FAQ on the AI Regulation

1.1 Scientific Privilege under Article 2(6) of the AI Regulation

The AI Regulation provides for an exception for AI systems that are developed and operated exclusively for scientific research and development . Does this exception also apply if the university later decides to use the system for purposes other than research—for example, in teaching, administration, or counseling? At what point does the exception no longer apply?

Only at purer Research. The scientific privilege applies only if an AI system is developed and operated exclusively for scientific research and development.

A planned Practical Application includes the Exception from. If future use in teaching, administration, consulting, or another practical field is already planned—or even if it is merely not ruled out—the exception does not apply. This applies even if the system is currently used exclusively in a research project.

None later Omission, but Exclusion from the very beginning. In this case, therefore, the privilege does not end at a later date. According to the expert opinion, it is ruled out from the outset as soon as clinical practice is planned or not ruled out.

***

1.2 Development Privilege under Article 2(8) of the AI Regulation

The AI Regulation distinguishes between an exception for purely scientific research—the so-called "science privilege"—and a rega provision governing the development and testing of an AI system prior to its market launch or commissioning, known as the “development privilege.” How do these two provisions differ, and when can a university invoke which provision?

Scientific Privilege. The research exemption applies to AI systems used exclusively for research purposes. There must be no plans for, nor any possibility of, subsequent practical application. If these conditions are met, the AI Regulation does not apply to the system.

Development Privilege. The development privilege, on the other hand, applies to the development and testing phase prior to market launch or commissioning. Subsequent practical use may be planned in this context. However, the privilege applies only until the system is put into service; from that point on, the relevant requirements of the AI Regulation must be met.

More decisive Transition. Once the AI system is released, put into operation, or used commercially, the privileges no longer apply, according to the underlying classification.

What specific implications does this distinction have for a university that initially develops an AI system as part of a research project but later makes it publicly available? or would like to implement it in practice? What requirements need to be addressed or documented even during the development phase?

None Scientific Privilege at planned Usage. If the AI is intended for future practical use, the university cannot invoke the academic privilege. The obligations under the AI Regulation become relevant no later than when the AI is put into operation.

Preparation already while the Development. To ensure that the requirements for commissioning can be met, the necessary measures must be prepared and documented as early as the development phase. The expert opinion cites the technical documentation required under Article 11 of the AI Regulation as an example.

***

1.3 Role of the Provider in Third-Party-Funded Projects

If an AI system is developed as part of a third-party-funded research project and is subsequently published or put into operation,rd: Under the AI Regulation, who is considered the „provider“—the university, the third-party funding agency, or another participating institution? How important is it under whose name the system is released, and whether it is accessible as open-source software?

Decisive is the Appearance according to outside. Under Article 3(3) of the AI Regulation, the “provider” is generally defined as the entity that places the AI system on the market or puts it into service under its own name or brand.

• College: She is considered a provider if she publishes or puts the system into operation under her own name.
• External Funding Sources: He is considered the provider if the system is published or put into operation under his name.
• Open Source: Open-source availability does not result in a complete exemption from the AI Regulation. Article 2(12) of the AI Regulation exempts open-source AI only from certain obligations. In the case of high-risk AI, as well as the prohibitions under Article 5 and the transparency obligations under Article 50, the requirements must still be observed.

2.1 Target Audiences for AI Education

Article 4 of the AI Regulation requires providers and operators of AI systems to ensure that individuals working with these systems have sufficient AI competence. For which groups of individuals must a university provide appropriate training or Provide other learning opportunities—for employees, faculty, external service providers, and students as well?

Two central Groups of People. According to the expert opinion, this obligation applies primarily to two groups: the university’s own staff and external individuals acting on behalf of or in the interest of the university.

• Own Staff: This includes, in particular, employees, civil servants, and teachers with permanent employment contracts.
• External People: This includes, for example, temporary workers or service providers who work with the AI system on behalf of the university.
• Students: In general, they must be included if the university specifically requires the use of an AI system, such as for an academic assignment.

***

2.2 Acceptable Training Formats

What measures can eHow can a university use these resources to teach AI skills? For example, are in-person training sessions, online courses, self-paced learning programs, informational materials, or AI-supported learning programs permissible? What factors are important when selecting and designing the format?

