Appearance
Commission Draft Worked Examples — High-Risk AI Classification
This page gives the examples in the Commission draft guidelines for the classification of high-risk AI systems in Article 6 of the EU AI Act. The AI Act is Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744. There is one section for each area of Annex III and one section for Annex I. For the classification rules, see High-risk classification.
How to use the examples
The Commission draft guidelines of 19 May 2026 give non-binding guidance on high-risk classification in Article 6 of the EU AI Act. In the draft, the examples are "not to be considered as exhaustive and may be updated over time" (general principles, paragraph (3)).
In the draft, a system is in an Annex III use case if its intended purpose includes it (Annex III part, paragraphs (70) and (77)). The intended purpose comes from the information of the provider (general principles, paragraphs (10)–(12)). Thus, compare the intended purpose of your AI system with each example, not only the name of the system. In the draft, "human involvement cannot change the purpose and area in which a system is intended to be used" (Annex III part, paragraph (70)).
The draft gives this rule for AI systems that are parts of a large AI system: "Where several AI systems form part of a more complex AI system, so that their combined intended purpose or joint outputs materially influence an individual decision, the combined configuration is treated as a single AI system for the purpose of high-risk classification" (Annex III part, paragraph (75)).
For each Annex III use case, the draft gives examples in three groups (Annex III part, paragraph (118)):
- In the use case. The AI system is high-risk in Article 6(2), unless the Article 6(3) exception applies.
- Not in the use case. The AI system is not high-risk in that point. A different Annex III point, Annex I or other Union law can apply to it.
- In the use case, with the Article 6(3) exception. The provider can find that the AI system is not high-risk, if all the conditions of Article 6(3) apply.
No result in these three groups shows that the use of an AI system is permitted by law. In the draft, the use must also comply with other applicable law (general principles, paragraph (3), and Annex III part, paragraph (82)).
The Annex III exception applies when one of the Article 6(3)(a)–(d) conditions is met without profiling natural persons. The provider must document this decision and register the system "before that system is placed on the market or put into service" (Articles 6(4) and 49(2)).
The examples apply only in the text of their Annex III point. Each section below gives the persons, the organizations and the conditions in its point. In points 1, 6 and 7, the use cases apply only "in so far as their use is permitted under relevant Union or national law". Thus, a high-risk classification in these points does not show that the use of the system is permitted.
For the dates from which the high-risk rules apply, see Application dates.
Annex III(1) — Biometrics (section 3.1)
Annex III point 1 of the EU AI Act gives three use cases for biometrics, "in so far as their use is permitted under relevant Union or national law":
- (a) Remote biometric identification systems. Point 1(a) does not include AI systems for "biometric verification the sole purpose of which is to confirm that a specific natural person is the person he or she claims to be".
- (b) AI systems for "biometric categorisation, according to sensitive or protected attributes or characteristics based on the inference of those attributes or characteristics".
- (c) AI systems for emotion recognition.
The Commission draft tells providers to examine the prohibitions in Article 5(1)(f), (g) and (h) and Article 5(2) to (5) first. If all the conditions of a prohibition apply, the AI system is prohibited. Then it is not necessary to examine its classification (Annex III part, paragraph (126)). In the draft, "Most AI systems that fall under an exception from the prohibitions provided in Article 5 AI Act will qualify as high-risk" (paragraph (126)). Each of the three prohibitions has different conditions, and more than one prohibition can apply to a biometric system. See Conditions for Article 5.
Remote biometric identification
Article 5(1)(h) prohibits "the use of ‘real-time’ remote biometric identification systems in publicly accessible spaces for the purposes of law enforcement, unless and in so far as such use is strictly necessary for one of the following objectives". Article 5(1)(h) gives three objectives, and Article 5(2) to (5) give more conditions for that use.
In the Commission draft, point 1(a) includes these uses not covered by the Article 5(1)(h) prohibition (paragraph (126) and section 3.1.2):
- use for one of the objectives of Article 5(1)(h),
- use that is not for law enforcement,
- use that is not real-time,
- use that is not in publicly accessible spaces.
For the prohibition, see Article 5(1)(h).
Biometric categorization
For the biometric-categorization prohibition and its exclusions, see Article 5(1)(g).
In the draft, point 1(b) includes biometric categorization not covered by the Article 5(1)(g) prohibition. The system must also infer a "sensitive or protected" attribute (paragraph (126)):
- In the draft, "sensitive or protected" attributes are "those protected under Article 9(1) GDPR", with Recital 54 of the AI Act. Thus, "the attributes and characteristics of ethnic origin, genetic data, and data concerning health are here of key relevance" (paragraph (169)).
