Protecting an AI Model by Patent or Trade Secret: What the Statutes and the Supreme Court Require

Protecting an AI model by patent or trade secret means choosing between two U.S. regimes with opposite conditions. A patent is a right to exclude others for up to 20 years from filing, granted on a specification that enables a skilled person to make and use what is claimed; a trade secret is information protected for as long as its owner keeps it secret by reasonable measures. In Kewanee Oil Co. v. Bicron Corp., 416 U.S. 470, 474 (1974), the Supreme Court held that state trade-secret law “is not pre-empted by the patent laws,” so both remain available. More on AI and patents is on the Tech & AI hub.

What is the basic trade-off between a patent and a trade secret?

A patent exchanges public disclosure for exclusivity against everyone; a trade secret keeps information private but protects only against wrongful taking. Kewanee put it this way: trade-secret law does not forbid discovery “by fair and honest means, e. g., independent creation or reverse engineering,” while patent law operates “against the world.” And: “Where patent law acts as a barrier, trade secret law functions relatively as a sieve.” (416 U.S. at 490.) The Court also concluded that trade-secret protection for “clearly patentable inventions does not conflict with the patent policy of disclosure.” (416 U.S. at 491.)

Can you patent an AI model?

Yes, if the claims meet the conditions that apply to any invention: eligibility under § 101, novelty under § 102, non-obviousness under § 103 and the disclosure rules of § 112 (see the explainer on what makes an invention patentable). On eligibility, the Federal Circuit held in Recentive Analytics, Inc. v. Fox Corp., No. 2023-2437 (Fed. Cir. Apr. 18, 2025) (precedential), “only that patents that do no more than claim the application of generic machine learning to new data environments, without disclosing improvements to the machine learning models to be applied, are patent ineligible under § 101.” (Slip op. 18.) The Supreme Court denied certiorari on 8 December 2025.

What must a patent application disclose?

Enough to enable the full scope of what is claimed. 35 U.S.C. § 112(a) requires “a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art” to make and use it, and the “best mode contemplated by the inventor.” In Amgen Inc. v. Sanofi, 598 U.S. 594, 610 (2023), the Court said “the specification must enable the full scope of the invention as defined by its claims. The more one claims, the more one must enable.” The explainer on full-scope enablement covers that decision.

Best mode is still required, but under 35 U.S.C. § 282(b)(3)(A), as amended by the America Invents Act, failing to disclose it “shall not be a basis on which any claim of a patent may be canceled or held invalid or otherwise unenforceable.” Because enablement is measured against the claims, how much of an architecture, training method, data set or set of trained weights an application must describe turns on what is claimed and on the state of the art; neither the statute nor these decisions address AI components by name.

When does a patent application become public?

Promptly after 18 months from the earliest filing date for which benefit is sought, or earlier on the applicant’s request (35 U.S.C. § 122(b)(1)(A)). Until then § 122(a) requires applications to “be kept in confidence,” subject to listed exceptions, and an application that is “no longer pending” is not published (§ 122(b)(2)(A)(i)).

An applicant may instead request nonpublication at filing by certifying that the invention “has not and will not be the subject of an application filed in another country, or under a multilateral international agreement, that requires publication of applications 18 months after filing” (§ 122(b)(2)(B)(i)). A later foreign filing must be notified within 45 days, or the application is “regarded as abandoned” (§ 122(b)(2)(B)(iii)).

How long does each kind of protection last?

A U.S. patent ends 20 years from its filing date, and a trade secret has no fixed term. Under 35 U.S.C. § 154(a)(2), the term ends “20 years from the date on which the application for the patent was filed in the United States,” or from the earliest U.S. application it relies on under §§ 120, 121, 365(c) or 386(c); the explainer on patent term covers adjustments and extensions.

Neither the DTSA nor the Uniform Trade Secrets Act sets a duration; information is protected while it meets the definition. In Ruckelshaus v. Monsanto Co., 467 U.S. 986, 1002 (1984), a Takings Clause case about data submitted to the EPA, the Court said: “If an individual discloses his trade secret to others who are under no obligation to protect the confidentiality of the information, or otherwise publicly discloses the secret, his property right is extinguished.”

What does the Defend Trade Secrets Act require?