No party required Format. The term „measures“ in Article 4 of the AI Regulation is deliberately worded in broad terms. Universities therefore have considerable discretion.

Various Formats are possible. Permitted options include, among others, in-person and online training, self-study courses, informational materials, and AI-supported training programs.

Scale is the Learning Outcomes. What matters is not the outward form, but whether the offering is tailored to the specific target audience, its prior knowledge, and the specific context of use, and whether it actually imparts the necessary skills.

***

2.3 Content of AI Competency Training

Specifically, what do university faculty and staff need to know and be able to do regarding AI, given that...Are the requirements for AI literacy being met? What technical, legal, social, ethical, and environmental topics should be taught, and how does the necessary knowledge vary depending on the task, prior knowledge, and field of application?

AI-Expertise is context-dependent. Article 3, No. 56 of the AI Regulation defines AI literacy as the skills, knowledge, and understanding required for the informed use of AI systems and for identifying opportunities and risks. The specific content required depends on the role, context, and prior knowledge of the respective target group.

• Technical Information Basic Understanding: Stakeholders should understand the basic functioning and limitations of the AI systems being used.
• More Impact: Legal, social, ethical, and environmental implications must be taken into account.
• Risks and Countermeasures: The training should address the specific risks of each operation and possible corrective measures.
• High Risk-AI: High-risk AI systems require more extensive expertise. This applies in particular to individuals who perform supervisory duties under Article 14 of the AI Regulation.

***

2.4 Mandatory Nature of Training Sessions

Article 4 of the AI Regulation requires that providers and operators act „to the best of their ability“ Ensure adequate AI proficiency. Does this mean that a university must require participation in training programs? Or is it sufficient to highlight voluntary opportunities and ask the individuals concerned to pursue continuing education on their own?

A mere Note is enough not from. According to the interpretation set forth in the expert opinion, the university must ensure the development of AI expertise in a binding manner. A non-binding reference to existing programs and a request for self-directed professional development are not sufficient.

Resources influence the Design, not that Pass the Requirement. While the phrase „to the best of its ability“ acknowledges that organizations have limited resources, it does not relieve the university of its obligation to organize appropriate and effective measures.

***

2.5 Duration of the Training Sessions

How comprehensive should AI literacy training be? Are short sessions of about 60 to 90 minutes sufficient, or are more extensive, multi-level, or certified programs necessary for certain target groups?

None flat rate Minimum duration. The AI Regulation does not specify any fixed requirements regarding duration or format. What matters is whether the AI competencies required for the respective activity are effectively taught.

The necessary Scope hangs from the Use from. A format lasting 60 to 90 minutes may be sufficient if it is appropriate for the target audience and the context of use and achieves the learning objectives. For complex tasks, high-risk AI, or special oversight responsibilities, more comprehensive formats may be required.

***

2.6 Burden of Proof

Must oneHow should the university document which employees, faculty members, or other individuals have participated in AI training? What evidence does the AI Regulation require, and what internal documentation is advisable even in the absence of an explicit legal requirement?

None express Burden of Proof. According to the interpretation adopted, Art. 4 of the AI Regulation does not contain an explicit requirement to provide proof of participation in training in a specific form.

Internal Documentation is nevertheless sensible. Nevertheless, it is recommended that institutions maintain internal documentation of the programs offered and student participation. This makes it easier for the university to demonstrate transparency in how it organizes its AI skills training.

***

2.7 External Training Programs

Can A college can fulfill its obligation to teach AI skills byDoes it merely refer to external offerings, such as state-wide training programs or freely accessible online courses? Or must it also ensure that the individuals concerned actually participate and acquire the necessary knowledge?

External Offers are possible, but not automatically sufficient. External training programs may be used. However, simply providing a link or reference to an existing program is not sufficient.

Participation and Knowledge Transfer must ensured be. According to the expert opinion, the university must ensure that the individuals concerned actually participate and that the necessary knowledge is imparted.

***

2.8 Delegation to State-Level Providers

If a state-level agency—such as a state data center, a university consortium, or a state initiative—develops an AI system or AI platform for meprovides to several universities: Can this agency delegate its responsibility for teaching AI skills to the universities that use its services? Who remains responsible for their own staff, external service providers, and end users in each case?