- Biometric data must be the input, and the attribute must be the output that the system infers (paragraph (170)).
- In the draft, "attributes such as age and gender are not considered as sensitive or protected under Article 9(1) GDPR" (section 3.1.3).
Emotion recognition
Article 5(1)(f) prohibits AI systems "to infer emotions of a natural person in the areas of workplace and education institutions, except where the use of the AI system is intended to be put in place or into the market for medical or safety reasons". See Article 5(1)(f).
In the Commission draft, emotion recognition not covered by the Article 5(1)(f) prohibition is in point 1(c). It is high-risk, unless the Article 6(3) exception applies (Annex III part, paragraph (126)). This includes:
- emotion recognition in areas other than the workplace and education institutions,
- emotion recognition in the workplace or in education institutions "for medical or safety reasons".
Point 1(c) applies only to emotion recognition systems in Article 3(39), which identify or infer "emotions or intentions of natural persons on the basis of their biometric data". In the draft, a system that infers emotions from "typed text" does not use biometric data (paragraph (124)). Emotions do not include "physical states, such as pain or fatigue" (paragraph (177)).
The deployer of an emotion recognition system or a biometric categorization system must also "inform the natural persons exposed thereto of the operation of the system" (Article 50(3)). Article 50(3) has an exception for systems "permitted by law to detect, prevent or investigate criminal offences". See Article 50(3).
Examples for point 1
| Example in the draft | Result in the draft | Draft section |
|---|---|---|
| CCTV cameras in a stadium. After an incident, the system compares the facial images with a biometric database "to identify offenders". | In point 1(a) | 3.1.2(a) |
| A system identifies, from a CCTV image on "private property", "an individual suspected of carrying out a burglary". | In point 1(a). Point 1(a) is "not limited to RBI in publicly accessible spaces". | 3.1.2(a) |
| Facial recognition at the entry to a stadium. Each person must go in front of the system to get access. | Not in point 1(a). The identification is not remote. | 3.1.2(b) |
| Biometric verification, for example to "unlock a smartphone" or for "logging in to a bank account online". | Not in point 1(a) | 3.1.2(b) |
| A system infers the health data of patients from their gait, and puts the patients in categories of disease. | In point 1(b) | 3.1.3(a) |
| A system infers the age or the gender of persons from facial images or keystrokes. | Not in point 1(b). In the draft, age and gender are not "sensitive or protected". | 3.1.3(b) |
| A system infers the emotions of customers from their voices in a call center. | In point 1(c) | 3.1.4(a) |
| A system in a vehicle finds when a driver "loses focus", for example because of fatigue. | Not in point 1(c). Emotions do not include fatigue. | 3.1.4(b) |
Annex III(2) — Critical infrastructure (section 3.2)
Annex III point 2 of the EU AI Act includes AI systems "intended to be used as safety components in the management and operation of critical digital infrastructure, road traffic, or in the supply of water, gas, heating or electricity".
The Commission draft gives these conditions for point 2 (Annex III part, section 3.2.1):
- The AI system must be a safety component in Article 3(14). The draft examines "whether the AI system directly protects the physical integrity of the infrastructure" (paragraph (180)).
- A system that is "merely supportive, informational, organisational or optimisation-oriented" is not a safety component (paragraph (181)).
- A component "used solely for cybersecurity purposes" is not a safety component in point 2 (paragraph (187)).
- The system is high-risk only if "used by an entity identified as a critical entity by a Member State under the CER Directive" (paragraph (190)). The CER Directive is Directive (EU) 2022/2557.
An AI system that is not in point 2 can be high-risk in Article 6(1), if its conditions apply (paragraph (179)). In the draft, point 2 excludes AI safety components in nuclear-specific elements covered by treaties and Union law on "nuclear power". AI safety components in "electricity components of nuclear power plants" can be in point 2, subject to its conditions (paragraphs (205)–(206)).
