality or performance as a substantial modification, so a developer that fine-tunes a publicly offered system on your records may need to post or update documentation. Whether a particular system is covered is a question for the developer's counsel.

Does the summary require the developer to publish our records?

No. The documentation is a high-level summary about datasets, not a release of the data. Your records stay governed by the license, including its access controls, permitted use and deletion terms.

Should we tell customers before a developer's documentation goes live?

That is a business decision worth planning. If your records came from customer interactions, a customer may recognize a description of them. Agreeing the wording in advance lets you decide whether to brief key customers, adjust your own privacy notice language, or leave things as they are.

Is AB 2013 the only law that asks for training data summaries?

No. The EU AI Act includes transparency obligations for providers of general-purpose AI models, including a public summary of training content, and other jurisdictions are considering similar rules. A disclosure clause in your license should cover any summary required by law, not only California's.

Related resources

See if your company qualifies

A short company assessment. No data uploads are needed.

See if you qualify