EU AI Act Ethics Guidelines

The EU High-Level Expert Group on AI (HLEG) defines seven requirements for trustworthy AI:1. Human agency and oversight: AI systems must enable human control and preserve user autonomy.2. Technical robustness and safety: Systems must be reliable, resilient, and protected against manipulation.3. Privacy and data governance: Processing of personal data must be GDPR-compliant and transparent.4. Transparency: AI decisions must be traceable and explainable.5. Diversity, non-discrimination, and fairness: AI must not disadvantage any group of people.6. Societal and environmental well-being: AI should have positive societal impact.7. Accountability: Clear responsibilities and audit mechanisms must be established.These requirements form the ethical foundation of the EU AI Act and apply particularly to high-risk AI systems.

AI ethics under the EU AI Act encompasses binding requirements for the development and deployment of AI systems in the EU. Unlike earlier voluntary guidelines, the EU AI Act creates the first legally binding framework.Core elements include:- Risk-based approach with four categories (unacceptable, high, limited, minimal risk)- Mandatory human oversight for high-risk AI (Article 14)- Transparency requirements for all AI systems- Prohibition of manipulative and discriminatory AI practices- Documentation and conformity assessment obligationsThe HLEG ethics guidelines from

2019 served as the direct basis for the legislation. Organisations that have already implemented ethical AI frameworks have a head start on EU AI Act compliance.

Human oversight is a core requirement of the EU AI Act, particularly for high-risk AI systems under Article 14. Implementation spans several layers:Organisational:- Appointing responsible persons for each AI system- Defining escalation processes for critical decisions- Training oversight staff on AI capabilities and limitationsTechnical:- Building stop mechanisms (human-in-the-loop or human-on-the-loop)- Dashboards for real-time monitoring of AI decisions- Alerting on anomalies or unexpected outputsProcess:- Regular review of AI outputs by domain experts- Documentation of all human interventions- Periodic assessment of whether oversight levels are appropriateThe degree of human oversight depends on the risk category of the AI system.

The EU AI Act entered into force on

1 August

2024 and becomes applicable in phases:- February 2025: Prohibited AI practices and AI literacy requirements apply- August 2025: Rules for general-purpose AI models (GPAI) apply- August 2026: Full applicability of all provisions, including ethics requirements for high-risk AI systems- August 2027: Certain high-risk AI systems under Annex I (e.g. in regulated products)Organisations should use the remaining time to inventory their AI systems, conduct risk assessments, and build ethical governance structures. Early implementation significantly reduces compliance risk.

Ethical AI and AI compliance are related but distinct concepts:AI compliance means meeting minimum legal requirements — under the EU AI Act, this includes conformity assessment, CE marking, and documentation obligations. It is binary: an organisation is compliant or it is not.Ethical AI goes further and asks whether an AI system acts responsibly, fairly, and in society's interest. The HLEG guidelines define ethics as a separate pillar alongside lawfulness and robustness.In practice, they complement each other:- Compliance provides the legal framework- Ethics addresses questions the law does not regulate (e.g. moral limits of automation)- Organisations with ethical AI frameworks find EU AI Act compliance easierADVISORI recommends thinking about ethics and compliance as integrated from the start, rather than as separate workstreams.

An AI ethics framework is a structured set of rules defining ethical principles for AI development and deployment within an organisation. It typically consists of:Principles: Guiding values based on the HLEG requirements (fairness, transparency, accountability, etc.)Governance structure: Roles and responsibilities — e.g. an AI ethics board, a Chief AI Ethics Officer, or an interdisciplinary committee.Processes: Ethics impact assessments before deploying new AI systems, regular fairness audits, bias monitoring, and complaint procedures.Technical measures: Bias detection tools, explainability mechanisms, audit trails, and documentation.Building one in practice:1. Inventory: What AI systems exist and what is their risk potential?2. Gap analysis: Where are ethical guidelines or technical safeguards missing?3. Framework development: Define principles, governance, and processes4. Pilot: Test the framework on a specific AI system5. Rollout and continuous improvement

Fairness testing is a central component of ethical AI governance. The EU AI Act explicitly requires measures against discrimination for high-risk AI systems.Testing methods:- Statistical fairness metrics: Demographic parity, equalised error rates, and individual fairness across groups- Data quality review: Are training data representative? Are there historical biases?- Adversarial testing: Targeted tests with edge cases and protected characteristics- Explainability analysis: Which features most influence decisions?Practical implementation:- Before deployment: Bias audit of training data and model- During operation: Continuous monitoring of decision distributions- On deviations: Escalation, root cause analysis, and model adjustment- Documentation: Record all tests and results transparentlyFairness should be understood not as a one-time audit but as an ongoing process.