Research exclusions in the AI Act – a fit(ness) check

A recurring feature of EU legislation on information technologies is the special treatment afforded to research. Legislators frequently carve out exceptions or introduce tailored rules that protect research activities, recognising their broad societal benefits and the comparatively limited risk associated with exploratory work. Examples of such protection include: the General Data Protection Regulation 2016/679 (GDPR), facilitating compliance with requirements tied to the purpose of processing and prolonged storage for scientific research; the Digital Services Act 2022/2065 (DSA), enabling researchers’ access to big online platforms’ and search engines’ data; and the Copyright in Digital Single Market Directive 2019/790 (CDSMD), providing a copyright exception for text and data mining conducted for the purpose of scientific research.

The Artificial Intelligence Act Regulation 2024/1689 (AI Act) - EU’s flagship legislative instrument for AI compliance - follows this trend, primarily through two exclusions placing AI systems and models tied to scientific research (Art. 2(6)) and applied research (Art. 2(8)) outside of the scope of this Regulation. This blog post presents both exclusions, explores the interpretative challenges they carry, and identifies their limitations.

AI systems and models

Before delving into the research-related exclusions themselves, it is useful to briefly recall the scope of the AI Act and the definitions that trigger its applicability. Not every computational model, software tool, or experimental system will necessarily qualify as an “AI system” or “AI model” under the Regulation.

The AI Act defines an AI system in Art. 3(1) as:

“a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.”

Meanwhile, Art. 3(63) defines a general-purpose AI model as:

“an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications, except AI models that are used for research, development or prototyping activities before they are placed on the market.”

The final part of this definition is particularly noteworthy from a research perspective. The exclusion of AI models used for research, development, or prototyping before they are placed on the market appears to overlap with the broader scope exclusions contained in Art. 2. Further discussion on this point follows later in this blogpost.

The definitions of AI system and general-purpose AI model are intentionally broad and technologically neutral. However, they also create interpretative difficulties, particularly in research settings where systems are experimental, modular, continuously evolving, or not yet intended for deployment.

This issue is especially relevant because the AI Act adopts a lifecycle-based approach to regulation. Yet, despite repeated references to the “lifecycle” or “lifetime” of AI systems (e.g. Art. 9), it does not  expressly clarify whether early research and exploratory development activities already fall within that lifecycle or, if not, at what point an AI system enters the regulated lifecycle for the purposes of the AI Act. This creates uncertainty as to where the boundary lies between excluded research activities and AI systems that are subject to the Regulation's substantive obligations. Guidance can be found in the OECD's conception of AI systems, which the AI Act largely follows. Unlike the AI Act, the OECD explicitly recognises planning and design as part of the AI lifecycle. This raises an immediate question for researchers: At what point does an experimental research artefact become an “AI system” or “AI model” subject to the AI Act?

The breadth of the definitions also explains why the research exclusions in Arts. 2(6) and 2(8) matter so significantly. Many AI research activities – including exploratory machine learning development, benchmark testing, synthetic data generation, or experimental model validation – may already fall within the material scope of the Regulation before any commercial deployment occurs.

Against this background, understanding the research exclusions becomes essential for a series of stakeholders, such as universities, hospitals, corporate laboratories, public authorities, and collaborative innovation partnerships seeking to determine when experimental AI activities remain outside the scope of the Regulation, and when they no longer do.

The scientific research exclusion – Art. 2(6)

The first of the two exclusions is phrased in the following manner:

“This Regulation does not apply to AI systems or AI models, including their output, specifically developed and put into service for the sole purpose of scientific research and development.”

