AI Act

A practical guide to the EU Artificial Intelligence Act for BUT. The page summarizes who is affected by the AI Act, how to distinguish the riskiness of AI systems, what obligations may arise during the development and deployment of AI, and how to proceed with the acquisition, testing or operation of AI tools at BUT.

The AI Act, i.e. Regulation (EU) 2024/1689 of the European Parliament and of the Council, introduces harmonized rules for artificial intelligence in the EU. It is based on a risk approach: most common AI tools fall under minimal or transparency risk, but some uses are prohibited and some systems in sensitive areas can be high risk. In the university environment, it is necessary to separately assess AI in the field of education, admissions, student evaluation, personnel agenda, biometrics, access to services, processing of personal data and research or operational systems with an impact on persons.

This page is a methodical guide. It does not replace a legal, GDPR, cyber security, ethical or substantive assessment of a specific system. If the AI system is to influence the rights, obligations, opportunities or evaluation of persons, or is to be connected to BUT information systems, the procedure must be consulted in advance with the relevant responsible departments.

Quick start

How to quickly determine whether the AI Act requires further review
The situation	Probable mode	Recommended procedure at BUT
I use an AI assistant for public text design, translation, brainstorming, stylistics or code review without personal and non-public data.	Usually minimal risk, or transparency obligations depending on the context.	Verify output, comply with BUT rules for transparency, copyright and data protection.
I want to publish a chatbot for applicants, students, employees or the public.	Transparency obligations; further assessment may be required depending on function and impact.	To state that the user is interacting with AI, to limit liability, to prevent decisions about the rights of individuals and to ensure control of content.
AI is supposed to evaluate study results, track cheating, sort applicants, decide on access to studies or fundamentally influence the educational path.	It can be a high-risk education and training system.	Do not deploy without prior legal, GDPR, ethical, technical and factual assessment, determination of human supervision and documentation.
AI is supposed to filter job applicants, evaluate employee performance, propose rewards, manage work or monitor workers.	It can be a high-risk system in the field of employment and management of workers; some practices may be prohibited.	Consult HR, legal, GDPR and security departments beforehand. Ensure human supervision, non-discrimination, information and records.
AI uses biometrics, emotion recognition, camera analysis or person identification.	High risk or prohibited practice according to purpose, place and method of use.	Do not use without express assessment and approval. Recognizing emotions in the workplace and in educational institutions is particularly risky.
A research team develops a model or prototype for scientific research only.	It may fall outside the scope of the AI Act if the system or model is developed and put into operation solely for the purpose of scientific research and development.	Separate research use from operational deployment, commercialization and real world testing. Other rules, especially GDPR, copyright and security, remain in effect.

What is the AI Act

The AI Act is a European regulation that establishes a uniform legal framework for general purpose AI systems and models in the EU. The goal is to promote trustworthy AI while protecting health, safety, fundamental rights, democracy and the rule of law. This is not a blanket ban on the use of AI. Duties vary according to intended purpose, impact on individuals, type of data, degree of autonomy and role of the organization.

The following principles are key for BUT:

Risk approach. Not only the technology is evaluated, but above all the specific purpose, context and impact of use.
Human responsibility. AI may not replace the responsibility of the user, teacher, manager, project investigator, system administrator or BUT authorities for the final decision.
Transparency. In relevant situations, it must be clear that a person is interacting with an AI or that the content has been created or edited by an AI.
Human supervision. For high-risk systems, it is necessary to ensure qualified human supervision, monitoring, records and response to incidents.
Continuity with other rules. The AI Act does not replace GDPR, copyright, cyber security, contracts, BUT internal standards or research ethics rules.

Application schedule

The AI Act does not apply all at once. Individual parts are launched sequentially. At the same time, the legislative process for the Digital Omnibus proposal is underway in 2026, which may change some deadlines, especially for high-risk AI systems. Until the change is formally adopted and announced, it is advisable to prepare according to the valid wording and at the same time monitor the final result of the Omnibus.

