Frequently asked questions about EU AI Act Commission guidance

Use these questions to check which EU AI Act rules and Commission interpretations apply to a system, model or use.

Disclaimer

This page gives general information. It is not legal advice.

Is European Commission guidance on the EU AI Act legally binding?

No. Commission guidelines are non-binding interpretations of the EU AI Act. The amended Act governs, and only the CJEU can ultimately give an authoritative interpretation (Commission guidelines on the definition of an AI system, paragraph (7)).

The high-risk guidelines are a draft that can change. The 20 July 2026 transparency Communication makes applicability conditional on subsequent formal adoption.

The Commission bases GPAI enforcement on its GPAI-scope Guidelines, but requires a "case-by-case assessment" (paragraph (9)).

See Commission guidance.

Which EU AI Act Commission guidance document should I use?

Select the Commission guidance for your EU AI Act decision:

Decision	Commission text	Page
Is the software an AI system?	Guidelines on the definition of an AI system	Definition
Is the practice prohibited?	Guidelines on prohibited AI practices	Prohibited practices
Is the AI system high-risk?	Draft guidelines on the classification of high-risk AI systems	Classification and Examples
Which Article 50 duties apply?	Guidelines on Article 50 transparency obligations	Transparency
Is the model a GPAI model, and which provider duties apply?	Guidelines on the scope of the obligations for providers of GPAI models	GPAI models

The overview links to each text and its legal status. This selection is not a complete inventory: training-summary materials and draft Article 73 incident guidance are other examples. The high-risk draft has three downloads: general principles, Annex I and Annex III. It is not necessary to read all three for each question.

See Commission guidance.

When were the EU AI Act Commission guidelines published and adopted?

The Commission published and adopted the EU AI Act guidance on different dates:

• Prohibited practices: published 4 February 2025, adopted 29 July 2025.
• AI system definition: published 6 February 2025, adopted 29 July 2025.
• GPAI scope: published 18 July 2025, adopted 19 November 2025.
• High-risk classification: draft published on 19 May 2026. Consultation closed 23 July 2026, with adoption planned by the end of 2026.
• Transparency: content approved 20 July 2026, with applicability conditional on subsequent formal adoption.

These are different from the Act's application dates.

See Commission guidance and Application dates.

Do Commission guidelines replace the EU AI Act?

No. The Commission guidelines give the Commission's interpretation of the EU AI Act. They do not replace or change the Act.

The definition guidelines, the guidelines on prohibited practices, and the GPAI-scope Guidelines are from 2025. Then Regulation (EU) 2026/1744 amended the Act, for example Articles 2(2), 5, 6, 50, and 56. Where a guideline from 2025 and the amended Act do not agree, the amended Act applies. For example, the 2025 guidelines do not give an interpretation of Article 5(1), points (ba) and (bb). Paragraph (96) of the GPAI-scope Guidelines refers to Article 56(6) before Regulation (EU) 2026/1744 replaced it. See Commission guidance.

How should an organization use Commission guidance in its EU AI Act assessment?

Use the Commission guidance to examine your EU AI Act role, system or model, intended purpose and applicable obligations. Examine each system individually.

Modulos recommends a record of the reasoning, and departures from the Commission's interpretation. This is not a general statutory duty. Specified duties, such as Article 6(4)'s "assessment" record, also apply. The EU AI Act Frameworks are OFF-1 at Organization level and MFF-1 at AI System level. The Requirement table and setup steps are on Operationalizing the EU AI Act in Modulos.

See Commission guidance and Use Modulos.

Does EU AI Act Commission guidance apply differently in Germany?

No. The Commission guidance on the EU AI Act is EU guidance. It is not binding in Germany or in a different Member State. The guidelines on prohibited practices are for the "consistent, effective and uniform application" of Article 5 (paragraph (4)).

The national law of a Member State can be important for some obligations of the EU AI Act. For example, Article 5(5) lets a Member State "provide for the possibility to fully or partially authorise" real-time remote biometric identification for law enforcement. The Member State does this in its national law, in the limits of Article 5. See Commission guidance and Real-time RBI authorization.

What counts as an AI system under Article 3(1) of the EU AI Act?

Article 3(1) of the EU AI Act gives the binding definition. The Commission divides it into seven elements:

• machine-based system
• varying levels of autonomy
• possible adaptiveness after deployment
• explicit or implicit objectives
• inference from inputs about how to generate outputs
• outputs such as predictions, content, recommendations or decisions
• possible influence on physical or virtual environments.

Adaptiveness is optional. Inference includes machine learning and logic-based or knowledge-based techniques. Examine the elements across the lifecycle (Commission guidelines on the definition of an AI system, paragraphs (9)–(10), (23), (39)). Article 2's scope exclusions and Section B rule also apply.

