Bankruptcy Issues Raised by Generative AI: Ownership, Licensing, and Asset Values

A New Wave of Distress

The intersection of artificial intelligence (AI) and bankruptcy law remains largely uncharted. The first prominent AI insolvency of the current cycle, the Chapter 7 case of Builder.ai's parent, In re Engineer.ai Corp., filed in the District of Delaware in June 2025, turned largely on allegations of fraud and inflated or misrepresented AI capabilities (also known as"AI washing") rather than on contested principles governing the restructuring of genuine AI assets. It therefore did little to settle how courts will treat model weights, datasets, and AI licenses in bankruptcy. Those questions remain open even as market conditions make them increasingly likely to be litigated: forecasters project failure rates among AI startups that would be extraordinary by historical standards, driven by inference costs that scale faster than revenue and by foundation-model providers absorbing the functionality of thinner application-layer companies.

Characterizing the AI Estate

The threshold difficulty is identifying what the debtor actually owns. A generative AI business typically depends on several distinct categories of assets: an underlying foundation model (often licensed from a third party), fine-tuned model weights derived from that foundation, training and evaluation datasets, the software and pipelines used to train and serve the model, and the outputs the model generates. Each category presents a separate ownership question.

Model weights, the numeric parameters that encode a trained model's capabilities, are commonly protected, if at all, as trade secrets rather than through registration. That characterization matters because the Bankruptcy Code's definition of "intellectual property" in Section 101(35A) includes trade secrets, patents, copyrights, and mask works, but pointedly omits trademarks and says nothing specific about AI artifacts. Whether a given set of weights qualifies as a protectable trade secret depends on the reasonable secrecy measures the debtor maintained, a fact-intensive inquiry that distressed companies, having often cut security spending and personnel, may struggle to satisfy.

Training data introduces a second layer. Much of it is licensed rather than owned, and those licenses frequently restrict use to the original licensee and prohibit transfer. The copyright status of AI-generated outputs remains unsettled and, to the extent human authorship is lacking, generally unprotectable; in Thaler v. Perlmutter, the D.C. Circuit affirmed in 2025 that a work generated autonomously by a machine, without a human author, is not eligible for copyright registration. That can leave a debtor's most commercially visible product outside the estate's registrable intellectual property. Counsel should map these categories early, because the answers drive every subsequent question about assumption, sale, and value.

Assumption and Assignment Under Section 365

Most AI arrangements are executory contracts, foundation model licenses, data licenses, cloud and GPU compute agreements, and API access terms all involve material unperformed obligations on both sides. Section 365 lets a debtor assume the contracts it wants to keep and reject those it does not, but the provision's limits are especially consequential here.

Section 365(c)(1) prohibits a debtor from assuming or assigning an executory contract without the counterparty's consent when "applicable law" would excuse the counterparty from accepting performance from anyone other than the debtor. Federal intellectual property law supplies exactly that kind of restriction: nonexclusive patent and copyright licenses generally are not assignable without the licensor's consent. Many foundation model and data licenses are nonexclusive and expressly non-transferable, which means they may fall squarely within Section 365(c)(1).

The circuits are split on how far this restriction reaches. Under the "hypothetical test" followed by the Ninth Circuit in In re Catapult Entertainment, a debtor in possession cannot even assume such a contract for its own continued use, the mere hypothetical inability to assign to a third party bars assumption. The "actual test," applied by the First Circuit and others, permits assumption where the debtor intends to keep performing itself and does not propose an actual assignment. For an AI company reorganizing around a licensed foundation model, the applicable circuit's rule can determine whether the business survives at all. A debtor in a hypothetical-test jurisdiction may find that the license anchoring its product cannot be retained without the licensor's cooperation, cooperation the licensor may withhold or price aggressively.

Open-source and "open weight" model licenses add nuance. Community licenses for widely used models often condition continued use on compliance with usage restrictions or monthly-active-user thresholds, and a change of control through a sale may trigger termination or renegotiation. The recent trend toward shared and pooled licensing compounds the issue: industry consortia, such as the Shared AI License Foundation launched in 2026 to give members non-exclusive access to foundational patent rights, reduce friction in ordinary times but can leave a debtor holding precisely the kind of non-exclusive, consent-dependent licenses that Section 365(c)(1) makes so difficult to preserve or assign. Counsel should not assume that a permissively or collectively licensed model travels freely with the estate.