These components trigger the following interpretative and/or practical challenges:

Inclusion of outputs

The exclusion explicitly covers the outputs of AI systems and models. This makes sense because outputs can have real-world effects independently of, or alongside, the operation of the underlying AI system or model. The AI Act itself provides guidance on what constitutes an "output". In the definition of an AI system, it gives a non-exhaustive list of examples, including "content, predictions, recommendations, or decisions that can influence physical or virtual environments". By extending the exclusion to outputs, the Act recognises that research often involves generating and analysing such outputs before an AI system is deployed or marketed. For outputs of AI models, the definition does not provide a similar list, referring instead to “performing a wide range of distinct tasks”. However, AI models are essentially mathematical models. For example, an AI model in the insurance sector may estimate a probability of default to be 53%, while an AI system would take this data onwards, to create a risk profile of the applicant or suggest whether a credit application should be successful. In most cases, an AI model involved in creation of outputs will mostly do it through the shell of an AI system.

An emerging question is whether the outputs of an AI system falling within Art. 2(6) have to be used for the sole purpose of scientific R&D to keep the AI system itself within the frame of this provision. In other words, if an AI system is developed exclusively for scientific research, does the research exemption cease to apply if its outputs are later used for a different purpose? The answer seems to be yes – otherwise, this provision could be used to bypass too much of AI Act’s obligations.

Development and putting into service

The AI Act does not define development in a dedicated provision; hence the general language definition of “creating or producing, especially by deliberate effort over time” could be endorsed. By contrast, the Act expressly defines putting into service as “supply of an AI system for first use directly to the deployer or for own use in the Union for its intended purpose” [Art. 3(11)].

Scientific research

Art. 2(6) applies only where the sole purpose is scientific research. However, the AI Act does not define this concept or refer to another EU instrument that does. To interpret it, it is therefore helpful to look at other areas of EU law that use the same term.

The first such point of reference lies within the GDPR. Recital 159 of this Regulation states that  “(f)or the purposes of this Regulation, the processing of personal data for scientific research purposes should be interpreted in a broad manner including for example technological development and demonstration, fundamental research, applied research and privately funded research”.

Recently published EDPB Guidelines 1/2026 on processing of personal data for scientific research purposes (covered in depth in another Timelex blog post) confirm that although there is no universally agreed definition of scientific research, there are six “common elements which can serve as key-indicative factors” towards the identification of scientific research:

§  1. Methodical and systematic approach – Including the formulation of a hypothesis, objective and a research plan;

§  2. Adherence to ethical standards – Focused here on protection of individuals from harm;

§  3. Verifiability and transparency – Based on verifiable results, peer review, sharing with other parties, legitimate limitations;

§  4. Autonomy and independence - In relation to differently sourced prejudices; including freedom to disseminate; presence of qualifications; transparency on assumptions and values; robustness of scientific evidence (these last two points overlap with no. 3);

§  5. Objectives of the research – Goal of the research should be to foster the growth of society’s general knowledge and well-being; but it may also seek to further commercial interests;

§  6. Potential to contribute to existing scientific knowledge or apply existing knowledge in novel ways – Including evaluation of this contribution by e.g., independent experts or committees.

EDPB explains that if all six factors can be confirmed, “they can be presumed to constitute scientific research”. If some are missing, the controller “needs to justify and be able to demonstrate why the activities should nonetheless be considered scientific research” (the more factors are present, the more convincing the overall justification).

The earlier mentioned Copyright in Digital Single Market Directive (CDSMD) also does not have a stand-alone definition of scientific research. Instead, in recital 12, it states that this term should cover “both the natural sciences and the human sciences”. A more operative definition can be found in the concept of a research organisation. Art. 2(1) defines such an entity as “a university, including its libraries, a research institute or any other entity, the primary goal of which is to conduct scientific research or to carry out educational activities involving also the conduct of scientific research.” This has to occur either “on a not-for-profit basis or by reinvesting all the profits in its scientific research” – or “pursuant to a public interest mission recognised by a Member State”. In any case, “the access to the results generated by such scientific research cannot be enjoyed on a preferential basis by an undertaking that exercises a decisive influence upon such organisation;”

For the DSA, it does not contain a firm definition of scientific research either. Instead, it sets out who can become a vetted researcher with access to online platforms’ and search engines’ data . Among other factors, they have to be affiliated to a research organisation; independent from commercial interest; with transparent funding; and a commitment to making their research results publicly available, free of charge. Interestingly enough, “research organisation” is expressly linked to the meaning of this term given in CDSMD.