| Example in the draft | Result in the draft | Draft section |
|---|---|---|
| A system that monitors road traffic and adjusts the traffic lights. | In point 2 | 3.2.3(a) |
| A traffic data platform for "flow optimisation functions" that does not directly change the traffic management. | Not in point 2 | 3.2.3(b) |
| A system "used for the detection of anomalies in data patterns when operating electricity grids for the purpose of monitoring and supporting decision-making in relation to critical functions". | In point 2 | 3.2.7(a) |
| A system "used for electricity grid optimization by forecasting the energy demand". "The core safety functions are handled separately". | Not in point 2 | 3.2.7(b) |
| A system that monitors the network traffic of an electricity grid "solely for cybersecurity purposes". | Not in point 2 | 3.2.7(b) |
Annex III(3) — Education and vocational training (section 3.3)
Annex III point 3 of the EU AI Act gives four use cases for educational and vocational training institutions at all levels. The use cases are AI systems intended:
- (a) "to determine access or admission or to assign natural persons to educational and vocational training institutions at all levels",
- (b) "to evaluate learning outcomes, including when those outcomes are used to steer the learning process of natural persons",
- (c) "for the purpose of assessing the appropriate level of education that an individual will receive or will be able to access",
- (d) "for monitoring and detecting prohibited behaviour of students during tests".
The Commission draft gives this guidance for point 3:
- Summative evaluation and formative evaluation. In the draft, point 3(b) includes only systems that make "an evaluation towards a final decision affecting educational and professional pathways based on a summative assessment" (paragraph (223)). Systems that give "feedback and learning support as part of the ongoing educational process" are not in point 3(b) (paragraph (224)). Thus, a system for formative evaluation only, with no effect on the summative evaluation or on a "final decision", is not in point 3(b) (section 3.3.3).
- Tests. Point 3(d) applies to behavior "during a process leading to an evaluation or a grade of a student", in a "controlled testing situation" (paragraph (234)). In the draft, "checking homework, essays, or other assignments for plagiarism" is not in point 3(d) (paragraph (235)).
- Profiling. Systems in Annex III point 3 that do profiling of natural persons are always high-risk (Article 6(3), last subparagraph). In the draft, point 3 systems "will often perform profiling of natural persons" (paragraph (215)).
- Emotion recognition. Article 5(1)(f) prohibits AI systems that infer the emotions of a natural person in education institutions, unless the system is "for medical or safety reasons". In the draft, emotion recognition in education can be high-risk in point 3 or 1(c) unless Article 5(1)(f) applies (paragraph (216)).
| Example in the draft | Result in the draft | Draft section |
|---|---|---|
| An admission system examines "student applications, transcripts, and test scores to determine eligibility for admission". | In point 3(a). Article 6(3)(d) can apply, if the system does only a preparatory task and does not do profiling. | 3.3.2(a) |
| A system puts students in "public schools" from their location and "other personal characteristics". | In point 3(a). The system does profiling. Thus, Article 6(3) does not apply. | 3.3.2(a) |
| A system that can "classify each application into one of several predefined categories" and "identify and flag duplicate applications". It makes no decision on admission. | In point 3(a). Article 6(3)(a) can apply. | 3.3.2(c) |
| A system examines "tests, quizzes, and exams which count towards a final evaluation" and gives grades. | In point 3(b) | 3.3.3(a) |
| "Language learning software applications" that students use "on their own volition", and that do not "lead to the attainment of a credential". | Not in point 3(b) | 3.3.3(b) |
| "AI-powered adaptive placement tools" give educators recommendations for "beginner, intermediate, or advanced classes". | In point 3(c) | 3.3.4(a) |
| A proctoring system monitors persons during a "certification exam" "by using a combination of facial recognition, keystroke analysis, and screen monitoring". | In point 3(d) | 3.3.5(a) |
| A proctor sees "suspicious behaviour" during a test and uses a system to examine it. The system gives a report to the proctor. | In point 3(d). Article 6(3)(b) can apply. | 3.3.5(c) |
Annex III(4) — Employment and worker management (section 3.4)
Annex III point 4 of the EU AI Act gives two use cases:
- (a) AI systems intended "for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates".
- (b) AI systems intended "to make decisions affecting terms of work-related relationships, the promotion or termination of work-related contractual relationships, to allocate tasks based on individual behaviour or personal traits or characteristics or to monitor and evaluate the performance and behaviour of persons in such relationships".
The Commission draft gives this guidance for point 4:
- Prohibitions first. In the draft, if a system is in an Article 5 prohibition and also in point 4, "the prohibition will prevail". The draft names the Article 5(1)(f) prohibition of emotion recognition in the workplace (paragraph (242)).
- Recruitment. For point 4(a), the system must "relate to or impact the substance of the recruitment process" (paragraph (245)).
- Targeted job advertisements. A system for targeted job advertisements that "relies on profiling" is always high-risk. In the draft, a system without profiling is also high-risk "if it meaningfully conditions access to employment opportunities or can lead to a risk of discrimination" (paragraph (250)).