None complete Delegation. It is not possible to fully transfer this obligation. The state central office remains responsible for its own staff and for third parties acting on its behalf.

The Areas of Responsibility exist side by side. Each individual institution of higher education is also responsible for its own staff. Delegation is therefore conceivable only to the extent that it concerns the practical implementation for personnel at the end-user institution; that institution retains its own responsibility in any case.

Typical Examples. Examples of state-level central agencies include state data centers, university consortia, state initiatives, state media centers, and education servers that centrally provide infrastructure, AI platforms, or training.

2.9 Delegation to Faculty Members

Can a college that prShould the practical teaching of AI skills to students be delegated to faculty members, for example by incorporating relevant content into courses? What organizational and legal responsibilities still rest with the university?

The Implementation can delegated will be. The practical implementation can be delegated to faculty members, particularly by incorporating appropriate content into courses.

The Responsibility remains at the university. This requires corresponding organizational guidelines or directives from the university. A complete release from responsibility is not possible: The university must continue to ensure that the necessary competencies are actually taught.

3.1 Access to GPAI Models

If a university provides its members with institutional access to a general-purpose AI model orThe AI system—known as general-purpose AI (GPAI)—provided by ChatGPT: Does this make the university a provider, an operator, or both? Does it make a difference whether the university merely makes the model accessible or further develops it technically?

As a rule Operator. If the university provides access to the system and manages its use, it is generally the operator of the AI system. This includes, for example, setting up access and determining who within the institution is authorized to use the system.

Provider only at own Further Development. The university will only become a provider if it expands the model into its own AI system or further develops it accordingly. Simply providing access is not sufficient for this.

***

3.2 Provision under a separate name, such as „GPT@HS“

Changes the legal role of aWhat is the legal status of a university if it provides an external AI service under its own name, such as „GPT@HS“? Does this naming alone make it the provider of the AI system, or does it remain the operator of a system developed by another company?

The Name alone is not crucial. According to the expert opinion, simply designating a general AI service with a university-specific name does not alter its status as a provider or operator.

The special The rule applies only for High Risk-AI. Article 25(1)(a) of the AI Regulation, which may establish provider status when a system is labeled with its own name, applies exclusively to high-risk AI systems. Since GPAI systems are generally not classified as high-risk AI, this provision does not apply here.

Practical Episode. A designation such as „GPT@HS“ therefore does not in and of itself make the university a provider. It remains the operator when the service is provided on a purely organizational basis.

***

3.3 Modification of GPAI Systems

A university can modify a general-purpose AI model, for example, through fine-tuning—that is, targeted adjustment using additional data—by incorporating its own knowledgeAI systems created using RAG or by replacing technical components. At what scale does such a change result in an independent AI system, such that the university is considered a provider? What significance do the newly emerging risks have in this context?

Crucial are Scope and Consequences the Change. A higher education institution may be designated as a provider if the changes it makes result in the creation of an autonomous AI system or give rise to new, unmitigated risks. Not every technical adjustment is sufficient to meet this criterion.

For GPAI missing one express Regulation. The AI Regulation explicitly addresses this change in roles only for high-risk AI systems in Article 25(1)(b) and (c). There are no direct provisions for GPAI and other non-high-risk systems. However, the legal opinion extends the underlying principle to these cases.

Possible Delimitation Criteria. Signs that a system is independent may include the need to update the technical documentation or training data specifications as a result of the change. New risks arising from fine-tuning, RAG integration, or the replacement of components must also be taken into account.

***

3.4 State-Wide Access

If a state provides, through a central agency, shared access to a general-purpose AI system for multiple colleges and universities: Who is considered The provider and the operator—is it the central office, the system’s original developer, or each individual college that provides access to its members?

Vendor Role the central Position only at one own System. The state agency becomes a provider only if it creates its own AI system through appropriate modifications, such as fine-tuning. If it merely makes an existing AI system available, the system’s original developer remains the provider.

The Colleges and Universities remain Female Operators. Individual institutions of higher education are considered operators as soon as they enable institutional use and make organizational decisions regarding the use of the resource by their members.