Another interesting point of sectoral reference lies in the European Health Data Space Regulation 2025/327 (EHDS). In setting out the purposes for which electronic health data can be processed, Art. 53(e) refers to scientific research purposes. These are specified as “scientific research related to health or care sectors that contributes to public health or health technology assessments or ensures high levels of quality and safety of healthcare, of medicinal products or of medical devices, with the aim of benefiting end-users, such as patients, health professionals and health administrators”. Two areas of activity are indicated as examples; “development and innovation activities for products or services”; and “training, testing and evaluation of algorithms, including in medical devices, in vitro diagnostic medical devices, AI systems and digital health applications”.

Overall, the guidance provided by EDPB in 2026 in the context of GDPR seems to be (currently) the most useful for AI Act’s needs. Just as the GDPR regulates the lifecycle of processing of personal data, the AI Act regulates the development and putting into service of AI systems and models, in a primarily sector-agnostic manner. The EDPB guidelines explain what counts as a scientific research purpose and provide a practical framework for assessing whether an activity falls within that concept, even if the framework remains broad.  At the same time, further guidance could be expected, as Art. 2(6) and the notion of scientific research purposes are likely to be tested by those wishing to obtain the pay-off of being excluded from the AI Act’s regime. This is particularly likely to take place shortly before the (revised) date of entry into force for stand-along high-risk AI systems – at the time of writing, 2 December 2027.

Interplay with prohibited AI practices

A final, though by no means least important question is whether the AI Act's scientific research exclusion also applies to prohibited AI practices under Article 5 (such as e.g., forms of predictive policing, vulnerability-based behaviour manipulation or social scoring). This provision prohibits certain AI practices by making it unlawful to place the relevant AI systems on the market, put them into service, or use them. At the same time, Article 2(6) excludes from the scope of the AI Act AI systems and models developed and put into service solely for scientific research and development. This creates an interpretative question. Firstly, because Article 2(6) excludes these activities from the scope of the AI Act, it also permits the development of AI systems that would otherwise fall within Article 5, provided they are developed solely for scientific R&D. It is unclear though whether the same applies to putting such systems into service, even if only for a purely scientific R&D purpose. Article 5 simply states that the relevant "AI practices shall be prohibited", without expressly limiting that prohibition to AI systems otherwise falling within the scope of the Regulation. This raises the question whether the research exclusion prevails, or whether Article 5 establishes a standalone prohibition that also applies in the research context. Official clarification of this dilemma would be most welcome.

The development phase exclusion – Art. 2(8)

Alongside the narrower scientific research exclusion in Art. 2(6), the AI Act introduces a broader development-phase exclusion in Art. 2(8), which states:

"This Regulation does not apply to any research, testing or development activity regarding AI systems or AI models prior to their being placed on the market or put into service. Such activities shall be conducted in accordance with applicable Union law. Testing in real world conditions shall not be covered by that exclusion." 

The provision therefore appears designed primarily to protect pre-commercial experimentation and innovation processes. Recital 25 confirms this logic by explicitly referring to “product-oriented research, testing and development activity regarding AI systems or models”. In practice, this exclusion is particularly relevant for commercial AI providers, start-ups, research partnerships, and innovation ecosystems where AI systems are still under development and have not yet crossed into deployment or commercialization.

Importantly, the exclusion is temporal rather than purposive. Contrary to Art. 2(6), Art. 2(8) does not require the activity to pursue exclusively scientific objectives. A project may therefore still fall within the exclusion even where future commercialization is anticipated.

