Most conversations about the EU AI Act in learning and development start with a single question: do our people understand AI well enough to use it responsibly?
That is the AI literacy question, and it matters.
A second question deserves equal attention, and it is asked less often:
Is the AI inside our learning tools itself regulated?
For some features, the answer is yes. The AI Act names education and vocational training as an area where certain uses are classed as high-risk, and it prohibits others outright. For L&D teams that buy or build learning technology, this changes the checklist.
Two Different Questions About AI in Learning
It helps to separate two topics that are often blended together.
AI literacy is about people. Do employees and learning professionals understand what the AI systems they use can and cannot do? This is covered by Article 4 of the AI Act.
AI system regulation is about the tools. Which AI-powered features in your learning stack fall into a regulated category, and what follows from that?
Both matter, and they call for different responses. The first is mainly a training design challenge. The second is mainly a procurement, governance and risk question that L&D needs to raise with legal, IT and vendors.
Where Learning Appears in the AI Act
The AI Act sorts AI systems by risk. For learning professionals, three areas are worth knowing.
1. Prohibited practices: emotion inference at work and in education. Since 2 February 2025, the Act has prohibited AI systems that infer emotions of people in the workplace and in educational institutions, except where used for medical or safety reasons. For learning teams, this is a clear red line for features that claim to read learner engagement, frustration or attention from faces, voices or other biometric signals.
2. High-risk uses in education and vocational training (Annex III). The Act lists several education-related use cases as high-risk, including AI systems intended to:
- determine access, admission or assignment to education and vocational training;
- evaluate learning outcomes, including when those outcomes are used to steer the learning process;
- assess the appropriate level of education an individual will receive or be able to access;
- monitor and detect prohibited behavior of students during tests.
The last point is where AI-based proctoring comes in. The second is the one many L&D teams overlook: it can reach adaptive systems that evaluate outcomes and use the result to decide what a learner sees next.
3. High-risk uses in employment and worker management. Annex III also covers AI used for recruitment and selection, for decisions affecting work conditions or promotion, and for monitoring and evaluating performance and behavior. Training that feeds into such decisions, for example assessments used to rank staff or inform promotion, deserves a closer look.
An important caveat: classification depends on the specific use and the exact wording of the Act, not on a product category. An AI feature that recommends optional reading is a very different matter from one that scores a learner and gates their progression or career path. Legal review is essential.
What Is Usually Not the Concern
It is equally useful to say what these provisions are not aimed at.
Using generative AI to draft a first version of a learning text, suggest quiz questions or propose a course outline does not, by itself, make a system high-risk under these education provisions. The main requirements here are practical rather than regulatory: accuracy, copyright awareness, confidentiality and human review before anything is published.
That said, transparency rules under Article 50 can still matter. They include informing people when they are interacting with an AI system such as a chatbot, and marking certain synthetic content. Application dates and any transitional arrangements should be checked against the current text.
A Moving Timeline
Timing is one of the trickiest parts of this topic.
The AI Act entered into force in August 2024 and applies in stages. Prohibitions and AI literacy provisions started to apply in February 2025. Obligations for general-purpose AI models followed in August 2025. The main high-risk obligations were originally scheduled for 2 August 2026.
In late 2025, the European Commission proposed changes under its Digital Omnibus package, including postponing parts of the high-risk regime and adjusting some provisions, among them the framing of AI literacy. Because legislative proposals evolve, the dates and obligations that apply to your organization should be confirmed against the latest adopted text before you plan around them.
The practical lesson is stable even if the calendar moves: organizations that understand where AI sits in their learning stack are better placed whichever dates finally apply.
Practical Steps for L&D Teams
You do not need to become AI lawyers. You do need a clear view of what you are using.
- Inventory AI features across the learning stack. Include the LMS, learning experience platform, authoring tools, assessment and proctoring tools, and any chatbot or recommendation engine. Note what each feature does and what decisions it influences.
- Separate low-stakes from high-stakes uses. Content suggestions are not the same as automated scoring, progression gating or anything feeding into hiring, promotion or performance evaluation.
- Ask vendors direct questions. Does the feature infer emotions? Does it evaluate learners or monitor test behavior? What documentation, human-oversight options and data-handling information can they provide?
- Keep humans in the decision. Where AI informs an outcome that matters to a person, define who reviews it, who can override it, and how a person can challenge it.
- Check data protection alongside the AI Act. Learning data is personal data, and automated decision-making rules under the GDPR may apply in parallel.
- Document your reasoning. Record what you assessed, what you decided and why. This links back to the broader trend of regulators expecting training and tooling decisions to be justified, as seen with cybersecurity training under NIS2 and DORA.
- Train the people involved. Article 4 literacy matters most for the staff who configure, buy and rely on these tools.
Content Governance Still Matters
Even when a tool is not regulated as high-risk, the content it produces or delivers still needs control.
If AI helps generate or adapt learning material, organizations benefit from knowing which content was AI-assisted, who reviewed it, which source it was checked against and which version was approved and delivered. These are ordinary content-governance questions, and they become more pressing as AI-assisted authoring spreads.
A learning content management system can support that discipline, through workflows, version history and review steps. It cannot decide whether a tool is compliant or whether a statement is true, and it should not be presented as doing so. Responsibility for those judgments stays with people.
The Takeaway
For L&D, the AI Act is not only about teaching people what AI is. It also asks what AI is doing inside the learning experience itself.
The most useful first step is rarely a new course. It is a simple map: where is AI used in our learning tools, what does it decide or influence, and who is accountable for the outcome?
Organizations that can answer those questions clearly will find both regulation and good practice easier to meet, whatever the final timeline looks like.
Sources and Further Reading
European Union, Regulation (EU) 2024/1689 (Artificial Intelligence Act), in particular Article 4, Article 5, Article 50 and Annex III.
European Commission, AI Act pages and Digital Omnibus proposal (check the latest status).
Editorial note: references reflect the author’s knowledge through mid-2026 and have not been checked against live sources. The AI Act timeline, the Digital Omnibus changes and Annex III wording should be verified on EUR-Lex and the European Commission website before publication.