Trade secrets in AI training data licenses: what attorneys should flag

The short answer: it depends on what is delivered and on what terms

Licensing company records as AI training data does not have to cost a client its trade secrets, but it can. The outcome turns mainly on two things counsel controls: what actually leaves the company, and the confidentiality, use and access terms that travel with it. Records with no secret content raise little trade-secret risk; formulas, source code, pricing logic or customer-specific know-how delivered on loose terms can raise a great deal.

In practice the protective work happens before delivery. Counsel narrows the scope, carves out what the client most needs to keep secret, agrees redaction rules, and writes use limits that stop the licensee from disclosing or exploiting the material beyond the agreed AI-training purpose.

What the law looks at

US trade-secret protection comes from federal and state law, and the wording of the elements varies by jurisdiction. In broad terms, counsel will ask two questions about any record set: does the information have value because it is not generally known, and has the owner taken reasonable steps to keep it secret? A license that hands secret material to a third party without confidentiality obligations makes the second question harder to answer. Confirm the exact test in the statute and case law that govern your client.

Three other bodies of law usually sit beside trade secrets in the same deal:

• Copyright ownership. Ownership of copyright may be transferred in whole or in part, and any of the exclusive rights may be transferred and owned separately, under 17 U.S.C. § 201. That divisibility is what lets a company grant an exclusive AI-training license for a term while keeping ownership of the underlying material.
• Promises to customers. FTC staff wrote in a January 2024 post on privacy and confidentiality commitments that a company's promises not to use customer data for undisclosed purposes, such as training models, are enforceable wherever they were made, including privacy policies, terms of service and marketing. It is staff guidance, not a rule, but it tells you to read what the client promised.
• Personal and health data. California's CCPA requires a written agreement limiting use when a business sells or shares personal information or discloses it to a service provider or contractor (Cal. Civ. Code § 1798.100 et seq.), and HHS describes the two methods for de-identifying protected health information under the HIPAA Privacy Rule in its de-identification guidance.

How it applies in common deal situations

Contract terms counsel typically negotiates

These points come up in most data licenses that touch confidential material. Treat them as a starting list, not a form.

1. A dataset schedule. Define exactly which systems, date ranges and record types are licensed, so anything not listed is excluded by default.
2. Exclusions and carve-outs. Name the categories that never leave the company, such as core source code, formulas, security documentation and board materials.
3. A redaction and de-identification specification. Fix the rules for names, identifiers, prices and third-party information before any preparation work begins.
4. Use restrictions. Limit use to the agreed AI training and evaluation purpose for the agreed term, with no onward licensing or disclosure.
5. Confidentiality and security obligations. Require stated protective controls and limit access to the licensee's people who need it.
6. No residuals clause. Strike language letting the licensee's staff freely use information retained in memory; it undercuts the confidentiality terms.
7. Output and extraction safeguards. Ask how the licensee handles the risk of a trained model reproducing distinctive passages, and prohibit attempts to re-identify people or reconstruct delivered records.
8. End-of-term handling. Say what happens to delivered data when the term ends, including deletion or return and a certificate.

In a SourceX deal, the company keeps ownership and licenses its records rather than selling them. Redaction and de-identification requirements are agreed with the company before any work begins, and nothing is delivered without an executed agreement and the company's authorization; the how it works page sets out the sequence. The client can walk away at any point before signing.

Disclosure and consent good practice

• Read the client's own commitments first: privacy policy, customer terms, master services agreements and employee notices. The piece on why consent matters in AI data explains why buyers check the same documents.
• Inventory third-party NDAs that could cover information sitting in email, chat or shared drives.
• Keep a written record of the client's protective measures: access controls, confidentiality agreements and the carve-out list. That record supports the trade-secret position if it is ever challenged.
• Bring the client's IT lead in early. Access controls, credential scanning and export logs are technical work; the guide to CIO peer groups shows how IT leaders approach these projects, and an EOS Integrator's responsibilities show who usually runs the internal workstream.

Questions to put to the client before any delivery

• Which five pieces of information would hurt most if a competitor saw them, and where do they live?
• Which systems hold third parties' confidential information under contract?
• Has the company ever promised customers, in writing, that their data would not be shared or used to train models?
• Who created the records: employees, contractors, agencies or clients?
• Who can authorize the license, and who will own the redaction review?

This is general information, not legal, tax or financial advice. Trade-secret, privacy and contract rules vary by jurisdiction and by deal; confirm with your own counsel or professional body before acting.

If you want to introduce a client yourself

Anyone can register as a SourceX partner, but referral compensation for lawyers is governed by each state's version of the professional conduct rules. The ABA's Model Rules of Professional Conduct address fees (Rule 1.5), professional independence (Rule 5.4), payments for recommendations (Rule 7.2) and business transactions with clients (Rule 1.8); check how your state adopted them, and whether client disclosure or consent is required, before accepting anything.

Partners earn 25% of the eligible platform fees SourceX actually collects from the referred company's licensing deals, capped at $100,000 per referred company, and the reward becomes payable only after the buyer pays and SourceX receives its fee. The reward is never deducted from what the client receives. Partners make introductions and share basic fit information only; they never export or describe a client's confidential records.

Next step

If a client holds years of operational records and the rights to license them, review the baseline on who qualifies. To make introductions, register as a partner; the client can also apply directly at sourcex.si/apply.