The main dates of the AI Act for BUT orientation
The date	What is applied or prepared	Impact on BUT
1. 8. 2024	The AI Act came into force.	It is advisable to build AI management, records of used tools, purchase rules and training.
2. 2. 2025	General Provisions, Definitions, Obligation of AI Literacy and Prohibited AI Practices apply.	BUT must have adequate measures for the AI literacy of persons who use or operate AI and must not use prohibited practices.
2. 8. 2025	Rules for general purpose models, the so-called GPAI, and governance at EU level apply.	Relevant to teams developing, publishing, or integrating generic models and applications on top of them.
2. 8. 2026	As it stands, the general applicability of much of the AI Act begins, including transparency rules and a number of obligations for high-risk systems.	By this date, it is advisable to have an inventory of AI systems, a risk classification process, deployment records, a process for assessing high-risk systems and training of relevant roles.
2. 8. 2027	According to the original schedule, selected obligations are running out, especially for some high-risk systems linked to regulated products and for the previously mentioned GPAI models.	Mainly relevant to development and research teams integrating AI into regulated products, healthcare, industrial or security solutions.
2026 – monitored change	The Digital Omnibus proposal may delay the application of some obligations for high-risk systems. Both the European Parliament and the Council in March 2026 supported later fixed dates for some of these systems.	It is safer for planning not to wait for a delay and to prepare inventory, risk classification and documentation now.

The role of BUT

The same tool may have different obligations depending on whether BUT only uses it, develops it for its own operation, integrates it into a service, provides it to third parties or markets it.

Typical roles under the AI Act
Role	What does it mean	Example at BUT	What to watch out for
Deployer	The organization uses the AI system under its authority outside of purely personal, non-professional use.	The BUT component uses a licensed AI tool for administration, teaching, project support, communication or HR process.	In the case of high-risk systems, obligations arise to use according to instructions, human supervision, logs, monitoring, informing and possibly impact assessment.
Provider	The entity develops an AI system or model and puts it on the market or into operation under its own name or brand.	A BUT research team or project will create an AI service for external users, spin-offs, partners or the public.	High-risk systems may incur extensive responsibilities for risk management, documentation, data quality, human oversight, cyber security and compliance assessment.
Integrator or downstream provider	The entity builds its own application on top of a foreign model of general purpose or a ready-made component.	The BUT project will create an application on top of the LLM that advises students, processes applications or provides analysis to partners.	It is necessary to distinguish between the responsibility of the model supplier and the responsibility of BUT for the specific purpose, integration, data, users and impacts.
Research organization	An AI system or model is developed or used exclusively for scientific research and development.	Laboratory model prototype, academic experiment or model evaluation without operational deployment in real decision-making.	An exemption does not mean an exemption from GDPR, copyright, ethics, contracts and cybersecurity. A new assessment is required during the transition to operation.

Risk classes of AI systems

AI Act's risky approach
Risk class	Description	An example in a university environment	Recommended procedure
Prohibited Practices	Use of AI considered unacceptable in terms of fundamental rights, manipulation, abuse of vulnerability, social scoring, certain biometrics or emotion recognition.	Recognizing the emotions of students during an exam or employees in the workplace, social scoring of people, manipulative system abusing people's vulnerability.	Do not deploy. If a similar function exists in the supplied tool, it must be disabled or the tool must not be used for that purpose.
High risk systems	Systems in sensitive areas or as safety components of regulated products.	AI for academic performance assessment, cheating monitoring, selection of study applicants, job applicant filtering, worker assessment, selected biometric and healthcare applications.	Before acquisition or deployment, perform classification, legal and technical assessment, GDPR/DPIA, or FRIA, determine human supervision, logs, monitoring and responsibilities.
Transparency risk	Systems where it must be clear to humans that they are interacting with AI or that the content is synthetic or AI generated.	Chatbot for applicants, AI generated image or video for communication, deepfake, synthetic voice, AI text informing the public about a matter of public interest.	Visibly identify AI, ensure content control, and respect copyright and personality rights.
Minimal or no risk	Common applications without a significant impact on the rights and safety of persons.	Spam filter, language proofreading of public text, auxiliary outline design, brainstorming, internal proof-reading without personal or sensitive data.	Comply with BUT rules, verify outputs, protect data and transparently state the use of AI where required.