See The seven elements of Article 3(1) and Scope of the AI Act.

Is a deterministic rule engine an AI system under the EU AI Act?

A deterministic rule engine can be an AI system as defined in Article 3(1) of the EU AI Act. Check its architecture and function against the seven elements of the definition.

Logic-based inference can qualify, with inference from rules encoded by human experts (Commission guidelines on the definition of an AI system, paragraph (39)). Mere execution of human-defined rules differs (paragraphs (40), (46)). Basic data processing and classical heuristics have their own conditions for exclusion (paragraphs (46)–(48)). The same outputs for the same inputs, or no learning after deployment, do not alone put a system out of the definition.

See Systems that are not AI systems.

Is a linear regression model an AI system under the EU AI Act?

Examine all Article 3(1) elements of the EU AI Act for a system that uses linear regression. The technique alone does not give the result.

The Commission guidelines on the definition of an AI system give a group with conditions: systems "used to improve mathematical optimisation or to accelerate and approximate traditional, well established optimisation methods, such as linear or logistic regression methods" (paragraph (42)). These systems can be out of the definition when "they do not transcend ‘basic data processing’" (paragraphs (42) to (45)). The guidelines give systems for basic predictions as a different group (paragraphs (49) to (51)). For all other uses, examine the inference, the function, and the other elements. See Systems that are not AI systems.

Does fine-tuning a GPAI model create a new AI system?

Fine-tuning alone does not by itself make a new AI system in the EU AI Act. A model needs more components to become a system (recital 97). Examine the resulting architecture and function against Article 3(1).

Provider status is a different question. A different downstream modifier can become a GPAI provider through a "significant change" in generality, capabilities or systemic risk (GPAI-scope Guidelines, paragraphs (57)–(59)). Development by or for the original provider follows the same-model lifecycle rule (paragraph (22)).

See AI systems and GPAI models and When a model modifier becomes a provider.

Is a wrapper around an external LLM API an AI system?

A wrapper around an external LLM API can be an AI system if the combined software meets Article 3(1) of the EU AI Act. Examine architecture, function, outputs and interaction with the environment (Commission guidelines on the definition of an AI system, paragraph (61)).

The name of the software does not show if it is an AI system. Check the combined software against the seven elements. A model needs more components to become a system (recital 97). Integration does not automatically make a new system or put it out of the definition. Model and system providers can be different actors.

See AI systems and GPAI models.

Must an AI system keep learning after deployment to fall under the EU AI Act?

No. Learning after deployment is not necessary for the definition in Article 3(1) of the EU AI Act. A system "may, but does not necessarily have to, possess adaptiveness or self-learning capabilities after deployment to constitute an AI system" (Commission guidelines on the definition of an AI system, paragraph (23)).

The other elements must apply. For example, the system must operate with "some reasonable degree of independence of actions" (paragraph (20)), and it must infer how to generate outputs. A model with weights that do not change in use does not put the system out of the definition. Examine the elements across the full lifecycle of the system. See The seven elements of Article 3(1).

Is the Commission's seven-element AI-system test binding?

No. The seven elements are the Commission's interpretation of the definition in Article 3(1) of the EU AI Act. This interpretation is not binding. Article 3(1) is the binding law.

The Commission guidelines on the definition of an AI system tell the reader that "each system must be assessed based on its specific characteristics" (paragraph (6)). Article 2 also applies, which gives the scope of the Act and its exclusions. See Commission guidance.

Which AI practices are prohibited by Article 5 of the EU AI Act?

Article 5(1) prohibits these practices only when each point's conditions apply:

• (a) specified subliminal, manipulative or deceptive techniques, with the objective or effect of materially distorting behavior "by appreciably impairing their ability to make an informed decision". The impairment must cause an otherwise-unmade decision in a manner that causes or is reasonably likely to cause significant harm.
• (b) exploitation of vulnerabilities due to age, disability or a specific social or economic situation, with the objective or effect of materially distorting behavior. The required causal connection to significant harm or its reasonable likelihood must also apply.
• (ba) realistic intimate material of identifiable persons without the specified consent.
• (bb) child material or performance defined in Directive 2011/93/EU, Article 2(c)/(e), subject to the national “without right” defense.
• (c) social scoring from behavior or personal characteristics across time, causing unfavorable treatment in unrelated contexts or unjustified or disproportionate treatment.
• (d) natural-person criminal-offense risk prediction based solely on profiling or "assessing their personality traits and characteristics".
• (e) untargeted internet or CCTV facial-image scraping to make or expand facial-recognition databases.
• (f) workplace or education emotion recognition, unless for medical or safety reasons.
• (g) biometric inference of race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation.
• (h) real-time remote biometric identification in publicly accessible spaces for law enforcement, with narrow exceptions.

