Key Takeaways
- GLBA regulates disclosure of nonpublic personal information to nonaffiliated third parties; none of its texts reviewed mentions AI or model training.
- A vendor that trains or fine-tunes for you needs notice and opt-out or an exception, usually the service-provider exception with the initial notice and a contract limiting use (interpretation).
- Reuse limits follow the data: a recipient under a processing or other exception may use it only for that activity, and a service-provider contract must limit use.
- GLBA rules exclude aggregate or blind data but set no de-identification test; test whether training sets and models can reveal customer records (interpretation).
- Regulation S-P's amended safeguards and disposal rules have applied to every covered institution since June 3, 2026 (disposal excepts notice-registered broker-dealers); applying them to training copies is interpretation.
Current as of October 30, 2026. This guide draws on the privacy provisions of the Gramm-Leach-Bliley Act (15 U.S.C. 6801-6809), the CFPB's Regulation P (12 CFR part 1016), subpart A and section 248.30 of the SEC's Regulation S-P (17 CFR part 248), the FTC Safeguards Rule (16 CFR part 314), the OCC's information security guidelines (12 CFR part 30, appendix B), the GLBA exemption in the California Consumer Privacy Act, NIST AI 600-1 and FINRA's 2026 oversight report, as published on that date. It covers US law only and is not legal advice.
None of the GLBA texts reviewed for this guide mentions artificial intelligence or machine learning. Their privacy rules regulate disclosure. The statute says a financial institution "may not, directly or through any affiliate, disclose to a nonaffiliated third party any nonpublic personal information, unless such financial institution provides or has provided to the consumer a notice that complies with section 6803 of this title." The questions for a training project are which data is in scope, whether any reaches a nonaffiliated party, under which exception, and on what terms.
Unless a step quotes a rule, it is a recommendation, labeled (interpretation). Where this guide applies the texts to training, fine-tuning, model weights or memorization, it is interpretation.
- Settled: Regulation P's service-provider exception (section 1016.13) lifts the opt-out requirement only if the institution gives the initial notice and enters into a contract "that prohibits the third party from disclosing or using the information other than to carry out the purposes for which you disclosed the information". Regulation S-P section 248.13 uses the same words.
- Settled: Regulation P excludes from personally identifiable financial information "Information that does not identify a consumer, such as aggregate information or blind data that does not contain personal identifiers such as account numbers, names, or addresses." Neither the statute nor Regulation P or subpart A of Regulation S-P sets a de-identification method or test.
- Settled: the 2024 Regulation S-P amendments are final, effective August 2, 2024, with compliance required by December 3, 2025 for larger entities and June 3, 2026 for smaller ones. A Federal Register search by CFR part found no document affecting parts 1016, 248 or 314 published since January 1, 2025.
- Not settled: whether in-house training is a "disclosure" (Regulation P does not define the word), whether trained model weights are themselves nonpublic personal information, and whether a vendor may use the data to improve models it offers to others. The texts reviewed do not address these questions.
On scope: the GLBA privacy provisions are binding federal law. Regulation P covers institutions within the CFPB's GLBA rulemaking authority, including "third parties that are not financial institutions but that receive nonpublic personal information from financial institutions with whom they are not affiliated"; it excludes registered broker-dealers and investment companies, SEC- or state-registered investment advisers (for advisory activities), persons under CFTC jurisdiction, insurers for supervised insurance activities and certain auto dealers (section 1016.1). Regulation S-P subpart A covers brokers, dealers, investment companies and SEC-registered investment advisers; section 248.30, as amended, also reaches registered transfer agents. Apart from section 248.30, both protect consumers only: Regulation P "does not apply to information about companies or about individuals who obtain financial products or services for business, commercial, or agricultural purposes." The FTC Safeguards Rule covers financial institutions within FTC jurisdiction, such as mortgage lenders, finance companies and "investment advisors that are not required to register with the Securities and Exchange Commission". The OCC guidelines apply to national banks, federal savings associations, federal branches and agencies of foreign banks, and certain subsidiaries; other banks should check their own regulator's version (interpretation).
