AI model weights, architectures, training data curation methods, and hyperparameter configurations can qualify as trade secrets under the EU Trade Secrets Directive (2016/943) and the US Defend Trade Secrets Act (18 U.S.C. Sections 1836-1839), provided the holder demonstrates commercial value from secrecy, has taken reasonable protective measures, and the information is not generally known or readily ascertainable by persons in the relevant field.
Protecting AI Models as Trade Secrets: Legal Framework and Practical Measures
Legal Framework for AI Trade Secrets
Trade secret protection requires meeting three conditions in both EU and US law. Under the EU Trade Secrets Directive Article 2(1), information qualifies if it: (a) is secret (not generally known or readily accessible to persons who normally deal with that kind of information); (b) has commercial value because it is secret; and (c) has been subject to reasonable steps to keep it secret. The US Defend Trade Secrets Act defines trade secrets similarly (18 U.S.C. Section 1839(3)), adding that the information must derive independent economic value from not being generally known. For AI systems, protectable elements include trained model weights, proprietary training data compilations, data preprocessing pipelines, hyperparameter configurations, model architecture innovations, and evaluation methodologies.
What AI Elements Qualify as Trade Secrets
Not every aspect of an AI system merits trade secret protection. Publicly disclosed architectures (such as published research papers) cannot qualify. However, the specific implementation, trained weights resulting from proprietary data and compute investment, and the combination of individually non-secret elements into a proprietary system can qualify. Courts in the US have recognized that compilations of information, even from publicly available sources, can be trade secrets when the selection, coordination, or arrangement adds value (Compulife Software v. Newman, 11th Cir. 2020).
| AI Element | Trade Secret Potential | Key Consideration |
|---|---|---|
| Trained model weights | High | Represent significant compute and data investment; not derivable from architecture alone |
| Training data compilation | High | Curation, cleaning, and labeling methodology; not the individual data points if publicly available |
| Hyperparameter configurations | Medium-High | Specific combinations resulting from extensive experimentation |
| Model architecture (novel) | Medium | Lost upon publication or independent discovery; patent may be more appropriate |
| Evaluation benchmarks (internal) | Medium | Proprietary test sets and evaluation criteria |
| Preprocessing pipelines | Medium | Data augmentation, filtering, and normalization procedures |
Reasonable Protective Measures
Demonstrating "reasonable steps" is essential and often determinative in litigation. For AI systems, reasonable measures include: technical controls (encryption of model weights at rest and in transit, access control with role-based permissions, audit logging of model access, secure enclaves for inference); contractual measures (NDAs with employees, contractors, and partners; IP assignment agreements; confidentiality clauses in licensing agreements); organizational measures (need-to-know access policies, classified information handling procedures, employee training on trade secret obligations, exit procedures including device wiping and reminder of ongoing obligations); and model deployment safeguards (rate limiting API access, obfuscating model architecture in deployed services, monitoring for model extraction attacks).
Tension with AI Transparency Requirements
The EU AI Act creates tension with trade secret protection. Article 13 requires high-risk AI systems to be designed with sufficient transparency for deployers to understand the system's output. Article 53 requires GPAI providers to publish training data summaries. Article 78 gives national authorities access to AI system information for enforcement. However, Article 78(3) and Recital 137 explicitly protect trade secrets and confidential business information in enforcement proceedings. The practical balance: provide sufficient information for regulatory compliance and deployer understanding without disclosing the specific implementation details (weights, exact data composition, hyperparameter values) that constitute the trade secret.
Litigation and Enforcement
Trade secret misappropriation claims for AI systems face practical challenges: proving access (was the defendant able to examine the model?), proving misappropriation versus independent development (did the defendant reverse-engineer or independently train a similar model?), and proving damages (how to value a trade secret that has been disclosed?). The EU Directive provides remedies including injunctions, recall of infringing goods, damages, and publication of decisions (Articles 12-15). The US DTSA provides federal civil remedies and, for willful misappropriation, exemplary damages up to twice the actual damages (18 U.S.C. Section 1836(b)(3)(C)). Ex parte seizure orders are available in extraordinary circumstances.
Practical Protection Strategy
Organizations should: (1) conduct a trade secret audit identifying all AI-related proprietary information; (2) classify information by sensitivity level and implement corresponding access controls; (3) require all personnel with access to execute tailored NDAs specifying AI-related trade secrets; (4) implement technical anti-extraction measures for deployed models; (5) maintain evidence of protective measures for litigation readiness; (6) document the commercial value derived from secrecy; (7) prepare for the disclosure-protection balance required by the AI Act by identifying the minimum information necessary for regulatory compliance; and (8) consider complementary IP protection (patents for novel architectures, copyright for code) to create layered protection.
Check your AI compliance readiness — free.
Take the Readiness Check 3 minutes · 10 questions · no signup requiredThis article is for informational purposes only and does not constitute legal advice. Regulatory requirements change frequently — verify current rules with official sources. Built by Sawai Gyoseishoshi Office, Hiroshima, Japan.