Each point has its own conditions for actors, actions, exclusions and exceptions. Article 2(2) excludes Article 6(1) high-risk systems related to Annex I Section B products from Article 5. Points (ba)/(bb) and paragraphs 1a/1b apply from 2 December 2026. The prohibitions from 2024 apply from 2 February 2025. The Commission guidelines on prohibited AI practices from 2025 predate the additions.

See Article 5 and the Commission guidance.

Does the EU AI Act ban personalized dark patterns?

Only if all Article 5(1)(a) conditions apply. The AI system must use the specified subliminal, manipulative or deceptive techniques, with the objective or effect of materially distorting behavior. The distortion must occur "by appreciably impairing their ability to make an informed decision" and cause an otherwise-unmade decision. Significant harm, or its reasonable likelihood, must follow through the required causal connection.

The effect can suffice without an objective to distort. The Commission distinguishes lawful persuasion and gives dark patterns with "significant harms" as examples (Commission guidelines on prohibited AI practices, paragraphs (127), (139)). Article 2's scope rules also apply.

See Article 5(1)(a).

Is fraud-detection profiling prohibited by the EU AI Act?

A fraud-detection name does not show if Article 5(1)(d) applies. The prohibition concerns natural-person criminal-offense prediction based solely on profiling or "assessing their personality traits and characteristics".

The qualifications in the Commission guidelines on prohibited AI practices include meaningful other facts (paragraph (202)), the narrow human-assessment exclusion, and ordinary private-business profiling without the purpose of predicting a "specific criminal offence" (paragraphs (208)–(211)). The human assessment must be "already based on objective and verifiable facts directly linked to a criminal activity" (Article 5(1)(d)). Criminal offense has an autonomous EU-law meaning (paragraph (218)). Annex III's financial-fraud exclusion is a different classification question.

See Article 5(1)(d).

Is anti-cheat AI in online games prohibited social scoring?

Only if all Article 5(1)(c) conditions apply. Evaluation or classification across a period of time must use social behavior or personal or personality characteristics. The resulting social score must cause detrimental or unfavorable treatment in unrelated contexts, or unjustified or disproportionate treatment.

The second limb can apply in the same game. The Commission guidelines on prohibited AI practices examine the period "taking all circumstances of the case into account", without a fixed minimum (paragraph (155)). Other Article 5 prohibitions stay different tests.

See Article 5(1)(c).

Is emotion recognition in call-center quality monitoring prohibited?

Yes, if the system infers workers' emotions from biometric data, unless Article 5(1)(f)'s medical or safety exception applies.

The Commission guidelines on prohibited AI practices give these results:

• Point (f) covers work in physical and virtual spaces, and candidates during recruitment (paragraph (254)).
• Point (f) can apply to emotion inference from voice or keystroke biometrics. It does not apply to written-text-only sentiment analysis (paragraph (251)).
• Point (f) does not apply to customer-only emotion analysis (paragraph (254)). The Commission also excludes unavoidable incidental employee emotion inference from point (f) when the system's primary objective does not target employees' emotions. Deployers must use safeguards to prevent employee screening and make sure employees are not "adversely affected" (paragraph (270)).
• Personal training is permitted only without sharing results with HR, effects on the work relationship, or circumvention of the prohibition (paragraph (254)).
• Point (f) does not apply to detection of physical states such as fatigue (paragraph (249)).

The medical or safety exception is narrow (paragraphs (256)–(258)). Article 2's scope rules also apply.

See Article 5(1)(f).

Does the medical or safety exception allow workplace wellness emotion recognition?

No. A wellness name does not show that a use qualifies for Article 5(1)(f)'s medical or safety exception. In the Commission guidelines on prohibited AI practices, that exception does not include general wellbeing, monitoring of "stress levels", burnout and depression detection (paragraph (257)).

Its interpretation limits therapeutic uses to CE-marked medical devices and safety reasons to protection of life and health (paragraphs (257)–(258)). The prohibition requires biometric emotion inference in workplace or education settings. Permitted uses can also have Annex III(1)(c) and Article 50(3) duties.

See Article 5(1)(f).

Are commercial face-search products banned by the EU AI Act?

Face-search capacity alone is not prohibited. Article 5(1)(e) concerns untargeted scraping from the internet or CCTV to make or expand facial-recognition databases.

Facial recognition does not have to be the database's exclusive purpose (Commission guidelines on prohibited AI practices, paragraph (226)). Point (e) has no consent exception. Targeted searches are different, but mixed searches keep the prohibition on their untargeted part (paragraphs (229)–(232)). Article 2(6) and (8) research exclusions have different conditions. Other biometric and data-protection rules can also apply.

