To Implement AI Ethics Requirements
This Portfolio provides a non-exhaustive collection of technical and organisational measures for implementing the seven HLEG ALTAI ethics requirements for AI systems.
The Portfolio is intended as a thinking and implementation aid for ethics evaluators, project managers, researchers, engineers, policy makers, or students.
About this portfolio and how to use it
Developed within the Horizon Europe project AIOLIA, this Portfolio was built through a bottom-up approach to operationalising AI ethics in three stages: (1) a co-creation process involving academic, policy, and industry partners who proposed measures for real-world use cases; (2) several rounds of internal review and adaptation; and (3) pilot testing during training sessions.
The Portfolio is not an exhaustive catalogue. Measures can be rephrased and complemented to adapt to specific contexts.
The Portfolio supports AI ethics operationalisation as an iterative reflective process to implement complex ethics requirements in practice. We explicitly encourage dialogue and discussion when selecting and prioritising practical measures.
Measures in the Portfolio should always be considered in relation to the specific design contexts, use situations, AI system features, and other organisational and societal factors. They may be, and often should be, rephrased or complemented with contextual requirements. Measures are structured under the seven ALTAI requirements which are, when necessary, further divided into components. Measures are divided into three types:
- ORG Organisational Measures
- Focus on organisational practices, structures, policies, standard operating procedures, or frameworks used to govern the development and deployment of AI systems.
- TECH Technical Measures
- Focus on the technical design of AI models or systems and refer to specific techniques, modules, routines, processes, or features.
- BOTH Both
- Includes both technical and organisational aspects. For example, adopting a safety policy and implementing technical controls to enforce it counts as BOTH.
119 of 119 measures shown
Browse by requirement, or select a measure ID for a direct link to that measure.
Human Agency and Oversight
Requirement 1a: Oversight
A. Human control
-
The AI system has built-in features that support human oversight and control for decision-making in critical areas
-
The organisation has defined a clear policy on high-risk or sensitive decisions that must involve human judgment
-
Automated decisions can be paused, overridden, or reversed by humans
-
Human oversight involves active review, validation, or challenge of system outputs
-
Oversight responsibilities are clearly defined and embedded into workflow
B. AI oversight
-
AI-assisted oversight is implemented when human capacity to perform meaningful oversight reaches its limits
-
Capabilities of AI oversight system are benchmarked and documented
-
Predefined thresholds trigger automated intervention, human escalation, or system interruption when specified risks, failures, or anomalous behaviour are detected
-
Controversial or morally ambiguous situations are referred to humans for decision making
C. Defining legitimate boundaries
-
When an AI system makes decisions with a high level of autonomy, it ensures that the decision stays within the scope or mandate defined by the provider or the user
-
The organisation has clearly defined what types of content, behaviour, or expression are restricted for AI use and why
-
AI system’s features and behaviours are aligned with its intended use and foreseen limitations
-
The scope of restrictions is documented and publicly communicated insofar as it is required by the existing regulation
D. Transparency and contestability
-
Restrictions are explained to users in a clear and accessible way
-
The organisation monitors patterns of restriction to detect systemic censorship or bias
-
Moderation policies are periodically reviewed to ensure that they remain relevant and effective
-
Moderation policies take contextual and cultural factors into account
-
The organisation has put in place a mechanism for users to report, question, or contest restrictions or moderation decisions
Requirement 1b: Human Autonomy and Agency
A. Risk of manipulation
-
When manipulation risks may arise, AI outputs and human-machine interactions are logged and documented
-
Human oversight exists to review AI outputs for manipulative effects in sensitive or high-impact contexts
-
Risk analysis considers psychological, emotional, and behavioural influences on users, not only user information or technical AI performance
-
Attention has been given during design to asymmetries of power, knowledge, or vulnerability between humans and the AI system
-
The organisation defines context-specific criteria distinguishing legitimate behavioural influence from unacceptable manipulation and monitors interactions against these criteria
B. Design safeguards against undue influence
-
Techniques that steer user behaviour through deception, coercion or exploitation are identified and prevented
-
Outputs are framed to support human reflection and to avoid immediate reaction, rather than pressure or exploit cognitive bias
-
Personalisation of AI interaction strategies that exploits user vulnerabilities is prevented
-
The organisation addresses, in a specific operational context, unacceptable manipulative practices through corrective actions, design changes, or restrictions on AI use
C. Human Agency
-
AI system encourages users to make meaningful choices among relevant alternatives
-
The organisation assesses whether users experience human-machine interaction as coercive or misleading
-
User feedback and complaints related to perceived manipulation are systematically collected and reviewed
D. Informed interaction
-
Users are informed when system outputs are intended to influence decisions or behaviours
-
The AI system is designed to ensure that users are informed at appropriately defined time intervals or checkpoints that they are interacting with an AI system
Technical Robustness and Safety
Requirement 2a: Robustness and Reliability
A. Technical robustness and reliability
-
AI systems display confidence scores or uncertainty metrics for AI-generated suggestions
-
Design is modular and layered to introduce system redundancy
-
Potential failure modes in components, data and algorithms are identified, and associated risks are assessed and mitigated
B. Traceability
-