Typical scenarios at BUT

Orientation assessment of selected university scenarios
The script	AI Act risk	What needs to be resolved
AI assistant for working with public texts, translation, summary or stylization.	Usually minimal risk.	Authorized use of data, verification of output, citation or indication of the scope of use of AI according to BUT rules.
AI tool for working with internal documents, contracts, project requests or minutes.	It depends on the data and purpose; often primarily GDPR, contractual and security risk.	Choice of appropriate tool, prohibition of entering sensitive data into unverified services, data minimization, contractual and security assessment.
Chatbot for applicants, students or employees.	Transparency risk; it can be increased if it provides individualized advice with an impact on rights.	AI Flagging, Approved Content, Escalation to Human, Hallucination Check and Information Timeliness.
AI for admissions, sorting applicants, assessing students, monitoring exams or deciding on access to education.	Potentially high-risk use in education and training.	Comprehensive pre-deployment assessment, human oversight, non-discrimination, explainability, records, GDPR/DPIA and legal review.
AI for personnel agenda, recruitment, CV filtering, employee evaluation or work distribution.	Potentially high-risk use in employment and workforce management.	Legal, personnel, GDPR and ethical assessment, notification of affected persons, human supervision, bias control, records and possibility of correction.
AI in research prototype with no operational deployment and no market release.	It may fall outside the scope of the AI Act if the research exemption is met or if it is pre-market or pre-operational R&D.	Describe the purpose and boundaries of the research, assess GDPR, ethics, contracts, copyright and security. Perform a new assessment when switching to operation.
AI in an industrial, healthcare, construction, transportation or cybersecurity product.	It may be a high-risk system by product or intended purpose.	Address already at the design stage: regulated products, conformity assessment, documentation, data quality, robustness, cyber security, partner liability and contracts.
AI agent that performs actions in BUT systems or on behalf of users.	It depends on the level of autonomy, agent permissions, data and impact. It can be a transparency or high risk system.	Limit permissions, implement human approval of steps, audit trails, testing, safety limits and a shutdown plan.

Prohibited AI practices

From 02/02/2025, bans on selected practices with an unacceptable risk apply. It is particularly important for BUT that the ban also applies to the use of ready-made tools, if they were used for a prohibited purpose.

Manipulative or deceptive techniques, which can significantly influence people's behavior and cause them harm.
Exploiting people's vulnerabilities, for example age, health status, disability or social situation.
Social scoring persons leading to unjustified or disproportionate disadvantage.
Selected biometric practices, including biometric categorization to derive sensitive characteristics and untargeted bulk collection of facial images for facial recognition databases.
Recognizing emotions in the workplace and in educational institutions, with narrow exceptions for health or safety reasons.
Predictive Police Assessment based only on profiling or personality traits, which is not a common university scenario, but belongs to the prohibited areas.

Practical rule: if a tool promises to reveal the emotions of students or employees, automatic assessment of personality, social reliability, riskiness of persons or covertly influencing behavior, it must not be used at BUT without an immediate legal and ethical assessment. In many cases, it can be directly a prohibited practice.

High risk AI systems

A high-risk AI system can arise in two ways: either the AI is part of a regulated product or security component, or it falls under one of the specific uses in Annex III of the AI Act. For the university environment, the most important areas are education, employment, biometrics, critical infrastructure, access to selected services and projects in regulated industries.

Areas of Annex III that may apply to BUT

Education and training. For example, AI for admissions to study, evaluation of results, direction of study path, allocation of access to educational opportunities or monitoring of cheating.
Employment and management of workers. For example, AI for targeted job advertising, analysis and filtering of applications, evaluation of applicants, performance evaluation, distribution of tasks or decision-making with an impact on the employment relationship.
Biometrics. For example, remote identification of persons, biometric categorization or some biometric authentication in operational systems.
Critical infrastructure and regulated products. Relevant to research and development in energy, transportation, medical devices, machinery, robotics, construction, cyber security and industrial systems.
Access to essential services or benefits. It may be relevant for systems making decisions about services, support, accommodation, social or health aspects, if AI would affect people's access to important services.