See Article 5(1)(e).

Does the EU AI Act ban AI-generated intimate images?

Yes, on conditions, from 2 December 2026. Article 5(1)(ba) covers realistic intimate material of identifiable persons without the specified consent. Point (bb) covers Directive 2011/93/EU Article 2(c)/(e) material or performance, subject to a national “without right” defense.

Article 5(1a) distinguishes supply from use. Supply requires the intended-purpose test or the alternative test in Article 5(1a)(a)(ii). That alternative requires reasonably foreseeable, reproducible generation or manipulation "without requiring significant technical modification". It also requires that the system has no specified "reasonable and adequate" safeguards to prevent that generation or manipulation and correct misuse.

Prohibited use requires the deployer's generation or manipulation purpose. Article 5(1b) limits only point (ba), for manipulation without increased intimate exposure or changed sexually explicit activities. The 2025 guidelines predate these additions.

See the material definitions and Article 5(1a)–(1b) conditions.

Does transparency make a prohibited AI use legal?

No. Information to persons in Article 50 of the EU AI Act does not make lawful a practice that Article 5 prohibits. In recital 137, compliance with the transparency obligations "should not be interpreted as indicating that the use of the AI system or its output is lawful".

• A prohibition in Article 5 does not remove the Article 50 obligations that apply to a system.
• When Article 5 does not apply to a practice, this does not show that Article 50 applies. Each Article 50 obligation has its conditions and exceptions.
• Article 2(2) removes Articles 5 and 50 for Article 6(1) systems related to Annex I Section B products.

See Article 50 transparency.

Is an AI system listed in Annex III automatically high-risk?

Annex III systems are high-risk unless Article 6(3) applies. The exception applies when one or more Article 6(3)(a)–(d) conditions are met. Profiling natural persons prevents the exception.

The Commission draft guidelines on high-risk classification say that the conditions "must be interpreted narrowly", in the light of the first subparagraph of Article 6(3): the system "should not materially influence the outcome of the decision". There is "no separate or independent assessment" of risk besides those conditions (Annex III part, paragraph (88)). A system that is part of a complex system stays high-risk if the joint outputs "materially influence an individual decision" (paragraph (90)). Article 6(3) does not apply to Annex I systems. Providers must document the "assessment" and complete registration before placing on the market or putting into service (Articles 6(4), 49(2)).

See Article 6(3).

What is a narrow procedural task under Article 6(3)(a)?

The Commission draft guidelines on high-risk classification give Article 6(3)(a) examples: structuring data, sorting incoming documents and duplicate detection (Annex III part, paragraph (91)). Scoring or ranking input data is different (paragraph (93)).

The draft reads this condition in the light of the first subparagraph of Article 6(3): the system "should not materially influence the outcome of the decision". There is "no separate or independent assessment" of risk besides the conditions (paragraph (88)). Profiling prevents the exception (Article 6(3)). Purely administrative judicial anonymization is not in Annex III(8)(a), and an exception is not necessary (paragraph (419)). A preparatory task is a different condition in Article 6(3)(d).

See Article 6(3).

Can an AI system that profiles people use the Article 6(3) exception?

No. An Annex III AI system that profiles natural persons is always high-risk, also when an Article 6(3)(a)–(d) condition applies.

This rule does not classify all profiling systems not in Annex III as high-risk. Article 3(52) gives the profiling definition. Simple classification by age, sex or height is not always profiling: examine its purpose and evaluation of personal characteristics (Commission draft guidelines on high-risk classification, Annex III part, paragraph (112)).

See Article 6(3).

Does an AI medical device fall under Annex I Section A or Section B?

The Medical Devices Regulation and In Vitro Diagnostic Medical Devices Regulation are in Annex I Section A of the EU AI Act.

A medical device AI system is high-risk in Article 6(1) only if all these conditions apply:

• The AI system is a product or an intended safety component of a product covered by Annex I legislation.
• That product must undergo third-party conformity assessment.

Article 6(1a) to (1c) give the safety-component conditions. Third-party assessments required solely for risks other than health and safety do not satisfy Article 6(1)(b). Thus, not all medical device AI systems are high-risk in Article 6(1). Chapter III Sections 1–3 apply to Annex I systems from 2 August 2028. Section B concerns different legislation and has its own Article 2(2) scope rule.

See Article 6(1) and Annex I products.

Must I register an Annex III system that I assess as not high-risk?

Yes, when relying on Article 6(3). Before placing on the market or putting into service, the provider must document its "assessment" (Article 6(4)). The provider or applicable authorized representative must register itself and the system by the same deadline (Article 49(2)). The provider must give the documentation "Upon request" of national competent authorities.