- Tasks. In the draft, systems that give tasks to workers and "rely only on objective, neutral, and external criteria tied directly to the requirements of a job" are not in point 4(b). An example is "allocating tasks based on availability as indicated by workers in scheduling tools" (paragraph (272)).
The Commission guidelines on prohibited practices limit Article 5(1)(f) to "inferences based on a person’s biometric data" (paragraph (245)). In those guidelines, an AI system "inferring emotions from written text" is not in the prohibition (paragraph (251)). See Article 5(1)(f).
Job descriptions
The Commission draft gives an example of a system that generates job descriptions (Annex III part, section 3.4.2, after paragraph (247)). The system uses "a list of tasks to be carried out and a set of necessary qualifications and skills previously defined by a human recruiter". From these, the system generates only the job description. A system of this type can be in point 4(a), "if it is within their intended purpose to play a role in the recruitment process in a manner that affects applicants’ prospects". In the draft, this example is a narrow procedural task in Article 6(3)(a). Thus, the system is not high-risk, if all the conditions of Article 6(3) apply, and the system does not do profiling.
The result is different if the system "itself generates the necessary qualifications and skills based on a high-level description of the job position". It is also different if the system has a function for "evaluating the candidates’ CVs against the job description that it created and to make recommendations of suitability". For these two examples, the draft tells providers that Article 6(3) does not apply, and that the system is high-risk.
Examples for point 4
| Example in the draft | Result in the draft | Draft section |
|---|---|---|
| An "automated job matching and ranking tool" compares CV data with job descriptions. It gives "quantitative scores, rankings" that recruiters use "to shortlist candidates". | In point 4(a). Article 6(3) does not apply. | 3.4.2(a) |
| A system to "place targeted job advertisements to specific users on social media". It analyzes navigation patterns and personal characteristics to select audiences. It does more than select audiences by professional accreditations or geographical location. It can materially change which candidates learn about vacancies. | In point 4(a). Article 6(3) does not apply. | 3.4.2(a) |
| A system to "perform background checks on job applicants" that gives risk scores. | In point 4(a). The system does profiling. Thus, Article 6(3) does not apply. | 3.4.2(a) |
| A system that only finds "non-inclusive or discriminatory wording in ad descriptions". | Not in point 4(a) | 3.4.2(b) |
| A system for "scheduling interviews" with candidates. | In point 4(a). Article 6(3)(a) can apply. | 3.4.2(c) |
| An "AI-enabled scheduler to assign shifts" uses "punctuality, non-show history, acceptance/decline rates for offered shifts, customer/manager ratings". The result is "allocating the highest performing workers to the most important shifts". | In point 4(b). Article 6(3) does not apply. | 3.4.3(a) |
| An "online tutoring platform" uses a "tutor performance score". If the score stays low, the system "automatically suspends the tutor’s account for a month", and then deactivates it. | In point 4(b). In the draft, this is a "functional termination of the work relationship". | 3.4.3(a) |
| "A consultancy firm uses an AI-enabled writing assistant to refine managers’ promotion reports after the evaluations are fully completed". | In point 4(b). Article 6(3)(b) can apply. | 3.4.3(c) |
Annex III(5) — Essential services (section 3.5)
Annex III point 5 of the EU AI Act gives four use cases. The use cases are AI systems intended:
- (a) "to be used by public authorities or on behalf of public authorities to evaluate the eligibility of natural persons for essential public assistance benefits and services, including healthcare services, as well as to grant, reduce, revoke, or reclaim such benefits and services",
- (b) "to be used to evaluate the creditworthiness of natural persons or establish their credit score, with the exception of AI systems used for the purpose of detecting financial fraud",
- (c) "to be used for risk assessment and pricing in relation to natural persons in the case of life and health insurance",
- (d) "to evaluate and classify emergency calls by natural persons or to be used to dispatch, or to establish priority in the dispatching of, emergency first response services, including by police, firefighters and medical aid, as well as of emergency healthcare patient triage systems".
Each point applies to different persons and organizations, for a different intended purpose. For example, point 5(a) applies only to use by or on behalf of public authorities. Examine each point and its Article 6(3) conditions independently.
