Full List of High‑Risk AI Systems Under Annex III and the EU AI Act
High‑risk AI systems are those listed in Annex III, pursuant to Article 6(2). These systems face strict obligations because they can significantly affect safety, rights, or access to essential services.
Below is the full list, with citations to the official Annex III text.
- Biometrics
(Only where permitted under EU or national law)
1(a) Remote biometric identification systems
Excludes simple biometric verification used solely to confirm identity.
1(b) Biometric categorisation systems
Systems that infer sensitive or protected attributes, such as:
- Race
- Ethnicity
- Political opinions
- Religious beliefs
- Sexual orientation
1(c) Emotion‑recognition systems
Used in any context (except where separately prohibited, e.g., schools/workplaces).
- Critical Infrastructure
AI systems used as safety components in managing or operating:
- Critical digital infrastructure
- Road traffic
- Water supply
- Gas supply
- Heating
- Electricity
- Education and Vocational Training
3(a) Determining access or admission
AI used to decide who gets into:
- Schools
- Universities
- Training programs
3(b) Evaluating learning outcomes
Including systems that steer or adapt learning pathways.
3(c) Assessing appropriate education level
Placement or tracking decisions.
3(d) Monitoring/detecting prohibited behavior during tests
E.g., AI proctoring systems that detect cheating.
- Employment, Worker Management, and Access to Self‑Employment
4(a) Recruitment and selection systems
Including:
- Targeted job ads
- CV screening
- Video‑interview analysis
- Candidate scoring
4(b) Workplace decision‑making systems
AI used to:
- Allocate tasks
- Evaluate performance
- Monitor behavior
- Make promotion or termination decisions
- Access to Essential Private Services and Essential Public Services & Benefits
5(a) Eligibility for essential public services/benefits
Used by or on behalf of public authorities to:
- Grant
- Reduce
- Revoke
- Reclaim
Examples:
- Healthcare access
- Social benefits
- Public assistance programs
5(b) Creditworthiness and credit scoring
AI used to evaluate:
- Creditworthiness
- Credit scores
Exception:
AI used solely for detecting financial fraud is not high‑risk.
- Law Enforcement
(Not included in your search results above, but part of Annex III — included here for completeness.)
High‑risk systems include AI used for:
- Individual risk assessments
- Polygraph/“deception detection”
- Crime analytics
- Predictive policing (non‑prohibited forms)
- Evidence evaluation
- Profiling for operational purposes
- Migration, Asylum, and Border Control
AI used for:
- Assessing visa/asylum applications
- Border‑control risk assessments
- Authenticating travel documents
- Detecting irregular migration patterns
- Administration of Justice and Democratic Processes
AI used to:
- Assist judicial decision‑making
- Interpret facts or law
- Predict legal outcomes
- Support democratic processes (e.g., redistricting, voting‑system integrity)
How are high risk systems added to Annex III?
High‑risk AI categories in Annex III are not fixed forever. The EU AI Act creates a formal mechanism that allows the European Commission to add, remove, or modify high‑risk use cases over time. This mechanism is intentionally flexible so the law can keep up with technological change without needing a full legislative rewrite.
Here’s a clear, structured explanation of how high‑risk systems can be added to Annex III and what criteria must be met.
⭐ How High‑Risk AI Systems Are Added to Annex III
The EU AI Act gives the European Commission the power to update Annex III through delegated acts. This means the Commission can expand the list of high‑risk AI systems without going back to Parliament and the Council for a full legislative process.
This is similar to how the EU updates technical annexes in product‑safety or chemicals regulation.
🧩 The Legal Mechanism: Delegated Acts
The Commission may adopt a delegated act to add a new high‑risk category if:
- The AI system is intended to be used in a sensitive context, and
- Its use is likely to pose a significant risk to health, safety, or fundamental rights.
The Commission must consult:
- The AI Office
- The European Artificial Intelligence Board
- Relevant stakeholders (industry, civil society, academia)
- National market surveillance authorities
The delegated act is then subject to:
- Scrutiny by the European Parliament, and
- Scrutiny by the Council
Either institution can veto the change, but they cannot rewrite it.
