The Missing Boundary in the Revised TTBER: AI Training Data and the Concept of Know-How

The adoption of the Technology Transfer Block Exemption Regulation (Commission Regulation (EU) 2026/877) and the accompanying Technology Transfer Guidelines mark the first comprehensive revision of the EU technology transfer regime in more than a decade. Most early commentary has concentrated on the revised treatment of licensing restraints, market definition and technology pools. Comparatively little attention has been paid to what may ultimately prove one of the reform’s most significant innovations: the introduction of a dedicated section on data licensing. That innovation deserves closer examination. Its significance does not lie in establishing a comprehensive legal regime for data licensing. Rather, it exposes an unresolved conceptual issue within the architecture of the TTBER itself.

Until now, TTBER has operated based on a relatively stable taxonomy of technology rights. Its scope extends to patents, software copyright, design rights, know-how and other recognized technology rights. The revised Guidelines preserve that structure while acknowledging commercially valuable data has increasingly become the object of licensing agreements. At the same time, however, they stop short of explaining where proprietary datasets fit within that taxonomy. Instead, the Guidelines distinguish between know-how, databases protected under Directive 96/9/EC, and "other forms of data" that may require an individual assessment under Article 101 TFEU.

This distinction raises a question that extends well beyond AI. Where, precisely, is the boundary between know-how and commercially valuable data? The revised Guidelines acknowledge that such a boundary exists. They provide little guidance, however, as to how it should be drawn.

The issue has become particularly significant because competitive advantage in AI markets increasingly derives from proprietary training datasets rather than from patents alone. Foundation models undoubtedly depend upon sophisticated algorithms. Yet commercial advantage increasingly depends on exclusive access to carefully curated training datasets. In many instances, those datasets constitute the undertaking’s most valuable commercial asset. They are licensed, protected through contractual arrangements and capable of generating substantial competitive advantages. Whether they should be analyzed as know-how is therefore no longer a purely theoretical question. It determines the legal framework through which AI licensing agreements are assessed.

The answer cannot be found simply by asking whether AI training datasets resemble the traditional examples of know-how. Such an approach assumes that the legal concept is defined by reference to technologies. TTBER adopts a different methodology.

Article 1(1)(i) TTBER defines know-how as “a package of practical information resulting from experience and testing” that is secret, substantial and identified. The accompanying Guidelines explain each requirement in turn. The information must not be generally known or readily accessible. It must be significant and useful to produce the contract products. Finally, it must be described with sufficient precision to verify that the requirements of secrecy and substantiality are met.

The structure of this definition deserves emphasis. Article 1(1)(i) does not define know-how by reference to industrial manufacturing, engineering processes or any other specific form of technology. Nor does it require that the information be protected by intellectual property rights. Instead, it defines know-how exclusively through the legal characteristics of the information itself. The concept is therefore functional rather than technological. Whether the licensed asset consists of manufacturing instructions, laboratory procedures or structured datasets is not, in itself, legally decisive. The decisive question is whether the information possesses the characteristics attributed to know-how by the Regulation.

This feature of the definition has received surprisingly little attention. Historically, the issue rarely arose because technology transfer agreements principally concerned patents, industrial processes and engineering expertise. AI changes the factual context in which the TTBER operates. It does not necessarily change the legal concept.

The revised Guidelines themselves support that conclusion. Paragraph 63 confirms that the TTBER applies where licensed data itself constitutes know-how within the meaning of Article 1(1)(i) or another recognized technology right. Paragraphs 64–67 apply the TTBER's analytical principles to protected databases while recognizing that other forms of data may require assessment under Article 101 TFEU. The Guidelines therefore acknowledge that some forms of data already fall within the concept of know-how. They do not explain, however, how that conclusion should be reached.

That omission represents the principal doctrinal issue raised by the revised framework. The Guidelines recognize that commercially valuable data cannot be treated as a homogeneous category. Some data constitutes know-how. Some benefits from database protection. Some fall into neither category. Yet the analytical boundary between those categories remains undefined. Proprietary AI training datasets expose that uncertainty more clearly than any other contemporary technology.

The question, therefore, is not whether AI training datasets represent a new form of commercial asset. They plainly do. The more interesting legal question is whether the existing concept of know-how is already capable of accommodating certain proprietary datasets without requiring any expansion of the TTBER itself. That inquiry requires closer consideration of the three cumulative elements contained in Article 1(1)(i). Those elements should not be understood as a mechanical checklist. Rather, they identify the legal characteristics that transform commercially valuable information into know-how for the purposes of EU competition law. Once the inquiry is framed in those terms, proprietary AI training datasets appear much closer to the traditional concept of know-how.

The three cumulative requirements contained in Article 1(1)(i) provide the analytical framework for answering that question. They should not, however, be applied mechanically. Their purpose is not to determine whether a particular asset is technologically innovative. Rather, they distinguish commercially valuable information that falls within the legal concept of know-how from information that remains outside the TTBER. The relevant inquiry is therefore functional. The question is not whether the licensed asset consists of “data”, but whether the information embodied in that dataset possesses the legal characteristics required by the Regulation.

