EU Tech Transfer Rules Leave AI Training Data in Legal Limbo

The European Union's first major overhaul of technology transfer rules in over a decade has introduced a significant ambiguity that directly affects how AI licensing agreements will be assessed under competition law.

The revised Technology Transfer Block Exemption Regulation (TTBER), adopted as Commission Regulation (EU) 2026/877, includes new guidance on data licensing—a recognition that commercially valuable datasets have become central to technology transfer agreements. Yet according to legal analysis first reported by Konstantin Voropaev of Gornitzky GNY on the Kluwer Competition Law Blog, the framework stops short of answering a fundamental question: where do proprietary AI training datasets fit within the existing legal taxonomy?

The know-how classification problem

The TTBER has historically covered patents, software copyright, design rights, and know-how. The revised Guidelines acknowledge that data licensing matters but distinguish between know-how, databases protected under EU Directive 96/9/EC, and "other forms of data" requiring individual assessment under Article 101 TFEU. What remains unclear is the boundary between these categories.

This matters because classification determines the analytical framework. If a proprietary dataset qualifies as know-how, established TTBER principles apply to questions of exclusivity, territorial restraints, and grant-back obligations. If it falls outside that definition, the agreement faces direct assessment under Article 101—a more uncertain path.

Why it matters

Competitive advantage in AI markets increasingly derives from curated training datasets rather than algorithms alone. Foundation model developers often guard their training data as their most valuable commercial asset, licensing it under strict contractual terms. Whether these datasets receive know-how treatment under EU competition law will shape how licensing agreements are structured and enforced across the AI industry. The uncertainty creates compliance risk for companies negotiating data access deals and may influence investment decisions in proprietary dataset development.

A functional definition under pressure

Article 1(1)(i) of the TTBER defines know-how as "a package of practical information resulting from experience and testing" that is secret, substantial, and identified. Notably, this definition is technology-neutral—it specifies legal characteristics rather than technological categories.

Voropaev argues that proprietary AI training datasets may already satisfy these criteria without requiring any expansion of the regulation. The secrecy requirement focuses on commercial accessibility, not the confidentiality of every data point. A dataset compiled from public sources can still qualify if competitors cannot readily reproduce the selection, cleaning, and validation work that makes it valuable.

The substantiality test asks whether information materially contributes to production or improves competitive position—a description that fits high-quality training datasets enabling more accurate AI systems. The identification requirement, demanding sufficient description to verify secrecy and substantiality, aligns well with modern dataset governance practices that document provenance, annotation methods, and quality assurance.

Enforcement will clarify boundaries

The revised Guidelines acknowledge at paragraph 63 that licensed data may constitute know-how under Article 1(1)(i), but they provide no methodology for making that determination. Not every proprietary dataset will qualify—publicly available information, easily reconstructed data, and commercially insignificant collections will fall short.

Future enforcement cases will necessarily draw these boundaries on a fact-specific basis. The broader implication, according to the analysis, is that EU technology transfer law may not need new legal categories to accommodate AI. Instead, the existing functional concept of know-how may prove flexible enough to evolve alongside technological change.

The details of this analysis were first reported by Konstantin Voropaev on the Kluwer Competition Law Blog.