Creditworthiness and credit scores
The Commission draft gives this guidance for point 5(b) (Annex III part, section 3.5.3):
- Essential services. In the draft, point 5(b) is for essential private services, for example consumer credit and mortgages (paragraph (302)). Some financial services are not essential in the draft, for example "special loans, such as leisure/travel loans" and "access to margin trading" (paragraph (304)). A provider that wants its system to be out of point 5(b) must show that the system is "solely intended to be used for non-essential private services". A system for essential services and for other services is high-risk (paragraph (305)).
- Fraud. For the fraud exception, "fraud detection must be the main intended use of the AI system, preceding all other purposes" (paragraph (307)). Such a system is not in point 5(b), "even if its output may be used for the creditworthiness assessment or establishing the credit score" (paragraph (308)).
- AML/CFT. Systems "for anti-money laundering and for countering the financing of terrorism" are not in the fraud exception. They are not in point 5(b) if their intended use does not include the "assessment of creditworthiness". They are in point 5(b) if they are "functionally linked and simultaneously intended to be used for the evaluation of creditworthiness or to establish a credit-score" (paragraph (309)).
- Pricing. A system intended only for pricing is not in point 5(b). A system that combines "assessing the creditworthiness of natural persons and the credit pricing" in one "integrated process" is in point 5(b) (paragraph (300)).
- Profiling. A system in point 5(b) that does profiling of natural persons is always high-risk (last subparagraph of Article 6(3)).
The fraud exception in point 5(b) applies only to point 5(b). It is not an exception from the Article 5 prohibitions. See Article 5(1)(d).
Life and health insurance
Point 5(c) has no fraud exception. In the draft, a system for risk assessment or pricing in life or health insurance is high-risk "even if it also has a feature that can be used for fraud detection". Only a fraud-detection "feature" that is "a distinct system in its own right" is not in point 5(c). The life or health insurance system for risk assessment or pricing remains high-risk (Annex III part, paragraph (321)). In the draft, "car or motor insurances, home insurance" are not in point 5(c) (paragraph (314)).
Prudential requirements
Section 3.5.5 of the Annex III part examines Article 144 of Regulation (EU) No 575/2013 (Capital Requirements Regulation) and Article 120 of Directive 2009/138/EC (Solvency II Directive):
- A system for a "creditworthiness assessment" or a credit score stays high-risk, if it is also used to calculate "capital requirements" "for prudential purposes" (paragraphs (322) and (324)).
- A system used to "calculate solely the risk-weighted exposure of portfolios of loans to natural persons under the internal ratings-based approach" is not in point 5(b), if it is not for credit scores or creditworthiness (paragraph (323)).
- If a financial institution uses two different systems, only the system for credit scores is high-risk (paragraph (325)).
- It is a "case-by-case assessment" if there is one AI system or two AI systems (paragraph (326)).
Examples for point 5
| Example in the draft | Result in the draft | Draft section |
|---|---|---|
| A system "intended to decide on the grant of unemployment benefits in individual cases" makes a recommendation to a case handler. | In point 5(a) | 3.5.2(a) |
| A system finds "fraud indicators" in applications for "essential public assistance benefits", and marks applications "for further investigation by a public authority". | In point 5(a) | 3.5.2(a) |
| A system intended "to translate applications for public assistance services made in a foreign language". | In point 5(a). Article 6(3)(a) and (d) can apply. | 3.5.2(c) |
| A system to "create a numerical representation of a natural person's creditworthiness, based on their payment behaviour and income", for "decision-making on a consumer credit or a mortgage". | In point 5(b) | 3.5.3(a) |
| A system puts customers in groups "to make personalised marketing offers", and the groups have no part in a "creditworthiness assessment". | Not in point 5(b) | 3.5.3(b) |
| An insurer uses a system that examines applications for life insurance with "mortality tables". The result "guides whether the insurer accepts the application and what conditions apply". | In point 5(c) | 3.5.4(a) |
| A system for "claims management in case of health insurance products". | Not in point 5(c) | 3.5.4(b) |
| A system with which "112 calls are classified to assign the level of urgency and to route responders". | In point 5(d) | 3.5.6(a) |
| A system "used in emergency departments that are intended to prioritise patients, without performing clinical assessment medical acts or diagnostic functions". | In point 5(d) | 3.5.6(a) |
| A system intended to "transcribe poor-quality emergency calls", but not to classify them. | Not in point 5(d) | 3.5.6(b) |
Annex III(6) — Law enforcement (section 3.6)
Annex III point 6 of the EU AI Act gives five use cases, "in so far as their use is permitted under relevant Union or national law". Each point is for AI systems intended for use by or on behalf of law enforcement authorities, or by "Union institutions, bodies, offices or agencies in support of law enforcement authorities". The use cases are AI systems intended:
- (a) "to assess the risk of a natural person becoming the victim of criminal offences",
- (b) "as polygraphs or similar tools",
- (c) "to evaluate the reliability of evidence in the course of the investigation or prosecution of criminal offences",
- (d) "for assessing the risk of a natural person offending or re-offending not solely on the basis of the profiling of natural persons as referred to in Article 3(4) of Directive (EU) 2016/680, or to assess personality traits and characteristics or past criminal behaviour of natural persons or groups",
- (e) "for the profiling of natural persons as referred to in Article 3(4) of Directive (EU) 2016/680 in the course of the detection, investigation or prosecution of criminal offences".