🎯 The Criteria for Adding a New High‑Risk System
The Commission must evaluate whether the AI system:
- Has a significant impact on fundamental rights
Examples:
- Privacy
- Non‑discrimination
- Access to services
- Freedom of expression
- Due process
- Has a significant impact on health or safety
Examples:
- Physical safety
- Psychological harm
- Systemic risks
- Is used in a context where individuals cannot easily opt out
Examples:
- Public services
- Employment
- Education
- Essential services
- Has a systemic or large‑scale effect
Examples:
- Used by millions of people
- Used by public authorities
- Used in critical infrastructure
- Cannot be adequately mitigated by existing rules
If existing EU law already covers the risk (e.g., medical devices, aviation), the Commission may decide not to add it.
🛠️ What the Commission Must Publish
When adding a new high‑risk category, the Commission must provide:
- A detailed justification
- A risk assessment
- Evidence of consultation
- An explanation of why the system meets the Annex III criteria
This transparency requirement is designed to prevent arbitrary expansion.
🔄 How Often Annex III Can Be Updated
There is no fixed schedule, but the Commission is expected to:
- Monitor emerging AI risks continuously
- Update Annex III when new high‑risk uses appear
- Use input from the AI Office, national regulators, and civil‑society reports
In practice, updates may occur every 1–3 years depending on technological developments.
📌 Summary
High‑risk systems can be added to Annex III through a delegated act adopted by the European Commission.
To add a new category, the Commission must show that the AI system:
- Poses significant risks to health, safety, or fundamental rights
- Operates in a sensitive or high‑impact context
- Cannot be adequately regulated through other EU laws
Parliament and the Council can veto the addition, but they cannot rewrite it.
This mechanism ensures the EU AI Act remains adaptable as AI evolves.
Explain what other high risk systems are included in Article 6(1)
Article 6(1) covers a separate category of high‑risk systems in addition to those listed in Annex III.
⭐ Short Answer (Summary)
Under Article 6(1), an AI system is automatically classified as high‑risk if it is a safety component of a regulated product or is itself a regulated product under specific EU product‑safety laws and must undergo a third‑party conformity assessment. This includes AI embedded in medical devices, machinery, toys, vehicles, aviation systems, and more.
🧩 What Article 6(1) Covers (Beyond Annex III)
Article 6(1) identifies high‑risk AI systems based on existing EU product‑safety legislation. These systems become high‑risk even if they are not listed in Annex III, because they are tied to products where safety is critical.
To qualify, both conditions must be met:
- The AI system is a safety component of a product (or is itself the product).
- That product is covered by one of the EU laws listed in Annex I and requires a third‑party conformity assessment.
🏗️ Examples of High‑Risk Systems Under Article 6(1)
These examples come from the types of products regulated under Annex I legislation.
- AI in Medical Devices
Examples:
- AI‑powered diagnostic tools
- AI used in radiology, pathology, or surgical robots
- AI that controls implantable devices (e.g., pacemakers)
These are high‑risk because medical devices require third‑party conformity assessments.
- AI in Machinery and Industrial Equipment
Examples:
- AI controlling robotic arms
- AI safety systems in manufacturing lines
- AI used in autonomous industrial vehicles
These fall under the Machinery Regulation.
- AI in Vehicles (Automotive)
Examples:
- AI used in advanced driver‑assistance systems (ADAS)
- AI controlling braking, steering, or stability systems
- AI used in autonomous driving components
These are regulated under EU vehicle safety legislation.
- AI in Aviation
Examples:
- AI used in aircraft navigation or control
- AI‑based safety monitoring systems
- AI in unmanned aircraft (drones) requiring certification
Aviation systems require strict conformity assessments.
- AI in Marine and Rail Transport
Examples:
- AI used in ship navigation or collision‑avoidance
- AI controlling rail signaling or braking systems
Transport safety systems are heavily regulated.
- AI in Pressure Equipment, Lifts, and Construction Products
Examples:
- AI controlling lift safety mechanisms
- AI monitoring structural integrity
- AI regulating pressure systems
These products require third‑party safety certification.