Step 1: Classify the training data as NPI, consumer or business, and by source
Regulation P defines nonpublic personal information as personally identifiable financial information plus "Any list, description, or other grouping of consumers (and publicly available information pertaining to them) that is derived using any personally identifiable financial information that is not publicly available." Personally identifiable financial information covers information "You otherwise obtain about a consumer in connection with providing a financial product or service to that consumer", and its examples include "Account balance information, payment history, overdraft history, and credit or debit card purchase information".
Before anything is copied into a training environment, tag every field and file, as NPI or not, consumer or business, and by source (interpretation).
Step 2: Decide whether training involves any disclosure to a nonaffiliated third party
The notice rule in 15 U.S.C. 6802(a) and the opt-out rule in 6802(b) apply to disclosures to a nonaffiliated third party, which Regulation P defines as "any person except" an affiliate or a jointly employed person. Regulation P defines "collect" but not "disclose". Training by the institution's own staff on infrastructure it controls, with no outside access, is on the face of these texts not such a disclosure (interpretation). An outside provider with access is a nonaffiliated third party; go to Step 3 (interpretation). Affiliate sharing has its own rules, outside this guide.
Internal use is not unregulated. Congress stated that each institution "has an affirmative and continuing obligation to respect the privacy of its customers and to protect the security and confidentiality of those customers' nonpublic personal information." Confirm the privacy notice still describes actual practices (interpretation).
Step 3: When a vendor trains or fine-tunes, name the GLBA exception that applies
The statute's service-provider exception allows providing NPI "to a nonaffiliated third party to perform services for or functions on behalf of the financial institution" if the institution "fully discloses the providing of such information and enters into a contractual agreement with the third party that requires the third party to maintain the confidentiality of such information." Section 1016.13 makes it an exception to the opt-out only; the initial notice is still required, and section 1016.6(a)(5) calls for "a separate statement of the categories of information you disclose and the categories of third parties with whom you have contracted" where no section 1016.14 or 1016.15 exception also applies.
Section 1016.14 covers disclosures "as necessary to effect, administer, or enforce a transaction that a consumer requests or authorizes", and section 1016.15 covers, among others, disclosures "To protect against or prevent actual or potential fraud, unauthorized transactions, claims, or other liability" and those made "With the consent or at the direction of the consumer, provided that the consumer has not revoked the consent or direction". A fraud model built by a vendor may fit the fraud exception; general model development is harder to fit within section 1016.14 (interpretation).
A vendor training a product it sells to others is not, for that use, acting on your behalf; that use would need notice and opt-out or consent (interpretation).
Step 4: Write the purpose limit and safeguards into the vendor contract
Limit the permitted purpose to building models for the institution; bar use of the data, prompts, outputs or fine-tuned weights to train models for the vendor or its other clients; bind subcontractors; and require deletion at the end (interpretation).
The FTC rule and the OCC guidelines both require service providers to be bound by contract to maintain safeguards. Regulation S-P section 248.30(a)(5) requires policies for oversight of service providers "including through due diligence and monitoring", designed to ensure they notify the covered institution within 72 hours of learning of a breach of a system they maintain; the SEC said it was "adopting the rule without a written contract requirement" for that provision. Practices described in FINRA's 2026 report for member firms include "adding language that prohibits firm or customer sensitive information from being ingested into a third-party vendor's open-source GenAI tool"; the report "does not create any new legal or regulatory requirements or new interpretations of existing requirements".
Step 5: Apply the reuse and redisclosure limits in both directions
The statute bars a recipient from passing NPI to a party unaffiliated with both it and the institution "unless such disclosure would be lawful if made directly to such other person by the financial institution." Under Regulation P section 1016.11(c), a third party that received NPI under section 1016.14 or 1016.15 may disclose and use it only "pursuant to an exception in § 1016.14 or § 1016.15 in the ordinary course of business to carry out the activity covered by the exception under which it received the information." A disclosure under section 1016.13 alone falls under section 1016.11(d), which limits redisclosure, while the section 1016.13 contract limits use (interpretation).
Where your institution is the recipient of data for account processing, the regulation's example says "You could not disclose that information to a third party for marketing purposes or use that information for your own marketing purposes." Training a model for an unrelated product on such data raises the same problem (interpretation). Regulation S-P section 248.11 is parallel.