Examine the Article 5 prohibitions first, for example Article 5(1)(d) and Article 5(1)(h). Article 5(1)(d) does not apply to AI systems "used to support the human assessment of the involvement of a person in a criminal activity, which is already based on objective and verifiable facts directly linked to a criminal activity". See Article 5(1)(d). An AI system in point 6(d) that does profiling of natural persons is always high-risk, because Article 6(3) does not apply to it.
The Commission draft gives this guidance for point 6 (Annex III part, section 3.6.1):
- In the draft, point 6 does not include AI systems "specifically intended to be used for administrative activities". Examples are systems of "tax and customs authorities", or of financial intelligence units "carrying out administrative tasks" (paragraph (348)).
- In the draft, points 6(a), 6(d) and 6(e) "do not cover the analysis or mapping of locations or geographic areas". Such a system is not high-risk in these points if it does not do profiling of persons (paragraph (355)).
| Example in the draft | Result in the draft | Draft section |
|---|---|---|
| A system uses "victim statements, police reports, previous incidents, restraining order data, social services records and other data" to find the risk to a possible victim of domestic violence. | In point 6(a). The system does profiling. Thus, Article 6(3) does not apply. | 3.6.2(a) |
| A system that examines "facial micro-expressions to assess the credibility of answers during an interrogation". | In point 6(b). Article 6(3) does not apply. | 3.6.3(a) |
| A system that monitors "police officers’ fatigue or stress levels" for "occupational health and safety reasons". | Not in point 6(b) | 3.6.3(b) |
| "AI-enabled digital forensics solutions" that find if images, audio or video "have been altered or manipulated". | In point 6(c) | 3.6.4(a) |
| A system used to "map or reconstruct crime scenes" that does not examine the reliability of evidence. | Not in point 6(c) | 3.6.4(b) |
| Police use a system "during the initial processing of detained youth suspects" to "calculate a risk score that indicates the likelihood of reoffending". Other information is also used. | In point 6(d) | 3.6.5(a) |
| A system used to predict "locations and timeframes of likely crimes", "without linking the prediction to identifiable individuals". | Not in point 6(d) | 3.6.5(b) |
| A system that "collects and pre-sorts prior convictions, probation reports, and other records to prepare a case file for human officers". | In point 6(d). Article 6(3)(d) can apply. | 3.6.5(c) |
| A system that helps to identify "potential suspects based on descriptions of witnesses and other personal data". | In point 6(e) | 3.6.6(a) |
Annex III(7) — Migration, asylum and border (section 3.7)
Annex III point 7 of the EU AI Act gives four use cases, "in so far as their use is permitted under relevant Union or national law". Each point is for AI systems intended for use "by or on behalf of competent public authorities or by Union institutions, bodies, offices or agencies". The use cases are AI systems intended:
- (a) "as polygraphs or similar tools",
- (b) "to assess a risk, including a security risk, a risk of irregular migration, or a health risk, posed by a natural person who intends to enter or who has entered into the territory of a Member State",
- (c) "to assist competent public authorities for the examination of applications for asylum, visa or residence permits and for associated complaints with regard to the eligibility of the natural persons applying for a status, including related assessments of the reliability of evidence",
- (d) "in the context of migration, asylum or border control management, for the purpose of detecting, recognising or identifying natural persons, with the exception of the verification of travel documents".
The Commission draft gives this guidance for point 7:
- Search and rescue. In the draft, AI systems "intended to be used exclusively for search-and-rescue (lifesaving)" are not in point 7. A system with an intended use that includes border control is in point 7 because of that use. This applies only if the other conditions of point 7 apply (Annex III part, paragraph (381)).
- Maritime systems. The draft has no general exception for maritime systems. A system "used for maritime surveillance that detects and tracks persons for migration management or border-control operations" is in point 7(d). A system used "exclusively for safety of navigation or lifesaving" is not in point 7(d) (section 3.7.5).