Obligations of providers of high-risk systems

If BUT or a BUT project acts as a provider of a high-risk AI system, it is necessary to plan in advance in particular:

risk management system throughout the life cycle,
data governance and data quality, including anti-discrimination measures,
technical documentation and instructions for use,
automatically generated logs and traceability,
transparency for deployers and affected persons,
human supervision,
accuracy, robustness and cyber security,
conformity assessment, registration where required, post-commissioning monitoring and reporting of serious incidents.

Obligations of deployers of high-risk systems

If BUT buys, licenses or uses a high-risk AI system, it typically acts as the deployer. Then you especially need:

use the system according to the provider's instructions,
identify persons responsible for human supervision and ensure their competence, training and authority to intervene,
ensure that input data is relevant and reasonably representative for the intended purpose, if BUT has control over it,
monitor the operation of the system and, in the event of a risk or serious incident, inform the provider and the competent authorities,
keep logs, if the system generates them, and comply with documentation obligations,
inform affected workers and, where appropriate, their representatives before using a high-risk AI system in the workplace,
for selected systems, carry out an assessment of the impact on fundamental rights, or use or follow up on the DPIA according to the GDPR.

FRIA: fundamental rights impact assessment

For some high-risk systems, the deployer must perform a Fundamental Rights Impact Assessment before first use. For BUT, this may be relevant especially for systems with an impact on students, applicants, employees or the public. The assessment should describe the purpose and process, time and frequency of use, groups of people affected, risks of harm, human supervision and measures in case risks materialize.

Transparency and generative AI

Transparency is key both according to the AI Act and BUT rules. At BUT, if the result was achieved using generative AI, this fact should be properly stated, including the scope of use. In operational practice, it is advisable to distinguish at least the following situations:

Transparency in the use of AI
The situation	What to do
A chatbot or AI assistant communicates with a human	Clearly inform that it is communicating with AI and offer a way to contact a human if it is an individual or binding matter.
AI generated or edited image, sound or video	Ensure labeling if it could be mistaken for real content, and address consent of depicted or impersonated persons.
Deepfake or synthetic voice	To be used only with clear labelling, legitimate purpose and consent of affected persons if necessary.
AI text informing the public about a matter of public interest	Provide human editorial review, fact-checking, and transparent labeling when context or rules require it.
Use of AI in academic, scientific or work output	Indicate the tool, purpose and scope of use according to BUT rules, instructions of the teacher, publisher or grant provider.

Research, prototypes and open-source

The AI Act contains exemptions for research and development, but they are not unlimited. It does not mean that any academic project is automatically outside the AI Act.

Exclusively scientific research and development. AI systems or models specifically developed and put into operation solely for scientific research and development may be outside the scope of the AI Act.
Research, testing and development before market or operation. These activities are generally outside the scope of the AI Act unless it involves testing in real-world conditions, which has a special regime.
Transition to operation or commercialization. Once the prototype is used for real decision-making, for service to third parties, for commercialization, or for integration into a regulated product, a reassessment is required.
Open-source is not an automatic exception. The release of a model or system as open-source may have a different regime depending on the type of model, riskiness, way of marketing and other conditions.
Other duties remain. GDPR, copyright law, contracts with partners, ethical approvals, rules of grant providers, cyber security and BUT internal standards will apply even where the AI Act does not directly apply.

Article 4 AI Literacy

The AI Act imposes an obligation on providers and deployers of AI systems to take reasonable measures to ensure sufficient AI literacy of persons who operate or use AI systems on their behalf. It's not just technical training. AI literacy includes knowledge of the opportunities and risks of AI, the ability to correctly interpret outputs, understanding roles and responsibilities, data protection, fundamental rights, ethics, human oversight and the rules of a specific tool.