Amended Annex VIII Section B retains registration but removes points 7 and 9. Article 49 applies from 2 August 2026, while Article 6(3) for Annex III systems applies from 2 December 2027. Systems not in Annex III have no registration duty from Article 49(2).

See Article 6(3).

Does intended purpose or actual use decide AI Act high-risk classification?

The provider's intended purpose governs high-risk classification in the EU AI Act. Article 3(12) includes instructions, promotional or sales materials, statements and technical documentation.

In the Commission draft guidelines on high-risk classification, the limits across those materials must not contradict each other. A terms-of-service exclusion cannot override a presentation that promotes high-risk use (general principles, paragraph (12)). Reasonably foreseeable misuse is use not in the intended purpose that can result from reasonably foreseeable human behavior or interaction with other systems, for example other AI systems (Article 3(13)). For high-risk systems subject to Article 9, providers must complete risk "estimation and evaluation" for intended use and reasonably foreseeable misuse. Article 2(2) excludes Article 6(1) high-risk systems related to Annex I Section B products from Article 9. A third party can become the provider if its purpose change makes the system high-risk (Article 25(1)(c)).

See Intended purpose.

When does modifying an AI system make me its provider?

Article 25(1) can make a distributor, importer, deployer or other third party a high-risk system's provider in three cases:

• Own branding of an existing high-risk system, subject to the contractual-allocation qualification in point (a).
• Substantial modification of a high-risk system that remains high-risk, point (b).
• Changed purpose that makes a previously non-high-risk system high-risk, point (c). This includes general-purpose AI systems.

Each concerns a system "already placed on the market or put into service". Article 3(23) defines substantial modification. Article 43(4) governs reassessment and the exception for predetermined, documented learning changes.

See Substantial modification.

Is CV-screening AI high-risk under the EU AI Act?

Usually. Annex III(4)(a) covers recruitment and selection AI, with filtering of job applications, candidate evaluation and targeted job advertisements. Article 6(3)'s full exception conditions apply, with no exception for profiling natural persons.

The Commission draft guidelines on high-risk classification distinguish candidate scoring from job-description drafting based only on human-defined qualifications and skills. The latter can be a narrow procedural task, not a general recruitment exemption. Chapter III Sections 1–3 apply to Annex III systems from 2 December 2027.

See Annex III(4).

Is AI exam proctoring high-risk under the EU AI Act?

Usually. Annex III(3)(d) covers AI intended to monitor and find prohibited student behavior during tests in educational or vocational training institutions. Article 6(3)'s full exception conditions apply, with no exception for profiling natural persons.

The Commission draft guidelines on high-risk classification distinguish controlled testing from subsequent plagiarism checks (paragraphs (234)–(235)). Biometric emotion inference can independently cause a prohibition in Article 5(1)(f), with its medical or safety exception. Chapter III Sections 1–3 apply to Annex III systems from 2 December 2027.

See Annex III(3).

Is consumer credit scoring high-risk under the EU AI Act?

Usually. Annex III(5)(b) covers natural-person creditworthiness evaluation and credit scoring, subject to its financial-fraud exclusion and Article 6(3). Profiling prevents the Article 6(3) exception.

The Commission draft guidelines on high-risk classification distinguish exclusively non-essential services from mixed essential/non-essential uses (paragraphs (302)–(305)). It also distinguishes a primary fraud purpose from creditworthiness evaluation and prudential-only systems from mixed credit-scoring purposes (paragraphs (307)–(309), (323)). These are conditional examples. Chapter III Sections 1–3 apply to Annex III systems from 2 December 2027.

See Annex III(5).

Is fraud-detection AI in banking high-risk?

Not automatically. Annex III(5)(b) excludes systems intended for financial fraud detection from its creditworthiness use case.

In the Commission draft guidelines on high-risk classification, fraud detection must be the primary purpose for this exclusion, also if creditworthiness evaluation uses its output (paragraphs (307)–(308)). AML/CFT is not this exclusion: mixed credit-assessment purposes stay in scope (paragraph (309)). Annex III(5)(c) life and health insurance pricing has no related fraud exclusion. Other Annex III uses and Article 5(1)(d) can also apply.

See Annex III(5).

Is HR sentiment analysis or worker-management AI high-risk?

Yes, if its intended purpose is in Annex III(4)(b): decisions on work relationships, behavior- or trait-based task allocation, or monitoring and evaluation of workers.

These limits also apply:

• Biometric emotion inference: Article 5(1)(f)’s prohibition prevails, with a narrow medical/safety exception. The exception does not include general wellbeing or general monitoring of "stress levels" at the workplace (Commission guidelines on prohibited AI practices, paragraphs (256)–(259)).
• Written-text-only inference: not in that prohibition (paragraph (251)).
• Article 6(3): all exception conditions apply, with no exception for profiling natural persons.
• Neutral task allocation: the exclusion in the Commission draft guidelines on high-risk classification requires job-related external criteria alone, without personal characteristics, nationality/ethnicity proxies or criteria covered by EU non-discrimination law (Annex III draft, paragraph (272)).