Two conditions, set out in 18 U.S.C. § 1839(3): the owner “has taken reasonable measures to keep such information secret,” and the information “derives independent economic value, actual or potential, from not being generally known to, and not being readily ascertainable through proper means by, another person who can obtain economic value” from it. The Defend Trade Secrets Act of 2016, Pub. L. 114-153, enacted on 11 May 2016, gives the owner a federal civil action if the secret “is related to a product or service used in, or intended for use in, interstate or foreign commerce” (18 U.S.C. § 1836(b)(1)).

A court may grant an injunction and damages and, for willful and malicious misappropriation, exemplary damages of up to twice the damages (§ 1836(b)(3)). Suit must be brought within 3 years after the misappropriation is or should have been discovered (§ 1836(d)). The Act does not displace state remedies (§ 1838); the Uniform Trade Secrets Act, a model act for the states, has a parallel two-part definition in § 1(4).

Are model weights a trade secret?

Neither statute names model weights, architectures or training data, so each component qualifies or not under the same two conditions as any other information. The DTSA’s definition reaches “scientific, technical, economic, or engineering information,” including “compilations,” “methods, techniques, processes,” and “programs, or codes”; the UTSA lists “a formula, pattern, compilation, program, device, method, technique, or process.” For any component, the statutory questions are whether the owner has taken reasonable measures to keep it secret and whether it is “readily ascertainable through proper means” by others.

What does trade-secret law not protect against?

Lawful discovery. Under 18 U.S.C. § 1839(6), improper means include “theft, bribery, misrepresentation, breach or inducement of a breach of a duty to maintain secrecy, or espionage,” but do “not include reverse engineering, independent derivation, or any other lawful means of acquisition.” Kewanee said trade-secret law “does not offer protection against discovery by fair and honest means, such as by independent invention, accidental disclosure, or by so-called reverse engineering.” (416 U.S. at 476.)

Can a model used in secret be patented later?

Only if the secret commercial activity has not become prior art before the effective filing date. 35 U.S.C. § 102(a)(1) bars a patent where the claimed invention was “in public use, on sale, or otherwise available to the public” before that date, subject to the one-year exception in § 102(b)(1) for the inventor’s own disclosures. In Helsinn Healthcare S. A. v. Teva Pharmaceuticals USA, Inc., 586 U.S. 123 (2019), the Court held (Part III) that “an inventor’s sale of an invention to a third party who is obligated to keep the invention confidential can qualify as prior art under §102(a).”

The question decided was the on-sale bar. The Court relied on pre-AIA Federal Circuit precedent, including Woodland Trust v. Flowertree Nursery, 148 F.3d 1368, 1370 (Fed. Cir. 1998): “an inventor’s own prior commercial use, albeit kept secret, may constitute a public use or sale under §102(b).” (Part II-B.)

What if someone else patents a model already in secret use?

The prior user may have a defense under 35 U.S.C. § 273. It covers “a process, or ... a machine, manufacture, or composition of matter used in a manufacturing or other commercial process” that the person, “acting in good faith, commercially used” in the United States at least one year before the earlier of the patent’s effective filing date or a public disclosure qualifying under § 102(b) (§ 273(a)), proved by clear and convincing evidence (§ 273(b)).

The defense is personal, passing only to the patent owner or with the entire enterprise or line of business to which it relates (§ 273(e)(1)); it is unavailable for subject matter “derived from the patentee” (§ 273(e)(2)); it “is not a general license” (§ 273(e)(3)); and a university exception applies (§ 273(e)(5)). Raising it does not alone make the patent invalid under § 102 or § 103 (§ 273(g)).

Which factors bear on the choice for an AI system?

The statutes and decisions point to factors practitioners may weigh component by component, without settling the choice for any system. Secrecy: trade-secret status depends on reasonable measures and on the component not being readily ascertainable by proper means (§ 1839(3)). Disclosure: a patent must enable the full scope of its claims (Amgen), and claims that do no more than apply generic machine learning to new data environments, without disclosing improvements to the models, are ineligible (Recentive).

Reach and duration: a patent runs 20 years from filing and operates “against the world” (Kewanee); a trade secret has no fixed term but ends on public disclosure (Ruckelshaus). Timing: publication at 18 months unless a nonpublication request applies (§ 122(b)), confidential sales as prior art (Helsinn) and the limited § 273 defense.