Aishwarya Tiwari, LL.M. Class of 2027
Here is the uncomfortable truth about AI model distillation: it is not obviously illegal. Model distillation is the practice of using the outputs of one AI system, its responses to thousands of carefully crafted prompts, to train a smaller, cheaper competing model. A competitor can pay for API access, comply with rate limits, maintain valid accounts, and still extract what makes a rival model worth a billion dollars. No lock is picked. No server is breached.
The existing legal response is a patchwork: copyright, trade secret doctrine, contract, and the Computer Fraud and Abuse Act. None were designed for distillation, and none are sufficient alone. This post examines each and argues that Congress should craft a narrowly tailored rule targeting the method and scale of extraction.
Copyright: No Protection for Machine-Generated Outputs
Copyright is the intuitive first theory, but it fails at the threshold. Under Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), copyright requires human authorship. AI-generated outputs, the very responses a distiller harvests, receive no copyright protection. The D.C. Circuit’s holding was cemented on March 2, 2026, when the Supreme Court denied certiorari.
Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025), complicates things further. The court held that training on legitimately obtained books was fair use, exceedingly transformative, producing a system capable of tasks no single book could accomplish. What was not fair use was retaining a permanent library of pirated copies. The court drew the line at how source material is handled, not training itself, a boundary that does little to constrain systematic API harvesting of AI outputs, which are often not copyrightable at all.
Trade Secrets: The Secrecy Problem and an Opening
Trade secret law is an even worse fit at first glance, because outputs from a public API are not secret. But Compulife Software Inc. v. Newman, 959 F.3d 1288 (11th Cir. 2020), aff’d, 111 F.4th 1147 (11th Cir. 2024), suggests that secrecy is not always the end of the analysis.
In Compulife, defendants hired a hacker to scrape 43 million insurance quotes from Compulife’s public website and built a competing service. Each quote was accessible to any user. But the Eleventh Circuit accepted that the database as a whole could be a trade secret, and reasoned that while Compulife gave the public “implicit permission” to pull as many quotes as humanly possible, “using a bot to collect an otherwise infeasible amount of data” could be improper means. In 2024, the court affirmed that the scraping was improper, reiterating that conduct can be improper “even if not independently unlawful.”
The logic maps onto distillation imperfectly. A model’s weights may qualify as a trade secret, but the API responses used to approximate them are public by design. Courts would need to extend Compulife’s framework, asking whether the scale and systematic nature of extraction transforms otherwise lawful API access into misappropriation under the Uniform Trade Secrets Act. Compulife suggests scale can be improper; the open question is whether extraction within a provider’s express permission ever can be.
Contract and the CFAA: Partial Tools with Structural Limits
Contract is today’s most direct tool. Major AI providers ban using outputs to train competing systems in their terms of service. When a direct customer violates that prohibition, there is a viable breach-of-contract claim.
But contract law has a privity problem. The prohibition binds the customer who signs the terms, not a third party who acquires outputs second-hand or builds a tool that sits between the customer and the provider. A competitor who routes queries through intermediaries, or purchases harvested outputs from a data broker, may breach no contract at all.
The Computer Fraud and Abuse Act, 18 U.S.C. § 1030, faces a parallel limit. The CFAA prohibits unauthorized access to computer systems. But a distiller with a valid API key and a paid subscription is not accessing anything without authorization; they are using the system as designed, just at massive scale and for an unauthorized purpose. Using a tool in an unintended way is not the same as breaking in.
Four doctrines, four partial answers. A competitor can pay for API access, route queries through intermediaries to avoid privity, and extract a rival’s core competitive asset without picking a lock, breaching a contract they signed, or accessing anything without authorization.
Congress Has Noticed—But Only Part of the Problem
Congress has recognized distillation as a national security concern, though the legislative response so far targets only part of the threat. H.R. 8283, the Deterring American AI Model Theft Act, was introduced in April 2026 and ordered reported out of the House Foreign Affairs Committee. S. 5252, the BLADE Act (Blocking Large-scale Adversarial Distillation Efforts Act), was introduced in August 2026 and referred to the Senate Banking Committee. Both bills remain in committee. In September 2026, NSA, CISA, and FBI issued Joint Advisory AA26-251A, designating industrial-scale distillation by six Chinese AI companies a national security threat.
Both bills, however, limit their scope to foreign adversaries and foreign entities of concern. The domestic distillation problem, a U.S.-based competitor quietly building a model on another American company’s outputs, falls entirely outside their reach.
What the Law Should Do
Closing the domestic gap requires a targeted federal statute. Compulife’s improper-means framework already reaches scale and gives courts a tool against bot-driven extraction. But the defendants there exceeded the implicit permission a public website extends, and the 2024 court noted they could not have done it without copying the order of Compulife’s copyrighted code. A distiller does neither: it is a paying API customer with express permission to query. Legislation should extend the improper-means concept to systematic, large-scale extraction through authorized access designed to replicate a competitor’s model—the gap Compulife does not reach on its own.
Any such statute should be narrowly drawn—it should not prohibit legitimate research, interoperability testing, or competitive benchmarking. The target conduct is distinguishable from ordinary API use by volume, pattern, and purpose.
Conclusion
AI model distillation is not obviously illegal, and that is precisely the problem. Copyright leaves AI outputs unprotected. Trade secret law requires secrecy that API-accessible outputs lack. Contract reaches direct customers but not intermediaries. The CFAA requires unauthorized access that a paid subscription forecloses.
The national security legislation in Congress acknowledges the threat but addresses only foreign actors. The domestic version of the same problem, one company quietly building on another’s outputs, remains unaddressed. Courts can use Compulife as a bridge where access was never permitted at that scale. Congress needs to build the road for when it was.