- Polygraphs. If the use is "by or on behalf of law enforcement authorities", these systems are in point 6(b) (paragraph (388)).
- Profiling. In the draft, systems in point 7(b) that give "a per-person risk flag or risk score" do profiling. Thus, Article 6(3) does not apply to them (paragraph (395)).
- Border crossing points. In the draft, "a border crossing point is not considered a publicly accessible space". Thus, the Article 5(1)(h) prohibition does not apply there. In the draft, "the street leading to a border crossing point" is usually such a space (paragraph (387)). See Article 5(1)(h).
| Example in the draft | Result in the draft | Draft section |
|---|---|---|
| A system uses "a traveller’s personal history, travel patterns and watch-list matches" and gives a risk score. Border guards use the score to send the person to "second line checks". | In point 7(b). Article 6(3) does not apply. | 3.7.3(a) |
| A system that uses data about groups "to detect trends and patterns in migration", "where no natural persons are identified or identifiable from the aggregated data". | Not in point 7(b) | 3.7.3(b) |
| A system that examines "the authenticity and consistency of documents or the plausibility of personal narratives" and gives "credibility signals" to the examiner. | In point 7(c) | 3.7.4(a) |
| A system that finds "document morphing, non-conforming inks, copied signatures" as "a step in the examination procedure" for a visa, "a residence permit or international protection". | In point 7(c) | 3.7.4(a) |
| A system intended "to transcribe interviews, translate interview transcripts or produce neutral summaries for the file, without advising on next steps or suggesting outcomes". | In point 7(c). Article 6(3)(d) can apply. It does not apply to a system that does more, for example by "introducing elements not contained in the applicant’s statements, making credibility assessments, or linking the applicant’s statements to the applicable legal framework". | 3.7.4(c) |
| A system that uses "live facial recognition for identity checks at border-crossing points". | In point 7(d) | 3.7.5(a) |
| A system intended to "verify travel or identity documents". | Not in point 7(d) | 3.7.5(b) |
| A system to "detect persons in distress at sea exclusively to coordinate search-and-rescue operations". | Not in point 7(d) | 3.7.5(b) |
Annex III(8) — Administration of justice and democratic processes (section 3.8)
Annex III point 8 of the EU AI Act gives two use cases:
- (a) AI systems "intended to be used by a judicial authority or on their behalf to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts, or to be used in a similar way in alternative dispute resolution".
- (b) AI systems "intended to be used for influencing the outcome of an election or referendum or the voting behaviour of natural persons in the exercise of their vote in elections or referenda". Point 8(b) has this limit: "This does not include AI systems to the output of which natural persons are not directly exposed, such as tools used to organise, optimise or structure political campaigns from an administrative or logistical point of view."
Judicial authorities and alternative dispute resolution
The Commission draft gives this guidance for point 8(a) (Annex III part, section 3.8.1):
- In the draft, "AI systems intended to be used by parties and their legal representatives" are not in point 8(a), because "parties and their representatives are not acting ‘on behalf of a judicial authority’" (paragraph (416)).
- AI systems for "purely ancillary administrative activities", for example "anonymisation or pseudonymisation of judicial decisions", are not in point 8(a) (paragraph (419)).
- In alternative dispute resolution, a system is in point 8(a) only "where the outcome of the ADR proceedings produces legal effects for the parties" (paragraph (425)).
- A provider can use Article 6(3) for a system in point 8(a), unless the system does profiling (paragraph (420)).
| Example in the draft | Result in the draft | Draft section |
|---|---|---|
| A system that "analyses the factual circumstances of the case, submissions received, identifies applicable law and case law and then can generate drafts of judicial decisions". | In point 8(a) | 3.8.1(a) |
| A system that "analyses the presented facts of the case (or their summary), then assists judges in identifying relevant laws and legal precedents and suggests how they apply to the facts of a case". | In point 8(a). Article 6(3) does not apply. | 3.8.1(a) |
| A system that "simply performs AI-enabled searches of a database on the basis of the general area of law involved" or on keywords. | Not in point 8(a). In the draft, it is a "search engine". | 3.8.1(a) |
| A system for "verifying the payment of court fees", or for the anonymization of judicial decisions. | Not in point 8(a) | 3.8.1(b) |
| A case law database with "summaries of the published decisions" and "searching based on natural language (NLP)", "used by public and by court personnel". | In point 8(a). Article 6(3)(a) can apply. | 3.8.1(c) |
Elections
The Commission draft gives this guidance for point 8(b) (Annex III part, section 3.8.2):
- In the draft, the system must be "specifically intended to have an effect on the electorate’s choice or turnout". A system with "non-intentional influence", for example "general content recommender systems", is not in point 8(b) (paragraph (437)).