See Annex III(4).

Is AI essay scoring for university admission high-risk?

Usually. Admission AI falls in Annex III(3)(a), subject to Article 6(3)'s full exception conditions and profiling override.

Learning-outcome evaluation is a different use case, point 3(b). In the Commission draft guidelines on high-risk classification, purely formative feedback is not in point 3(b) only where it does not contribute to summative evaluation or "final decisions affecting the educational pathways" or professional pathways (paragraphs (222)–(224)). Function and intended purpose control, not the system name. Chapter III Sections 1–3 apply to Annex III systems from 2 December 2027.

See Annex III(3).

Is AI translation in an asylum interview high-risk?

It can be. Annex III(7)(c) covers specified help to competent authorities examining asylum, visa or residence applications, where Union or national law permits the use.

The Commission draft guidelines on high-risk classification apply Article 6(3)(d)'s preparatory-task exception for transcription, translation or neutral summaries without advice on next steps or outcomes (section 3.7.4). Added facts, credibility judgments or links to the applicable law do not qualify for that example. All Article 6(3) conditions stay, with no material influence and no profiling.

See Annex III(7).

Are the Commission's Article 50 transparency guidelines adopted and binding?

The Commission approved the content on 20 July 2026, but the guidelines are not binding law. Communication C(2026) 5054 final makes applicability conditional on subsequent formal adoption when all language versions are available.

The 20 July library page calls them adopted, while the 31 July release IP/26/1714 calls them draft. These descriptions do not resolve the condition in the Communication that approved the content. Article 50 itself applies from 2 August 2026, subject to Article 111(4)'s transitional period for Article 50(2).

See Commission guidance.

When must a chatbot tell me I am interacting with AI?

Providers must make sure each person is informed that they interact directly with AI, "at the latest at the time of the first interaction" (Articles 50(1) and 50(5)).

The Article 50(1) rules include:

• The conditional obviousness exception, interpreted restrictively with consideration of vulnerable audiences (Commission Article 50 Guidelines, paragraph (45)).
• In-context disclosure, not terms alone or vague names (paragraph (38)).
• AI-agent disclosure of artificial nature and the person represented (paragraph (31)).
• The criminal-law-enforcement exception for systems "authorised by law", with safeguards and no exception for public crime-reporting systems (Article 50(1)).
• Recommended disclosure at each session’s start (paragraph (143)).

The Section B exclusion applies on Article 2(2)’s conditions. Article 50(1) applies from 2 August 2026 without a transitional period.

What AI-generated content must be marked under Article 50(2)?

Article 50(2) applies to providers of systems that generate synthetic audio, image, video or text. Providers must make sure that outputs are marked in a machine-readable format and detectable as artificially generated or manipulated.

Article 50(2) has exceptions for standard editing help, changes that do not substantially change input data or its semantics, and specified criminal-law-enforcement purposes authorized by law. Article 2(2)'s Section B exclusion also applies.

The Commission Article 50 Guidelines give more limits with conditions:

• Not in the scope: short outputs such as single words or image captions, source code, and outputs that go only from machine to machine, with no exposure to persons (paragraph (68)).
• Not in the scope: outputs used only in closed-loop industrial and product-development workflows. The final output of the workflow must be marked and detectable (paragraph (68)).
• Less robust metadata can be sufficient for physical products that keep their outputs in a technically controlled and closed environment, for example vehicle navigation (paragraph (86)).
• No marking and detection is necessary for some industrial or business-to-business outputs that are "strictly technical in nature" and are for "a limited pre-defined number of natural persons acting in a professional capacity", only if all the conditions of paragraph (87) are met.
• Real-time content that is ephemeral and is not recorded, stored or disseminated "may also be exempted" when marking is not technically feasible and "the persons exposed to the content are made aware that the content is AI-generated or manipulated" (paragraph (88)).

For systems placed on the market before 2 August 2026, providers must comply with Article 50(2) by 2 December 2026 (Article 111(4)).

See Article 50(2)'s conditions and Application dates.

Does the EU AI Act require a specific watermarking technology?

No. Article 50(2) requires machine-readable marking and detection, without naming one technology.

Providers' solutions must satisfy the effectiveness, interoperability, robustness and reliability requirements, subject to technical feasibility, content limitations, costs and the state of the art. Using a model provider's or third party's solution does not remove the provider's responsibility (Commission Article 50 Guidelines, paragraph (74)). The voluntary Code of Practice is one compliance method, not a harmonized standard or the only method.