***

3.5 External AI tools, such as plagiarism detection software

If a university uses an AI tool developed by an external party and makes it available to its members...For example, consider AI-powered plagiarism detection software: Who is the provider, and who is the operator of the system? Does this division of roles change depending on whether the system is hosted on the university’s servers or operated by an external service provider?

• Provider: The external developer who develops and places the AI tool on the market is a provider under Article 3(3) of the AI Regulation.
• Operator: The university is the operator because it uses the tool under its own responsibility and exercises organizational control over its use.
• Hosting: The physical location of the servers does not affect this classification. The decisive factor is who decides how the system is used.

***

3.6 Contractual Agreements

Can universities transfer their legal role and the associated obligations as providers or operators to the manufacturer through a contract?Is ownership transferred to the creator or developer of the AI system—for example, to OpenAI? Or does the legal classification depend on the actual circumstances, regardless of the contract? What effects might such agreements have internally and with respect to third parties?

None contractual Postponement the statutory Role. The role of provider or operator cannot be transferred to another company solely through a contract.

Crucial is the actual Division of Responsibilities. The decisive factors are objective criteria and the actual circumstances—such as who develops or markets the system and who exercises organizational control over its use.

Contracts have an effect only in the Internal ratio. Contractual agreements may establish rights, obligations, or responsibilities in the internal relationship between the parties involved. However, they do not alter the legal classification as a provider or operator with respect to third parties.

***

3.7 Open-Source AI and Subsystems

Do the so-called “original provider obligations” under Article 25(2) of the AI Regulation—that is, the obligations of the original provider set forth therein—apply, Does this also apply to AI systems that are made available as open source? And who is responsible if only individual technical components—such as a RAG pipeline for integrating proprietary information sources—are integrated into a high-risk AI system?

Open Source at High Risk-AI. In the case of open-source AI, the obligations of the original provider under the applicable classification apply only if the system is already classified as high-risk AI. The fact that a system is made available as open-source does not generally exempt it from these obligations.

Responsibility for Subsystems. Developers of individual subsystems are not themselves subject to the original provider’s obligations. However, pursuant to Article 25(4) of the AI Regulation, the original provider of the entire high-risk system must ensure that it has all the information necessary to fulfill these obligations. This includes information about integrated subsystems.

***

3.8 Integration into Your Own Applications

If a university integrates a third-party AI system into its own applications—for example, into its intranet or a learning management system: WhatWhen does the university remain merely the operator of a third-party system, and when does integration or a significant modification result in the creation of its own AI system, for which the university is considered the provider? Does it matter where the system is technically operated?

Scale is the functional Distinguishability. The classification depends on whether the third-party AI system remains functionally distinct from the university's own application.

• Sure delimited Third-party system: If the third-party AI system remains independently identifiable, the university does not become a provider merely by integrating it.
• Comprehensive or significant changed System: If the integration results in a new, usable AI system, or if the third-party system is significantly modified and put into operation under the university’s name, the university may become the provider.
• Operator Role: As soon as the university takes sole responsibility for using the system, it becomes the operator. The location of the hosting server is irrelevant in this regard.

***

3.9 AI in Third-Party Software, such as Microsoft 365

If a university enables AI features in third-party software it already uses—for example, in Microsoft 365—does that make it the provider of the AI system, the operator, or both?

None Vendor Role. The university is not a provider because it neither develops nor markets the AI system itself.

Operator Role through Approval and Usage. However, she is the operator because she independently authorizes the use of the AI function and decides on its institutional use.

4.1 GPAI Chatbots

Can A university may stipulate in its terms of use that a generally applicableDoes the fact that an AI chatbot like ChatGPT is not permitted to be used for high-risk purposes rule out its classification as a high-risk application? Or does it depend on how the chatbot is actually used, regardless of the terms of use?

GPAI is not automatically High Risk-AI. A GPAI chatbot, as a general-purpose system, is classified as a high-risk AI system not solely because of its general applications.

Terms of Use change the Classification No. Terms of use can specify permitted and prohibited uses and govern contractual matters. However, they cannot circumvent the regulatory classification.

