WIPO's Guide to Protecting AI Systems: Balancing Patents and Trade Secrets

WIPO's guide to protecting AI systems supports a complementary role for patents and trade secrets.

World Intellectual Property Organization – WIPO's new guide for small businesses, Learning Machines, explains how an AI model "will be protected by a combination of copyright and patents law, and possibly by trade secrets."

Patent Eligibility Across Jurisdictions

Patents can offer significant value, the guide notes, "sometimes even as pending applications," but eligibility remains a constraint. The European Patent Office treats AI algorithms as abstract and unpatentable by themselves but may grant patents on computer-implemented inventions with technical character, such as a heart-monitor neural network identifying irregular heartbeats.

The U.S. standard is less certain, as the guide finds "no clear rule." Yet recent decisions separate applying a model from improving it. In Recentive Analytics v. Fox (2025), the Federal Circuit held ineligible claims applying generic machine learning to new data "without disclosing improvements to the machine learning models." The USPTO's precedential Ex parte Desjardins decision (2025) vacated an eligibility rejection of claims to a training method that improved the model itself by helping it retain performance on earlier tasks. That decision binds USPTO examiners and the PTAB but not courts, so a patent allowed at the USPTO can still face an eligibility challenge in litigation under Recentive. GoTV v. Netflix (2026), a non-AI case, illustrates the risk: the Federal Circuit reversed a $2.5 million verdict, holding that only claimed features, not unclaimed specification details, can lift a claim above an abstract idea.

The Role of Trade Secrets in AI

Trade secrets protect information kept out of patent applications, especially model internals outsiders cannot observe or derive. The guide's example is a model's "weights and parameters," which "could be protected as confidential information." Protection lasts only while the owner takes "reasonable steps to keep the information secret." Examples of such steps include technical measures and NDAs, which the guide calls "common and essential," particularly in contracts with employees and independent contractors. The Defend Trade Secrets Act treats reverse engineering as lawful, so an improvement a competitor could reverse engineer is a stronger patent candidate.

Takeaways for AI Companies

• Patent improvements a competitor could detect, and claim and describe how each improvement changes the model.
• Protect weights, datasets, and training methods with NDAs, access controls, and logs.

Whatever a U.S. patent application discloses becomes public at publication, generally 18 months after the earliest filing date, so the patent-or-secret choice for each improvement is best decided before filing.

Sources

1. WIPO, 𝘓𝘦𝘢𝘳𝘯𝘪𝘯𝘨 𝘔𝘢𝘤𝘩𝘪𝘯𝘦𝘴: 𝘈𝘯 𝘐𝘯𝘵𝘳𝘰𝘥𝘶𝘤𝘵𝘪𝘰𝘯 𝘵𝘰 𝘈𝘐 𝘢𝘯𝘥 𝘐𝘗 𝘧𝘰𝘳 𝘚𝘮𝘢𝘭𝘭 𝘢𝘯𝘥 𝘔𝘦𝘥𝘪𝘶𝘮-𝘴𝘪𝘻𝘦𝘥 𝘌𝘯𝘵𝘦𝘳𝘱𝘳𝘪𝘴𝘦𝘴 (2026), Developing and protecting AI systems: https://www.wipo.int/web-publications/learning-machines-an-introduction-to-ai-and-ip-for-small-and-medium-sized-enterprises/en/developing-and-protecting-ai-systems.html
2. 𝘙𝘦𝘤𝘦𝘯𝘵𝘪𝘷𝘦 𝘈𝘯𝘢𝘭𝘺𝘵𝘪𝘤𝘴, 𝘐𝘯𝘤. 𝘷. 𝘍𝘰𝘹 𝘊𝘰𝘳𝘱., 134 F.4th 1205 (Fed. Cir. 2025), cert. denied, No. 25-505 (U.S. Dec. 8, 2025)
3. 𝘌𝘹 𝘱𝘢𝘳𝘵𝘦 𝘋𝘦𝘴𝘫𝘢𝘳𝘥𝘪𝘯𝘴, Appeal No. 2024-000567 (PTAB Sept. 26, 2025) (Appeals Review Panel) (precedential Nov. 4, 2025): https://www.uspto.gov/sites/default/files/documents/memo-desjardins.pdf
4. EPO Guidelines for Examination, G-II, 3.3.1
5. 18 U.S.C. § 1839(3), (6); 35 U.S.C. § 122(b)
6. 𝘎𝘰𝘛𝘝 𝘚𝘵𝘳𝘦𝘢𝘮𝘪𝘯𝘨, 𝘓𝘓𝘊 𝘷. 𝘕𝘦𝘵𝘧𝘭𝘪𝘹, 𝘐𝘯𝘤., 166 F.4th 1053 (Fed. Cir. 2026)