AI outputs and system configurations are logged in a way that allows for external reconstruction of decision making
-
Data is traceable via dataset source, pre-processing, labelling, and version control
Requirement 2b: Human safety and non-maleficence
A. Safety monitoring
-
AI safety assessment is used to automatically classify content into safety categories (tiers of risks, harms) at scale
-
Safety categories are developed and reviewed collaboratively by technical and legal teams
-
Safety criteria and categories (risks, harms) are clearly defined, and content falling within prohibited categories is filtered out
-
Safety criteria are continuously assessed and updated against emerging patterns
-
Training data and model behaviour are steered or limited to safe domains
-
Measures are in place to prevent misuse and attempts to circumvent safeguards
B. “Know thy user”
-
Vulnerable users are identified, including via analysis of user behaviour and interaction patterns across multiple sessions
-
Additional safety measures are implemented for vulnerable users (e.g., indicators of acute personal crisis that trigger escalation to a support service)
-
A progressive escalation process is implemented from highlighting prohibited content to warning and user ban
-
Human moderators are involved in decision-making on high-risk content and user bans
C. Human moderators
-
Human moderators receive regular training on representative examples and data
-
The organisation has implemented monitoring and checks against moderators’ bias and over-reliance
-
The organisation implements a policy to protect moderators’ well-being
D. Safety culture
-
The organisation promotes a safety-first culture
-
The organisation embeds safety and traceability measures throughout development, deployment, and operation of AI systems
-
The organisation conducts independent audits of safety policies
-
Within existing regulatory limits, the organisation publicly discloses anonymised statistics on high-risk behaviour and significant safety risks
Privacy and Data Governance
A. Personal data collection, use, and processing
-
The organisation clearly defines and documents the purpose of personal data processing
-
Individuals are informed in clear and accessible language about personal data collection and use
-
Consent procedures are adapted to context to ensure informed and meaningful choice
-
Limitations of informed consent during AI use are assessed and documented
B. Data minimisation and protection
-
Only data necessary for the stated purpose is collected and processed
-
Technical measures are in place to protect data from unauthorised access, loss, or misuse
-
Data retention periods are defined and enforced
-
Sensitive data is subject to appropriate additional safeguards
C. Rights of data subjects under applicable law
-
Data subjects can refuse or withdraw consent without penalty
-
Clear and accessible processes exist for individuals to exercise their data protection rights (e.g., access, correction, deletion), including easy access to contact information for the data controller
-
Data processing is documented to support accountability and audits
-
Data protection practices are reviewed following incidents, complaints, or regulatory changes
Transparency
A. Openness about intent and purpose
-
The organisation clearly communicates the intended purpose, scope, and use of the AI system to relevant stakeholders
-
The organisation clearly communicates how the AI system may affect individuals or groups
-
The organisation monitors departures from the AI system’s intended purpose and scope of use and informs relevant stakeholders
-
Users are informed that their interaction is monitored and/or controlled by AI systems
B. Accessibility of information
-
Information about AI system functioning, capabilities, limitations, and safeguards is accessible to all relevant stakeholders to the degree needed and possible
-
Communication channels and formats are adapted to different roles, languages, digital literacy, and accessibility needs
C. Transparency beyond disclosure
-
Additional interaction safeguards beyond disclosure are implemented when user manipulation or psychological, emotional, and behavioural influences may arise
-
The organisation has a culture of considering non-standard design measures beyond user information (e.g., disrupting user experience or gamifying user interaction)
D. Explainability of outputs and decisions
-
Explanations are relevant and address the main factors influencing the output or decision, at a level appropriate to the audience, context and consequences
-
The organisation evaluates whether information provided to users produces adequate understanding, including key behavioural factors, limitations, and uncertainties in their interaction with AI system
-
Information and explanations provided about AI-supported decisions are sufficient to support meaningful review, justification, or contestation
Diversity, Non-discrimination, and Fairness
A. Diversity and representativeness
-
Data used to design, train or configure the AI system is assessed alongside relevant fairness metrics for representativeness across relevant groups and contexts
-
Data audits are used to detect underrepresented cases (edge scenarios, rare hazards)
-
The organisation considers diversity criteria and their intersection across roles, locations, languages, or cultural characteristics
-
The organisation considers multiple stakeholder perspectives on fairness during AI system design
-
The organisation clearly communicates limitations on data diversity and fairness
B. Objective and consistent assessment of fairness
-
Sensitivity analysis is used to check fairness across different inputs
-
AI decisions are assessed for consistency to ensure that comparable cases receive comparable outcomes
-
AI outputs are reviewed for unintended disparate impacts
-
The organisation conducts appropriate fairness checks to assess and prevent unfair or stigmatising profiling
C. Monitoring and bias correction
-
AI outputs are monitored for emerging bias over time, using context-appropriate fairness indicators
-
Bias identification triggers corrective action (e.g., adjustment to data or organisational practices)
-
The organisation conducts focus groups (e.g., ‘disagreement analysis’ sessions) to identify the sources of unfairness or bias
-
Mechanisms exist for users to challenge outcomes perceived as unfair or discriminatory
Environmental and Societal well-being
Since Environmental well-being is not in the scope of AIOLIA project, no measures have been added.