See Article 50(2).

What is a deepfake, and does the artistic exception remove labeling?

A deep fake is AI-generated or manipulated image, audio or video resembling existing persons, objects, places, entities or events and falsely appearing "authentic or truthful" (Article 3(60)). Deployer disclosure is also necessary for artistic works (Article 50(4)).

The Article 50(4) qualifications include:

• Evidently artistic or analogous works: disclosure "in an appropriate manner" without hampering display or enjoyment.
• Mixed informative/creative content: the informative character prevails (Commission Article 50 Guidelines, paragraph (122)).
• Human-perceivable disclosure: machine-readable marks alone are not sufficient (paragraph (117)).
• Exceptions for criminal law enforcement "authorised by law" and for natural-person deployers using AI systems in purely personal non-professional activities (Articles 50(4) and 2(10)).
• No retroactive labeling requirement for deep fakes generated or manipulated before 2 August 2026 (paragraph (154)).

See Article 50(4).

When is AI-generated public-interest text exempt from labeling?

Article 50(4) exempts AI-generated or manipulated public-interest text after human review or editorial control only if a person holds editorial responsibility for publication.

The duty depends on publication to give information to the public, not on using AI to publish. Superficial checks are not sufficient, and substantive AI changes after human review defeat the exception (Commission Article 50 Guidelines, paragraphs (135)–(136)). Criminal law enforcement "authorised by law" has a different exception. The historical-content rule requires generation or manipulation together with publication before 2 August 2026 (paragraph (154)).

See Article 50(4).

Who enforces Article 50, and what are the EU AI Act fines?

National market surveillance authorities usually enforce Article 50, with AI Office competence subject to Article 75(1)’s system exceptions and deployer limits. The EDPS enforces it for Union institutions, bodies, offices and agencies.

The Article 50 fine limits are:

• EUR 15 000 000, or for undertakings 3% of total worldwide annual turnover for the preceding financial year, "whichever is higher" (Article 99(4)(g)).
• The lower maximum for SMEs, which include start-ups, and small mid-cap enterprises (Article 99(6)/(6a)).
• EUR 750 000 for Union institutions, bodies, offices and agencies (Article 100(3)).

Article 5 and GPAI fines are different. Some systems have more time for Article 50(2) (Article 111(4)).

Do systems released before August 2026 have an Article 50 grace period?

Only for Article 50(2). Article 111(4) of the EU AI Act applies to providers of AI systems that generate synthetic content. It applies to the systems that they placed on the market before 2 August 2026. These providers must comply with Article 50(2) by 2 December 2026.

• The transitional period applies to all of Article 50(2): marking and detection. It applies also to interactive systems that generate content.
• Article 50(1) has no transitional period. It applies from 2 August 2026.
• Paragraph (154) of the Article 50 guidelines gives different rules for content made before 2 August 2026. These rules include text on matters of public interest that is published on or after that date.

See Application dates.

Is my model a general-purpose AI model under the EU AI Act?

Article 3(63) requires "significant generality" and competence across "a wide range of distinct tasks". The model must also be one that "can be integrated into a variety of downstream systems or applications". The pre-market research, development or prototyping exclusion applies.

The Commission's indicative test is more than 10^23 FLOP plus language, text-to-image or text-to-video generation (GPAI-scope Guidelines, paragraph (17)). Exceptionally, a model that meets the criterion is not a GPAI model if it lacks "significant generality" or cannot competently do "a wide range of distinct tasks". Exceptionally, a model below the criterion is a GPAI model if it has the two properties (paragraph (20)). The full Article 3(63) definition also governs, "regardless of the way the model is placed on the market". The 10^25 FLOP systemic-risk presumption is different.

See The GPAI model test.

Does fine-tuning someone else's GPAI model make me a provider?

A different actor not acting for the original provider can become a GPAI provider through "a significant change" to generality, capabilities or systemic risk (Commission GPAI-scope Guidelines, paragraphs (57)–(59)).

Modification compute above one-third of training compute of the original model is indicative, not a safe harbor below it (paragraph (60)). The fallback applies only where the modifier cannot reasonably know the original model's compute and cannot estimate it: one-third of 10^25 FLOP for systemic-risk models, for other models one-third of 10^23 (paragraph (61)). Development by or for the original provider follows a different lifecycle rule.

See When a model modifier becomes a provider.

Do safety measures remove a GPAI model's systemic-risk classification?

No. The Commission rejects implemented or planned mitigations as grounds for an exclusion of a model that has high-impact capabilities but allegedly no systemic risks (GPAI-scope Guidelines, paragraph (39)).