- In the draft, "natural persons must be directly exposed to the outputs of the AI system" (paragraph (436)). This "generally excludes AI systems used by political parties or candidates to increase back-office efficiency" (paragraph (445)).
- Use in a political campaign is not sufficient. The intended purpose of the system must be to have this effect (paragraphs (437) and (445)).
| Example in the draft | Result in the draft | Draft section |
|---|---|---|
| A system "intended to optimize the targeting and ad delivery of political advertising". | In point 8(b). If the system does profiling, Article 6(3) does not apply. | 3.8.2(a) |
| A chatbot "developed for use by political actors" that speaks with natural persons "to persuade them to support a candidate or policy". | In point 8(b) | 3.8.2(a) |
| "AI systems that optimise campaign staff (including volunteers) management", or systems for "scheduling rallies". | Not in point 8(b) | 3.8.2(b) |
| A system that "checks the tone of campaign texts already written by humans". | In point 8(b). Article 6(3)(b) can apply. | 3.8.2(c) |
Annex I sectoral examples
The Commission draft’s Annex I examples illustrate safety-component status, which establishes high-risk classification only when all Article 6(1) conditions of the EU AI Act apply.
Article 2(2) limits the provisions for Article 6(1) high-risk AI systems related to Annex I Section B products.
| Example in the draft | Annex I legislation | Result in the draft | Draft paragraph |
|---|---|---|---|
| A computer vision system that "detects human presence in a robot cell and triggers a safe stop or speed reduction where the intended purpose is to prevent injury". | Regulation (EU) 2023/1230, Section B point 21 | Safety component, because of its safety function | (45) |
| A system to "monitor gas concentrations and command shutdown". | Directive 2014/34/EU, Section A point 5 | Safety component, because of its safety function | (45) |
| "Lifts: an AI system managing door closing timing and obstacle detection". | Directive 2014/33/EU, Section A point 4 | Safety component, because its "failure" can cause injury | (46) |
| A system to "optimise combustion efficiency in household gas appliances". | Regulation (EU) 2016/426, Section A point 10 | Safety component, if its "failure" can cause "carbon monoxide formation, explosion or fire". A system that "merely optimises heating schedules based on household habits" is not a safety component if its "failure" causes only "discomfort or higher energy bills". | (43) |
| A system that "recommends music in a connected toy". | Directive 2009/48/EC, Section A point 2 | Not a safety component | (47) |
| A "smart thermostat covered by the Radio Equipment Directive", intended "to enhance user comfort and energy efficiency". | Directive 2014/53/EU, Section A point 6 | Not a safety component, if a "failure" causes only "discomfort or higher energy bills". It is a safety component if it has a safety function, or if a "failure" can "endanger health and safety". | (48)–(49) |
In Annex I, the entry for Regulation (EU) 2018/1139 is point 20 of Section B. It applies only "in so far as the design, production and placing on the market of aircrafts referred to in Article 2(1), points (a) and (b) thereof, where it concerns unmanned aircraft and their engines, propellers, parts and equipment to control them remotely, are concerned".
Use Modulos
The EU AI Act Frameworks in Modulos are OFF-1 at Organization level and MFF-1 at AI System level. The Requirement table and the sequence of work are on Operationalizing the EU AI Act in Modulos.
Sources
- Regulation (EU) 2024/1689 (AI Act), consolidated version of 27 July 2026: Articles 3(14), 3(39), 5, 6, 49(2) and 50(3), and Annexes I and III. EUR-Lex
- Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI), Official Journal of 24 July 2026: Article 1(41). EUR-Lex
- Commission draft of 19 May 2026, "Draft Commission guidelines on the classification of high-risk AI systems under Article 6 of Regulation (EU) 2024/1689 (AI Act) for stakeholder consultation":
- Draft Commission guidelines on the classification of high-risk AI systems, Commission page of 19 May 2026. Commission
- "Commission Guidelines on prohibited artificial intelligence practices established by Regulation (EU) 2024/1689 (AI Act)",
C(2025) 5052 finalof 29 July 2025: paragraphs (245) and (251). PDF
Disclaimer
This page gives general information. It is not legal advice.