Requirement 6a: Individual and Societal well-being
A. Improved well-being
-
AI system detects contexts indicating potential tension between AI agents / users or potential social conflict
-
AI system detects and avoids AI-generated content that may reduce individual or societal well-being, such as fake news, disinformation, or misinformation
-
AI system detects and discourages user behaviour that may reduce individual or societal well-being, such as violent actions or polarised opinions
-
The organisation collects and audits user reports on well-being impacts
B. Behavioural constraints
-
AI system avoids interactional extremes that may cause undue influence on human behaviour (e.g. sycophancy or exaggeration)
-
In clearly defined contexts, AI system promotes friction-by-design in human-machine interaction (e.g., realistic conversations where the AI system may contradict the user)
-
AI system applies behavioural nudging only in documented and auditable contexts with specified trigger thresholds and objectives
-
The organisation has a policy and procedures for monitoring, assessing, and addressing unforeseen impacts of the AI system on individual and societal well-being
Requirement 6b: Limit Over-reliance and Deskilling
A. Managing dependence on AI outputs
-
The organisation has defined clear boundaries and guidance for intended use of AI system, including when AI outputs should be questioned, verified, or supplemented with human input
-
AI system detects indicators of excessive psychological or emotional dependence and activates safeguards to limit over-reliance
-
AI system identifies high-impact decision points in the professional workflow and enforces human participation in decision-making
-
The organisation explicitly discourages treating AI outputs as definitive or infallible
B. Preserving human judgment
-
AI interface discourages automatic or hasty user reaction (e.g., default approvals without human review)
-
AI system does not penalise users for rejecting, ignoring, or questioning AI suggestions
C. Skills, training and competence
-
Training exists to support users in understanding AI capabilities and limitations, as well as related human capabilities and limitations
-
The organisation periodically reassesses reliance patterns as systems evolve or scale
Accountability
A. Responsibility assignments
-
Responsibility for AI outputs is explicitly defined and documented (e.g., ownership of decisions, approvals, and incident response)
-
Responsibility for monitoring harm indicators and responding to them is explicitly assigned
-
Responsibilities at operational, managerial and governance levels are defined to provide effective resolution of conflicts and tensions
-
An escalation procedure is defined with the organisation’s hierarchy for resolving unclear or disputed responsibility assignments
-
“No gaps”: responsibility assignments are maintained across system updates, handover, or organisational changes
-
Responsibility assignments are integrated into validation, governance, and risk management processes
B. Auditability
-
Auditable Standard Operating Procedures exist for AI design and validation
-
Changes to models, data or AI system are documented and traceable
-
Records are sufficient to support internal reviews, external audits, or regulatory scrutiny
-
Audit findings are translated into corrective design action
C. Learning, remediation and improvement
-
Lessons from incidents, errors, or near-misses are systematically used to improve AI design and governance processes
-
Accountability assignments are periodically reviewed to remain aligned with legal obligations
-
Reflexive governance ensures that accountability mechanisms evolve at an appropriate pace as technologies, contexts, risks, or societal expectations change
Reference
High-Level Expert Group on Artificial Intelligence, 2020. Assessment List for Trustworthy Artificial Intelligence (ALTAI) for self-assessment.
Source: A. Prégent and A. Grinbaum, Developing Competency in AI Ethics: A training framework. arxiv:2610.13319