Step 6: Treat de-identification as a claim you must test, not a label
The aggregate or blind data exclusion is the only de-identification language in the GLBA definitions reviewed, and a grouping derived from NPI remains NPI. Removing names and account numbers does not by itself show that a data set or model no longer identifies anyone (interpretation).
NIST AI 600-1 (July 2024), a profile of the AI Risk Management Framework, which NIST describes as intended for voluntary use, warns that "Models may leak, generate, or correctly infer sensitive information about individuals." It suggests to "Use techniques such as anonymization, differential privacy or other privacy-enhancing technologies to minimize the risks associated with linking AI-generated content back to individual human subjects." Test models for extraction of training records, and treat a model as covered data until testing supports otherwise (interpretation).
Step 7: Apply safeguards, retention and disposal to training sets and models
Regulation S-P section 248.30(a)(1): "Every covered institution must develop, implement, and maintain written policies and procedures that address administrative, technical, and physical safeguards for the protection of customer information." Section 248.30(b) requires covered institutions, other than notice-registered broker-dealers, to "properly dispose of consumer information and customer information" with reasonable measures, but does not itself require keeping or destroying any record. The FTC rule requires disposal of customer information "no later than two years after the last date the information is used in connection with the provision of a product or service to the customer to which it relates", unless, among other exceptions, it "is necessary for business operations or for other legitimate business purposes"; the OCC guidelines also require disposal measures.
Inventory training data, checkpoints and vendor-held copies, and apply retention and incident response to each (interpretation).
Step 8: Check state privacy law against the GLBA exemption, data set by data set
The California Consumer Privacy Act, as compiled by the California Privacy Protection Agency effective January 1, 2026, "shall not apply to personal information collected, processed, sold, or disclosed subject to the federal Gramm-Leach-Bliley Act (Public Law 106-102), and implementing regulations", and it adds: "This subdivision shall not apply to Section 1798.150." The exemption covers information, not institutions, so data not subject to GLBA may still be covered (interpretation). This guide reviewed the CCPA, not the California Financial Information Privacy Act.
Checklist
Items restate the recommendations above (interpretation).
- Training fields are tagged by NPI status, consumer or business, and source
- Each data flow is classified as internal, affiliate or nonaffiliated third party
- The GLBA exception relied on is recorded for every vendor flow
- The privacy notice still describes actual practices
- Vendor contracts limit use to the institution's purposes and bar training for others
- Reuse limits are checked for data from other institutions
- De-identification claims are tested
- Training copies follow retention and disposal schedules
- State law is checked for data not subject to GLBA
How we sourced this guide: every factual statement in it is drawn from published rules, regulatory guidance and other public documents, and we keep a full set of sources and citations for it. If you would like them, or want to question a point, get in touch through our contact page and we will share them.
Frequently Asked Questions
Does GLBA require customer consent before we train an AI model on customer data?
Not as such. The GLBA notice and opt-out rules apply to disclosures of nonpublic personal information to nonaffiliated third parties, and none of the texts reviewed mentions model training. Training done in-house with no outside access is, on the face of the texts, not such a disclosure (interpretation). A vendor working on your behalf can rely on the service-provider exception with the initial notice and a contract limiting use; consent is one route under section 1016.15 where no other exception fits.
Can our AI vendor use our customer data to improve its own models?
Under the service-provider exception, the contract must prohibit the vendor from using the information other than to carry out the purposes for which you disclosed it. Improving models the vendor offers to others is not, on our reading, a service performed for you, so it would need the notice and opt-out path or consumer consent (interpretation). The texts reviewed do not address model training specifically.
Is de-identified or aggregate customer data outside GLBA?
Regulation P and Regulation S-P exclude information that does not identify a consumer, such as aggregate information or blind data without personal identifiers. Neither sets a de-identification method or test, and a grouping of consumers derived from nonpublic personal information is itself nonpublic personal information. Whether a data set or a trained model still identifies anyone should be tested and documented (interpretation).
Does the California Consumer Privacy Act apply to our training data?
The CCPA does not apply to personal information collected, processed, sold or disclosed subject to GLBA and its regulations, except for section 1798.150, which addresses personal information security breaches. The exemption covers information rather than institutions, so data that is not subject to GLBA may still fall under the CCPA (interpretation). This guide reviewed the CCPA, not the California Financial Information Privacy Act.
