Article 2 excludes AI systems and models specifically developed and put into service for the sole purpose of scientific research and development, and it treats research, testing and development activity prior to placing on the market or putting into service separately from the duties that follow market placement. The exclusion is tied to sole purpose and to timing. Once a system or model is placed on the market or put into service for another purpose, or a research prototype is developed into a supplied product, the exclusion no longer describes the activity and the full analysis must be done.
What does the scientific research exclusion actually say?
The Act states that it does not apply to AI systems or models specifically developed and put into service for the sole purpose of scientific research and development. The words specifically and sole purpose do heavy work. A system built for research, documented as research, used only for research, and not supplied for a commercial or operational purpose fits the wording more clearly than a system built for research but also used to serve customers. The exclusion is about the purpose for which the system was specifically developed and put into service, not about the background of the people who built it. A research team inside a commercial company does not automatically bring every prototype within the exclusion.
How is pre-market research, testing and development treated?
Research, testing and development activity that takes place before a system is placed on the market or put into service is distinguished from the regulated market act. Testing in a laboratory, evaluating a prototype, and running experiments that do not supply the system to a deployer or user are not the same as placing the system on the market. The boundary is crossed when the system is first made available or supplied for first use for its intended purpose. Internal testing that involves real users, real affected people or real operational decisions needs careful handling, because the facts may show that the system has in substance been put into service even if the team still calls the work a trial.
For the broader regulatory framework, see EU AI Act fundamentals.
What about regulatory sandboxes and real-world testing?
The Act provides separate regimes for AI regulatory sandboxes and for testing in real-world conditions, with their own conditions, supervision and safeguards. Those regimes are not the same as the scope exclusion. A sandbox provides a controlled environment for developing and testing innovative systems under supervision, and real-world testing has conditions including informed consent and oversight in the situations the Act specifies. A team should not treat participation in a sandbox, or a decision to call work a pilot, as proof that the scope exclusion applies. The applicable regime must be identified by its own article, and its conditions must be met and recorded.
What happens when research becomes a product?
The transition from research to product is the point where most scope errors occur. A prototype that is productised, integrated into a customer-facing system, or supplied under a contract has changed its purpose and its market status. At that point the threshold question, the role analysis, the Article 5 screen and the Article 6 classification must all be done on the product facts. Earlier research documentation can help, because it may show what was tested and what was known, but it does not substitute for classification, documentation and conformity work for the supplied system. The record should show the date and decision that marked the transition, and who approved it.
What other exclusions are often confused with research?
Free and open-source AI systems and models benefit from a distinct, limited treatment in the Act, and purely personal non-professional use is outside scope, but neither is the research exclusion. An open-source release can still engage Article 5 or Article 50 or high-risk duties in the situations the Act specifies, and a personal user is not the same as a research organisation. Military, defence and national security exclusions are also separate and narrow. Each exclusion should be tested on its own wording, and a system should not be described as excluded in general terms when only one narrow limb has been considered. If two exclusions are relied on, both should be recorded separately.
Continue with scope and territorial reach.
What can this guide not confirm?
This guide outlines the research and development exclusions. It does not confirm that any particular project, prototype, sandbox activity or open-source release is excluded. That confirmation needs the development purpose, the funding and governance arrangements, the users and affected people, the point of market placement, and the conditions of any testing regime relied on. Where a prototype touches real people or real decisions, or where the line between research and supply is unclear, the position should be confirmed with a competent authority or legal adviser before the system is used beyond the research setting.
Frequently asked questions
Does labelling work as research exclude it from the AI Act?
No. The exclusion turns on whether the system was specifically developed and put into service for the sole purpose of scientific research and development, and on the timing of market placement, not on the label a team gives the work.
Can a prototype be tested with real users under the research exclusion?
Testing with real users or real decisions may amount to putting the system into service or to real-world testing, each with its own conditions. The facts of the testing, not the word prototype, decide which regime applies.
Does the research exclusion cover a model released openly?
Open release is governed by separate provisions on free and open-source systems and models, with limited exceptions that do not cover every duty. The research exclusion and the open-source treatment should not be conflated.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689 ↗AI Act Service Desk timeline ↗European Commission regulatory framework for AI ↗