In response to the actual Application comes it. What matters is what the system is actually used for and what risks are associated with that use. Therefore, simply prohibiting high-risk uses in the terms of service does not preclude such a classification.

***

4.2 Assessment of Learning Outcomes

The AI Regulation classifies certain AI systems in the education sector as high-risk when they evaluate learning outcomes. What constitutes such an evaluation: only fully automated grading, or also AI-generated feedback, Performance predictions and grade recommendations? How should we view this practice when a teacher reviews the AI's results and makes the final decision?

Recorded are Rating and forecast. The term encompasses the automated assessment of work or test performance—such as assigning grades—as well as the prediction of performance, for example, in the form of grade predictions.

High Risk. According to the classification in the expert opinion, automated grading or performance prediction is to be considered a high-risk application.

None High Risk at more human Final Decision. Non-binding feedback or a suggested grade, on the other hand, does not result in a high-risk classification if a human grader makes the final decision. Although the concept of grading may encompass preparatory AI-assisted support, the human final decision is decisive for the classification.

***

4.3 Guiding the Learning Process

The AI Regulation also refers to AI systems for „controlling the learning process.“ What kinds of applications does this refer to? Do they includeDoes this apply only to systems that make binding decisions regarding placement, course assignments, or the recognition of academic achievements, or does it also include voluntary self-study programs and personalized feedback tools whose results cannot be viewed by instructors?

Binding Decisions about the Educational Background. According to the report, this refers to AI systems that make binding decisions regarding educational pathways. Examples include placement tests that assign students to a course level, or decisions regarding the recognition of academic credits.

High Risk at more binding Control. Such systems, which are used by institutions and are mandatory, should be classified as high-risk applications.

Volunteers Study Aids are not recorded. Optional self-study resources and personalized feedback tools that do not result in binding decisions are not included under this classification. This applies in particular when they are intended solely for students and are not accessible to instructors.

***

4.4 Assessment of the Appropriate Level of Education

What does „assessment of the appropriate educational level“ mean in the AI Regulation? What decisions regarding a student’s future educational path or placement in a specific educational program are referred to here? How Is this different from the assessment of individual learning outcomes, such as an exam, a term paper, or other form of assessment?

Parent Decision on Educational Path. The assessment of an appropriate educational level involves overarching decisions that influence a student’s future educational path. As an example, the bill cites placement based on final exams, such as for recommendations regarding high schools or college programs.

Delimitation to individual Performance Evaluation. This should be distinguished from the assessment of individual learning outcomes. The latter pertains to specific achievements, such as a grade on an exam or a term paper.

***

4.5 AI-Powered Academic Advising

When is an AI-powered chatbot considered effective? Is a chatbot for choosing a major or academic advising considered a high-risk application? How should such a chatbot be designed and implemented so that it merely provides information and support, but does not make binding decisions about a person’s educational path?

Supportive Consulting is not high-risk. According to the report's assessment, a chatbot designed to assist with choosing a major or provide academic advising is not a high-risk application as long as it only provides support and does not make binding decisions.

Design Scale. The chatbot should therefore provide information or non-binding recommendations. The system must not make any binding assignment to a degree program or any other binding decision regarding a student's educational path.

What legal considerations must universities take into account when using open-source AI for teaching, advising, or other academic tasks, or when building their own applications based on it? What are the obligations of AI providers—Do the regulations continue to apply despite open access, particularly with regard to high-risk applications, prohibited practices, and transparency requirements? Who is responsible if the university changes the system’s original intended use?

None complete Exception. Open-source AI is only granted limited privileges. Its open availability does not result in a general exemption from the AI Regulation.

Important still in effect Duties. These requirements continue to apply, in particular, in cases involving high-risk AI or where prohibited practices under Article 5 or transparency obligations under Article 50 of the AI Regulation are involved.

Full Responsibility at Change of Purpose. If the university modifies the original purpose in such a way that a high-risk application results, it bears full responsibility under the terms of the license. At the same time, there is no entitlement to have the original open-source developer assist the university in fulfilling its obligations.

Practical Exam Question
In cases of uncertainty, special attention should be paid to determining whether the AI system makes binding decisions or whether changes to the system give rise to new risks.