Licensee Protections: Section 365(n) and Mission Product

When the debtor is the licensor, different protections apply. Section 365(n) allows a licensee of "intellectual property" to elect to retain its license rights even after the debtor-licensor rejects the agreement, continuing to pay royalties in exchange. Because Section 101(35A) omits trademarks, Section 365(n) does not by its terms protect trademark licensees, a gap the Supreme Court addressed in Mission Product Holdings v. Tempnology (2019), holding that rejection operates as a breach rather than a rescission, so a licensee may generally continue exercising the rights it held under non-bankruptcy law.

For AI licensees, the key question is whether the licensed subject matter is "intellectual property" within Section 101(35A). A license to trade-secret-protected model weights likely qualifies, giving the licensee a Section 365(n) election. A license framed around trademarks, access services, or unprotectable outputs may not, leaving the licensee to rely on the narrower protection of Mission Product. Licensees dependent on a distressed vendor's model should assess which regime governs before distress arrives, not after.

Section 363 Sales and the Data-Privacy Dimension

Distressed AI businesses are frequently sold as going concerns under Section 363, often free and clear of liens and interests. Several cautions apply. First, a Section 363 sale cannot be used to strip a licensee of the retention rights that Section 365(n) preserves; those protections generally survive the sale. Second, the free-and-clear order is only as good as the estate's title, if the debtor's rights to weights or data rest on non-transferable licenses, the buyer may acquire far less than the deal assumed.

Third, and increasingly central, is data privacy. The very training data and user records that give an AI business much of its value are often laden with personal information, and Section 363(b)(1) directly restricts their sale. Codifying the Federal Trade Commission's position in the Toysmart.com matter, that provision bars a trustee from selling personally identifiable information contrary to a privacy policy the debtor had in effect on the petition date, unless the sale is consistent with that policy, or a consumer privacy ombudsman is appointed under Section 332 and the court, after notice and hearing, finds no showing that the sale would violate applicable law.

The 23andMe bankruptcy brought these provisions to the fore. In its 2025 Chapter 11 case, the debtor sought to sell the genetic and personal data of millions of customers; state attorneys general and the FTC intervened, an ombudsman was appointed, and the court ultimately approved a sale to a nonprofit affiliated with the company's founder. Two aspects of the court's analysis are instructive. It held that Section 363(b)(1) does not preempt applicable non-bankruptcy law, so state privacy statutes continue to constrain the sale, and that the relevant "policy" is the formal privacy policy in effect on the petition date, not the accumulated history of website statements. The case also exposed a statutory gap: the Code's PII provisions do not expressly address genetic or biometric data, an omission that has since prompted legislative proposals for heightened protection.

For AI estates, these principles carry directly. Training corpora and user-interaction logs frequently contain personal information, and selling them into a buyer's model may conflict with the debtor's own privacy representations and with regimes such as the California Consumer Privacy Act and the European General Data Protection Regulation (GDPR), which restrict secondary use and cross-context transfer. A buyer intent on ingesting acquired data into a model should expect to inherit those constraints, to agree to use limitations or re-consent mechanisms, and, where the debtor's policy prohibits transfer, to proceed only through the ombudsman process.

Valuation

Valuation compounds every one of these difficulties. Traditional IP valuation methods strain against assets whose legal status is uncertain, whose value depends on continued access to compute and talent, and which depreciate rapidly as newer models emerge. A model's worth may collapse if the training data cannot be conveyed for privacy reasons, if key engineers depart, or if the underlying foundation license terminates on the sale. A specialized market of distressed-asset buyers has emerged for exactly these assets, which makes accurate valuation both more feasible and more consequential: surfacing and properly pricing data repositories can materially affect creditor recoveries, while overlooking transferability constraints risks the estate promising something it cannot deliver.

Practical Takeaways

Several steps can reduce risk before distress materializes. License agreements should address bankruptcy expressly, including consent-to-assignment provisions, Section 365(n) acknowledgments, and source-code or model-weight escrow triggered by insolvency events, a protection gaining traction as vendor failures mount. Companies should document ownership of weights, datasets, and outputs; maintain the secrecy measures necessary to preserve trade secret status; and keep privacy policies clear about whether data may be transferred in a sale or bankruptcy. Lenders should confirm that their security interests attach to, and are perfected in, the specific AI assets, recognizing that weights and data may not fit neatly within standard collateral descriptions. Licensees should identify their most business-critical AI dependencies and negotiate durable protections, including escrow and data-return rights, in advance. And acquirers should treat both transferability and privacy compliance, not headline capability, as central diligence questions. Generative AI has outpaced the statutory framework built to govern intellectual property and data in bankruptcy; careful contracting and early planning remain the most reliable protections.

If you have questions about AI-related bankruptcy risks or would like assistance evaluating these issues, please contact Jake Adams, Edward D. Lanquist, or any member of Baker Donelson's Intellectual Property Group.