Recommended minimum structure of AI literacy at BUT
Target group	Minimal content	Documenting
All users of AI in a work or study context	Basics of AI, limits and hallucinations, BUT rules, transparency, copyright, prohibition of entering sensitive data into unverified tools.	Completed training, methodical material, agreement with the principles or records of familiarization.
Executives and agenda owners	AI Act risk approach, person impact assessment, tool procurement responsibilities, human oversight, contracts and incidents.	Record of training, records of deployment decisions, regular review of the system.
System administrators, developers and project teams	AI system classification, data governance, logging, cybersecurity, monitoring, documentation, supply chain, API, agents, testing and validation.	Technical documentation, test reports, model cards, data sheets, configuration and change records.
Persons overseeing high-risk systems	Provider instructions, system boundaries, interpretation of outputs, human intervention, escalation, incidents, non-discrimination, work with affected persons.	Designation of responsible person, documented training, supervision procedure and operational records.

Recommended procedure at BUT

Describe the purpose. What should the AI system do, for whom, in what agenda and with what effects?
Determine the role of BUT. Is BUT a deployer, provider, researcher, integrator, or just an end user of a foreign service?
Map the data. Is it public, internal, personal, sensitive, contractually protected, research or security data?
Exclude Prohibited Practices. Check manipulation, social scoring, emotion recognition, biometrics and other unacceptable scenarios.
Assess high risk. Review the areas of education, employment, biometrics, critical infrastructure, access to services and regulated products.
Assess transparency. Should a person be informed that they are dealing with an AI? Is the content synthetic or AI generated?
Address GDPR, DPIA and FRIA. If personal data is processed or the system may affect fundamental rights, involve the responsible persons.
Verify suppliers and contracts. Review license terms, data processing, data placement, retention, training on inputs, audit rights, data export and incident liability.
Prepare human supervision. Designate those with the competence, training, authority, and procedure to approve, change, or stop AI output.
Ensure records and reviews. Record owner, purpose, risk class, dates, vendors, settings, logs, training and date of next review.

Control checklist

Checklist before acquiring or deploying an AI system
Question	Yes / No / Note
Is the purpose of the AI system and the responsible agenda owner clearly described?
Is it determined whether BUT acts as an implementer, provider, integrator or research organization?
Has the system been verified for prohibited practices?
Has it been assessed whether the scheme may be high risk under Annex III or regulated products?
Are the affected persons and possible impacts on their rights, opportunities or security described?
Is it clear what data will enter the system and whether it contains personal, sensitive, non-public or contractually protected information?
Has GDPR, possibly DPIA, FRIA or ethics approval been assessed?
Are the license, contractual and security conditions of the supplier checked?
Is human supervision, responsibility for output control and the ability to stop the system specified?
Are the rules of transparency towards users and affected persons established?
Are access rights, logging, monitoring, incident procedure and regular reassessment set?
Are users and supervisors trained according to their role?

AI system record template

For any AI system or pilot deployment that goes beyond individual assistance work with public data, it is advisable to keep a simple record. It will help with risk review, GDPR compliance, cybersecurity, procurement, contracts and AI literacy.

Recommended AI system registration fields
Field	What to fill out
System / tool name	The name of the product, model, internal application or pilot.
Owner and responsible component	Faculty, institute, department, project, responsible person.
Purpose and intended use	A description of the process, users, outputs and decisions that the system supports.
The role of BUT	Deployer, provider, integrator, research prototype, other role.
Affected persons	Students, applicants, employees, research participants, the public, suppliers, partners.
Data	Data category, source, legal title, personal data, sensitive data, retention, location, accesses.
Risk classification	Out of Scope, Minimal Risk, Transparency, Suspected High Risk, High Risk System, Prohibited Practice Excluded.
AI Act obligations	Transparency, human supervision, logs, FRIA, DPIA, registration, documentation, incidents, regular review.
Supplier and contracts	Service provider, contract, DPA, license, API, audit rights, data training, subcontractors, data location.
Human supervision	Who controls the outputs, when they can intervene, how errors escalate and how the system shuts down.
Training and AI literacy	Who is trained, on what date, to what extent and for what role.
Date of approval and further review	Decision date, approver, review date, trigger when model, purpose, data or supplier change.

VUT

Faculties and university institutes

Parts