The first requirement is secrecy. Article 1(1)(i) requires that know-how be “not generally known or easily accessible.” The emphasis is therefore placed on the commercial accessibility of the information rather than on the secrecy of every individual element from which it is composed. This distinction is particularly important in AI markets. Proprietary training datasets frequently incorporate information originating from publicly available or licensed sources. Their value lies in the effort invested in selecting, cleaning and validating the data. Competitors may have access to similar raw material. They are often unable, however, to reproduce the commercially valuable package of information resulting from those cumulative processes. What remains confidential is therefore not necessarily the underlying data, but the practical knowledge embodied in its organization.

The second requirement-substantiality-points in the same direction. The Guidelines explain that know-how must be significant and useful to produce the contract products or otherwise capable of improving the competitive position of the licensee. That definition is deliberately functional. It asks whether the information contributes materially to production rather than whether it belongs to a particular technological category. Historically, substantiality was most frequently associated with manufacturing techniques because those represented the predominant forms of technology transfer. Nothing in the wording of Article 1(1)(i), however, confines the concept to industrial production.

Viewed from that perspective, proprietary AI training datasets frequently perform the same economic function as traditional know-how. They enable the development of more accurate, reliable and commercially valuable AI systems by providing structured information that competitors cannot readily recreate through independent effort. The competitive advantage does not arise from legal exclusivity comparable to a patent. It arises from accumulated commercial experience embodied in the dataset itself. That is precisely the rationale underlying the protection of know-how within technology transfer law.

The third requirement, identification, reinforces that conclusion. Article 1(1)(i) requires that the information be described sufficiently to permit verification that it satisfies the requirements of secrecy and substantiality. The Guidelines further clarify that know-how need not always be embodied in exhaustive written documentation. Practical knowledge transmitted through training may also satisfy the requirement, provided that the information can be identified with sufficient certainty. Modern AI development practices appear particularly compatible with this criterion. Dataset governance increasingly relies upon detailed documentation concerning provenance, annotation methodologies, version histories, quality assurance procedures and permitted uses. Such materials frequently identify commercially valuable information with greater precision than many traditional forms of industrial know-how.

None of these observations suggest that every proprietary dataset should automatically qualify as know-how. Such a conclusion would be inconsistent with both the TTBER and the Guidelines. Publicly available datasets, information that can readily be reconstructed by competitors or data lacking independent commercial significance are unlikely to satisfy the cumulative requirements of Article 1(1)(i). The concept remains a narrow one. Its boundaries are defined by secrecy, substantiality and identification.

The more limited proposition advanced here is that the existing legal definition is capable of accommodating certain proprietary AI training datasets without requiring any modification of the TTBER itself. The decisive issue is therefore not whether AI has created a new category of commercial assets. It is whether the information embodied in the dataset performs the legal and economic function that Article 1(1)(i) associates with know-how.

The practical implications of that conclusion are significant. Classification determines the analytical framework through which licensing agreements are assessed. Where a proprietary dataset constitutes know-how, the established principles of technology transfer law become directly relevant. Questions concerning exclusivity, territorial restraints and grant-back obligations can then be analyzed under established TTBER principles. By contrast, where a dataset falls outside the concept of know-how, the agreement must be assessed directly under Article 101 TFEU. The distinction therefore affects the methodology of legal analysis rather than merely its terminology.

More broadly, the revised Guidelines illustrate a gradual evolution in EU competition law. Technology transfer has traditionally been associated with patents, industrial processes and other recognized technology rights. AI demonstrates that competitive advantage increasingly derives from proprietary information rather than formal intellectual property rights. The Commission has acknowledged that development by introducing dedicated guidance on data licensing. It has not, however, redefined the concept of know-how. Instead, it has preserved a definition that is sufficiently technology-neutral to accommodate new forms of commercially valuable information where the statutory requirements are satisfied.

The principal contribution of the revised TTBER may therefore lie not in the rules that it introduces but in the conceptual question that it leaves open. The Guidelines acknowledge that data may constitute know-how while distinguishing it from protected databases and other forms of commercially valuable information. They do not identify the criteria by which proprietary datasets should be allocated between those categories. AI training datasets expose that unresolved boundary with particular clarity. They are unlikely to be the last category of information to do so.

Whether future enforcement practice ultimately classifies AI datasets as know-how will necessarily depend upon the facts of individual cases. The broader doctrinal point is, however, already apparent. The revised TTBER does not require the creation of a new category of technology right to accommodate developments in AI. Properly interpreted, Article 1(1)(i) already contains a functional concept of know-how capable of evolving alongside technological change. The future development of EU technology transfer law may therefore depend less upon the invention of new legal concepts than upon a renewed appreciation of the breadth of one of its oldest ones.