A provider can give "sufficiently substantiated arguments" based on "specific characteristics" of the model with its Article 52 notification. The Commission decides (Article 52(2)–(3)). An argument that the model has no high-impact capabilities is different. Mitigations stay applicable to Article 55(1) duties.

See Classification of systemic risk.

Does updating my own GPAI model create a new model?

The Commission treats development by or for the original provider after the large pre-training run as the same model's lifecycle (GPAI-scope Guidelines, paragraph (22)).

Applicable duties continue: updated documentation, copyright policy, public training summary and its required updates, and continuous systemic-risk "assessment" and mitigation (paragraph (23), Articles 53 and 55). Changed compute or capabilities can cause Article 52 notification and other duties. A different downstream modifier follows a different provider-status test.

See Model lifecycle.

What training compute counts for GPAI thresholds, including synthetic data?

The GPAI-scope Guidelines use parameter-update compute for models without systemic risk. They use cumulative training compute for models with systemic risk and when assessing if a model has systemic risk (paragraph (115)). The same distinction governs modification compute (paragraph (60)).

Cumulative compute includes lifecycle runs that enhance model capabilities and non-public synthetic-data generation, with discarded candidates used to select training data (paragraphs (116)–(118)). Paragraph (119) gives conditional exclusions for public synthetic data, diagnostics that do not enhance capabilities, human-learning-only experiments, parent/auxiliary-model training and activation recomputation.

Providers can use one of the two illustrated estimation methods or a different method. In their best judgment, the estimate must be accurate "within an overall error margin of 30% of the reported estimate". The Commission expects documented assumptions and uncertainties (paragraph (120)).

See Training compute.

How do autonomous agents affect GPAI systemic-risk classification?

The "level of autonomy" of a GPAI model and "the tools it has access to" are criteria in Annex XIII, point (e), of the EU AI Act. The Commission uses the Annex XIII criteria when it designates a GPAI model with systemic risk (Article 51(1)(b)).

• These criteria are for a decision of the Commission. They do not automatically cause the classification.
• The GPAI-scope Guidelines do not give guidance on AI agents.
• The rules for GPAI models and the rules for AI systems are different. The rules for GPAI models apply also when the model is part of an AI system (recital 97).

See Classification of systemic risk.

Are GPAI obligations enforceable now under the EU AI Act?

Yes. Chapter V applies from 2 August 2025, and Commission enforcement powers from 2 August 2026. Providers of models placed on the market before 2 August 2025 have until 2 August 2027 (Article 111(3)).

Commission information requests, measures and fines are different from AI Office evaluations after Board consultation (Articles 91–93, 101). The Article 101 maximum is EUR 15 million or 3% of annual worldwide turnover in the preceding financial year, "whichever is higher". The first year without enforcement did not stop applicable obligations (Commission GPAI-scope Guidelines, paragraph (112)).

See Enforcement and penalties.

Are free and open-source GPAI models exempt from the EU AI Act?

Only in part. Article 53(2) exempts qualifying free and open-source models from Article 53(1)(a)/(b) documentation duties. It requires license permissions and public parameters, weights, architecture and information on model use. Systemic-risk models cannot use it. Copyright policy and training-summary duties continue.

In the Commission GPAI-scope Guidelines, monetization prevents the exceptions. Required payments for model "access, use, modification, and distribution" count as monetization (paragraphs (82)–(83)). Required personal-data processing for model access, use, modification or redistribution also counts, with an exception for processing "exclusively and strictly limited to the security of the model" without "commercial or financial gain" (paragraph (84)). Purely optional paid services do not count as monetization if model usability, free use and free and open access continue (paragraph (85)). Transactions between microenterprises also do not count as monetization (paragraph (86)). Also check the AI-system duties.

See Free and open-source models.

When must a GPAI provider notify the Commission about systemic risk?

Article 52(1) requires notification "without delay and in any event within two weeks" after Article 51(1)(a)'s condition is met or becomes known to be forthcoming. Article 51(2) presumes high-impact capabilities above 10^25 FLOP.

The duty can occur before training ends when the presumption condition is reasonably foreseeable (Commission GPAI-scope Guidelines, paragraph (30)). Rebuttal arguments accompany rather than replace notification and do not stop obligations (paragraph (40)). Commission designation is a different method in Articles 51(1)(b) and 52(4).

See Classification of systemic risk.

Do Articles 4, 5 and 50 apply to Annex I Section B AI systems?

No, for high-risk Article 6(1) systems related to Annex I Section B products. Article 2(2) limits applicable provisions to Article 6(1), Article 60a and Articles 102–112. Articles 4, 5 and 50 do not apply.

Articles 57–59 apply only insofar as high-risk requirements are integrated into the Section B legislation. Mere use in a Section B industry does not satisfy the Article 6(1) product conditions.

See Annex I Section B scope.