- AI in Toys (When Safety‑Critical)
Examples:
- AI‑enabled toys with autonomous movement
- AI systems that control physical behavior of toys
If the toy requires a third‑party safety assessment, the AI becomes high‑risk.
📌 Why These Systems Are High‑Risk
Because they are embedded in products where malfunction can cause physical harm, the EU automatically classifies them as high‑risk. This is separate from Annex III, which focuses on fundamental rights risks.
Are there cases where Annex III examples are not high risk, and how must they be documented by an organization?
There are indeed cases where an AI system that looks like it falls under Annex III is not treated as high‑risk, and the Act requires organizations to document and justify that conclusion through a formal process.
Let’s break this down clearly and practically.
⭐ When Annex III Systems Are Not High‑Risk
This situation is governed by Article 6(3) — the “rebuttable presumption” clause.
Annex III lists use‑case categories that are presumed to be high‑risk.
But Article 6(3) allows an organization to prove that its specific system does not pose a significant risk to health, safety, or fundamental rights.
If the organization can demonstrate this, the system does not have to be treated as high‑risk, even if it fits an Annex III category.
This is intentional: Annex III is broad, and the EU wanted to avoid over‑classification.
🧩 Examples of Annex III Systems That Might Not Be High‑Risk
Here are realistic examples where the presumption can be rebutted:
- Education — Low‑impact, non‑automated assessment tools
Annex III includes AI used to evaluate learning outcomes.
But a system that merely suggests quiz difficulty levels, with no impact on grades or access to programs, may not be high‑risk.
- Employment — Tools with no material effect on decisions
Annex III covers hiring and worker management.
But a system that only summarizes CVs for a human recruiter, without ranking or filtering candidates, may not be high‑risk.
- Essential services — Non‑determinative creditworthiness tools
Credit scoring is high‑risk.
But a system that provides non‑binding affordability suggestions to consumers (not used by lenders) may fall outside high‑risk classification.
- Law enforcement — Tools not used for operational decisions
If an AI system is used only for administrative analytics (e.g., resource planning), not for profiling or risk scoring, it may not be high‑risk.
- Migration/asylum — Tools not used for eligibility decisions
If AI only helps applicants fill out forms, not assess eligibility, it may not be high‑risk.
These examples reflect the EU’s intent: classification depends on the actual function and impact, not the label.
📘 What Organizations Must Document (Required Under Article 6(3))
To rebut the presumption, an organization must prepare a formal, written assessment showing that the AI system:
- Does not materially influence decisions affecting rights or safety
You must show the system has no meaningful effect on:
- access to education
- employment
- essential services
- legal status
- safety‑critical outcomes
- Does not pose significant risks to fundamental rights
This includes risks of:
- discrimination
- unfair treatment
- privacy violations
- lack of transparency
- autonomy harms
- Is not used in a way that creates systemic or large‑scale impact
If the system is low‑impact, small‑scale, or advisory only, this supports the rebuttal.
- Has effective human oversight that prevents harmful outcomes
The organization must show:
- humans retain meaningful control
- AI outputs are not determinative
- oversight is active, not symbolic
- Has safeguards that reduce risk to a non‑high‑risk level
This can include:
- limiting the system’s scope
- restricting automated decision‑making
- ensuring outputs are non‑binding
- removing sensitive attributes
🗂️ Required Documentation Must Include:
A compliant rebuttal file should contain:
- System description
- Intended purpose
- User group and deployment context
- Risk assessment (fundamental rights + safety)
- Impact analysis
- Human oversight description
- Safeguards and mitigations
- Justification for why the system is not high‑risk
- Decision and sign‑off by senior management
This documentation must be:
- kept internally
- made available to regulators upon request
- updated if the system or its use changes
📌 Key Takeaway
Annex III creates a presumption of high‑risk — not an automatic classification.
Under Article 6(3), organizations can demonstrate that their specific system:
- does not materially affect rights or safety
- does not create significant risk
- is advisory or low‑impact
- has strong safeguards
…and therefore does not need to comply with the full high‑risk regime.
But this requires formal, structured documentation that regulators can audit.