This temporal focus also sheds light on the exclusion contained in the definition of a general-purpose AI model under Art. 3(63), which excludes AI models used for research, development or prototyping activities before they are placed on the market. Since Art. 2(8) already excludes research, testing, development and prototyping activities concerning AI systems and models prior to their placement on the market or putting into service, the carve-out in the definition of a general-purpose AI model is better understood not as an autonomous exclusion but as a clarification that experimental models should not trigger the specific regime applicable to general-purpose AI models while they remain in the development phase. Art. 2(6), which excludes AI systems and models developed and put into service solely for scientific research and development, points in a similar direction, although it focuses on the purpose of the activity rather than the developmental stage of the system or model. Today, many research initiatives, such as Horizon Europe projects, operate precisely at the intersection of scientific research and market-oriented innovation. Research activities are frequently conducted with the expectation that successful outputs may eventually evolve into marketable products. Usually, partners in research consortia are expected to describe in their project’s exploitation plan specific actions for market uptake of research results and this is directly linked to project’s funding approval. Therefore, the EU itself embraces a mindset of commercial development incorporated in research activities.

Given the above, Art. 2(8) may prove particularly important for research consortia or large corporate research units, whose activities often resemble academic scientific research while simultaneously remaining connected to future commercial deployment.

Limitations of the exclusion from Art. 2(8)

Unlike Art. 2(6), the exclusion ends once the AI system or model is either made available on the market or put into service.

The AI Act defines making available on the market as supplying an AI system or AI model for distribution or use on the Union market in the course of a commercial activity, whether for payment or free of charge (Art. 3(10)). Putting into service, by contrast, means supplying an AI system for its first use by a deployer, or for the provider's own use in the Union, for its intended purpose (Art. 3(11)). 

At first glance, these concepts appear familiar from EU product safety law. Yet their application to AI research environments is far from straightforward.

For instance, it is unclear whether releasing an AI model or related software under an open-source licence amounts to making it available on the market. The European Commission’s AI Act FAQ suggests that an open-source release can indeed amount to placing on the market because commercial supply may occur free of charge. This creates significant uncertainty for AI researchers, particularly in academic and open research ecosystems, where publication of models, codes, and datasets is considered an integral part of scientific methodology.

The uncertainty is further reinforced because the AI Act does not define the concept of "commercial activity", even though this forms part of the definition of making available on the market under Art. 3(10). Its meaning must therefore be derived from the broader framework of Union product legislation. The Blue Guide – which informs the interpretation of market-related concepts under Union product harmonisation legislation – explains that commercial character of an activity must be assessed contextually, taking into account factors such as the regularity of supply, the characteristics of the provider, and the intentions behind dissemination. Even non-profit organisations may, in certain circumstances, operate within a commercial context.

This becomes especially relevant for collaborative AI research environments involving multiple actors who work together within shared research infrastructures but with potentially different motivations. In such settings, the distinction between scientific dissemination and commercial availability becomes increasingly difficult to maintain.

Real-world testing

While the previous limitation concerns an AI system or model being made available on the market – a concept that indirectly raises questions regarding "commercial activity" through Art. 3(10) – Art. 2(8) also establishes a distinct limitation. Independently of whether an AI system has been made available on the market or put into service, the final sentence of Art. 2(8) provides that "testing in real world conditions shall not be covered by that exclusion”. This carve-out demonstrates the legislature’s concern that developers might otherwise circumvent the AI Act by characterising deployment activities in the real world as mere “research”.

The Regulation defines “testing in real world conditions” as temporary testing of an AI system for its intended purpose outside a laboratory or otherwise simulated environment, with a view to gathering reliable and robust data and to assessing and verifying the conformity of the AI system with the requirements of the AI Act [Art. 3(57)]. The AI Act also establishes detailed procedural safeguards for such testing, including approval procedures, testing plans, oversight obligations, and informed consent requirements.

However, the Regulation provides limited guidance on what constitutes “real-world conditions” in practice. This creates particular difficulties in sectors such as healthcare, where AI validation frequently requires interaction with real data, operational systems and live environments before full deployment.

Development, validation and deployment in modern AI ecosystems – A use case from the health domain

Consider, for example, researchers evaluating an AI triage system in a live hospital emergency department. The system analyses real patient data and generates prioritisation recommendations. However, clinicians continue making triage decisions independently and without relying on the AI system’s outputs. The deployment is intended to assess the system’s accuracy, robustness, and clinical performance under operational conditions before full implementation.

Would such deployment remain protected under the development phase exclusion (Art. 2(8))? The answer likely depends on the intended purpose of the system. If an AI system designed to support emergency triage or clinical prioritisation is tested through operation within a functioning hospital environment – even where its outputs are not relied upon in practice – this may still amount to testing the system for its intended purpose outside a laboratory or simulated environment. In that case, the activity could qualify as “testing in real-world conditions” and therefore fall outside the Art. 2(8) exclusion. If that interpretation prevails, the activity would trigger the AI Act’s specific framework for real-world testing, including prior approval requirements, real-world testing plans, oversight obligations, informed consent safeguards and time limitations.

By contrast, if the system relies exclusively on historical emergency department records or synthetic patient data within a secure laboratory environment, allowing clinicians to retrospectively compare the system’s recommendations against past triage decisions, the activity would be more likely to remain within the Art. 2(8) exclusion. This is because the AI system is not connected to the hospital’s live intake infrastructure and does not process real-time patient information during ongoing emergency care.

The example illustrates a broader challenge underlying Art. 2(8). The AI Act appears to treat research, development and deployment as distinct stages, whereas contemporary AI innovation is often iterative, collaborative and continuous. Universities, hospitals, industry partners and open-source communities frequently operate within shared innovation ecosystems where validation and deployment increasingly overlap. Consequently, the boundary between development activities covered by Art. 2(8) and real-world testing remains uncertain and will likely require further clarification through guidance, enforcement practice and judicial interpretation.

Until then, developers and researchers may face a difficult compliance dilemma: either adopt overly cautious interpretations that risk hampering innovation or rely on an uncertain exclusion that may later prove inapplicable.

Conclusion

The AI Act’s research exclusions reveal the broader tension at the heart of contemporary EU digital regulation. On one hand, the EU seeks to foster innovation, scientific freedom, and technological development. On the other hand, it simultaneously aims to protect health, safety, and fundamental rights against the risks associated with increasingly powerful AI systems. Arts. 2(6) and 2(8) represent a legislative attempt to reconcile these objectives. Yet, as the preceding analysis demonstrates, the practical operation of these exclusions remains far from straightforward. Art. 2(6) establishes a narrow exclusion for AI systems and models developed exclusively for scientific research purposes. Art. 2(8), meanwhile, creates a broader development-phase exclusion designed to protect research, testing, and innovation activities before commercialization or deployment. We summarize the differences in the table below.

However, both provisions rely on distinctions that increasingly struggle to reflect the realities of modern AI ecosystems. Contemporary AI research rarely occurs in isolation. As a result, several core concepts within the AI Act remain legally uncertain. This uncertainty has important practical consequences. For AI developers, research institutions, hospitals, and public authorities, understanding the limits of the research exclusions is essential for designing compliant innovation strategies from the earliest development stages. In practice, many organisations will need to anticipate AI Act obligations well before commercialization occurs.

At the same time, regulators and supervisory authorities will likely need to provide further clarification through guidance, standards, enforcement practice, and eventually judicial interpretation. Without such clarification, the AI Act’s research exclusions risk functioning less as a genuine innovation safe harbour and more as a legally unstable grey zone of uncertainty, discouraging the actors the Regulation seeks to encourage.

If your organisation is developing AI systems or conducting research activities with commercial potential, early assessment of AI Act obligations is becoming increasingly important. Timelex supports research actors and other key stakeholders in navigating the AI Act throughout the entire innovation lifecycle – from research and testing phases to deployment and commercialization. Our specialised team combines extensive expertise across AI regulation and the broader field of digital law to deliver practical, tailored, and proactive compliance support.

The authors would like to thank Magdalena Kogut-Czarkowska for her insightful feedback on this post.