UNESCO Recommendation on the Ethics of AI

Member States are encouraged to use the UNESCO Readiness Assessment Methodology to evaluate institutional, legal, technical, and capacity readiness for ethical AI, and to translate the Recommendation into national strategies and laws.

• Self-assessment against UNESCO RAM dimensions
• Roadmap of gaps with owners and timelines
• Mapping between UNESCO Principles and internal policies
• Periodic re-assessment evidence

• Self-assessment completed once and never refreshed
• Roadmap items lack owners or deadlines
• Mapping to internal policies absent, making the Recommendation aspirational only

Member States should develop and implement mechanisms for ethical impact assessment of AI systems, including regulatory frameworks and oversight mechanisms.

• Ethical Impact Assessment (EIA) reports for AI systems
• AI literacy and awareness training records
• AI risk register with human-rights and sustainability lenses
• AI ethics policy aligned to UNESCO values and principles

• Environmental footprint of AI not measured
• Bias testing absent for protected attributes
• Lack of multi-stakeholder governance forum
• EIA scope limited; missing high-risk use cases

Member States should address AI impacts on employment, labour markets, and economic inequality, and promote policies for just transitions in AI-affected sectors.

• Ethical Impact Assessment (EIA) reports for AI systems
• AI risk register with human-rights and sustainability lenses
• Stakeholder consultation records for high-impact AI
• AI ethics policy aligned to UNESCO values and principles

• Lack of multi-stakeholder governance forum
• No traceable link between principles and operational controls
• EIA scope limited; missing high-risk use cases
• Environmental footprint of AI not measured

Member States should leverage AI for health and social well-being while ensuring equitable access, safety, and ethical use of AI in healthcare and social services.

• AI risk register with human-rights and sustainability lenses
• AI ethics policy aligned to UNESCO values and principles
• Algorithmic transparency disclosures
• Ethical Impact Assessment (EIA) reports for AI systems

• Environmental footprint of AI not measured
• Bias testing absent for protected attributes
• No traceable link between principles and operational controls
• EIA scope limited; missing high-risk use cases

Member States should adopt regulatory frameworks to ensure ethical governance of AI, with appropriate institutions and oversight mechanisms.

• Stakeholder consultation records for high-impact AI
• AI risk register with human-rights and sustainability lenses
• Ethical Impact Assessment (EIA) reports for AI systems
• Algorithmic transparency disclosures

• Bias testing absent for protected attributes
• Environmental footprint of AI not measured
• No traceable link between principles and operational controls
• EIA scope limited; missing high-risk use cases

Member States should implement data governance frameworks that ensure data used in AI systems respects privacy, is representative, and is collected with informed consent.

• AI literacy and awareness training records
• Algorithmic transparency disclosures
• Stakeholder consultation records for high-impact AI
• AI risk register with human-rights and sustainability lenses

• Environmental footprint of AI not measured
• Lack of multi-stakeholder governance forum
• Bias testing absent for protected attributes
• No traceable link between principles and operational controls

Member States should promote international cooperation in AI development, particularly supporting developing countries in building AI capacity and infrastructure.

• AI ethics policy aligned to UNESCO values and principles
• Algorithmic transparency disclosures
• AI literacy and awareness training records
• AI risk register with human-rights and sustainability lenses

• No traceable link between principles and operational controls
• Lack of multi-stakeholder governance forum
• Environmental footprint of AI not measured
• Bias testing absent for protected attributes

Member States should leverage AI for environmental sustainability while reducing the environmental footprint of AI systems through energy-efficient practices.

• Algorithmic transparency disclosures
• AI ethics policy aligned to UNESCO values and principles
• Stakeholder consultation records for high-impact AI
• AI literacy and awareness training records

• Lack of multi-stakeholder governance forum
• Bias testing absent for protected attributes
• Environmental footprint of AI not measured
• EIA scope limited; missing high-risk use cases

Member States should address gender bias in AI systems and promote gender equality in AI development, deployment, and access to AI benefits.

• AI ethics policy aligned to UNESCO values and principles
• AI risk register with human-rights and sustainability lenses
• Algorithmic transparency disclosures
• Stakeholder consultation records for high-impact AI

• EIA scope limited; missing high-risk use cases
• Bias testing absent for protected attributes
• No traceable link between principles and operational controls
• Lack of multi-stakeholder governance forum

Member States should ensure AI systems respect and promote cultural diversity, linguistic diversity, and indigenous knowledge while preventing cultural homogenization.

• AI risk register with human-rights and sustainability lenses
• Ethical Impact Assessment (EIA) reports for AI systems
• AI literacy and awareness training records
• AI ethics policy aligned to UNESCO values and principles

• Bias testing absent for protected attributes
• Lack of multi-stakeholder governance forum
• EIA scope limited; missing high-risk use cases
• Environmental footprint of AI not measured

Member States should integrate AI ethics into education curricula and research agendas, promote AI literacy, and support interdisciplinary research on AI ethics.

• AI ethics policy aligned to UNESCO values and principles
• AI literacy and awareness training records
• Stakeholder consultation records for high-impact AI
• Algorithmic transparency disclosures

• Bias testing absent for protected attributes
• Environmental footprint of AI not measured
• No traceable link between principles and operational controls
• EIA scope limited; missing high-risk use cases

Member States should address AI impacts on media, information integrity, and freedom of expression, including risks from AI-generated disinformation.

• AI risk register with human-rights and sustainability lenses
• Algorithmic transparency disclosures
• AI ethics policy aligned to UNESCO values and principles
• Ethical Impact Assessment (EIA) reports for AI systems

• Lack of multi-stakeholder governance forum
• No traceable link between principles and operational controls
• Bias testing absent for protected attributes
• Environmental footprint of AI not measured

Member States and AI actors should develop and apply data policies that ensure data quality, openness where appropriate, privacy, security, interoperability, and stewardship across the AI lifecycle.

• Data governance policy covering AI training and inference data
• Data quality KPIs and monitoring
• Open data assessments where applicable
• Data stewardship roles assigned and trained

• Data quality KPIs limited to operational data, ignoring training corpora
• No stewardship for synthetic data pipelines
• Open data potential not assessed

Member States and AI actors should assess the direct and indirect environmental impact of AI systems, including carbon footprint, energy consumption, environmental impact of raw material extraction, and reduction of e-waste.

• Carbon and energy reporting for AI workloads
• Hardware lifecycle policy including responsible disposal
• Targets for reducing AI environmental footprint
• Procurement criteria favouring lower-impact compute

• No measurement of training versus inference footprint
• Hardware disposal handled by general IT with no AI-specific tracking
• Targets aspirational with no accountable owner

Member States and AI actors should ensure that the potential of AI to advance gender equality is fully maximised, while addressing harms and biases that may reproduce or amplify gender inequalities.

• Gender-disaggregated bias and fairness testing
• Gender impact assessments for AI use cases
• Workforce diversity metrics for AI teams
• Targeted measures to expand access for women and non-binary users

• Gender-disaggregated testing absent or limited to binary categories
• AI team diversity metrics not collected
• No targeted measures to address gender access gaps

Member States should introduce frameworks for impact assessments to identify and address impacts of AI systems on human rights, the rule of law, and the environment. Organisations should perform such assessments throughout the lifecycle.

• Ethical impact assessment methodology adopted internally
• EIA reports for material AI systems
• Mitigation tracking from EIAs to closure
• Mechanism to revisit EIAs after material changes

• EIAs done at launch only and never refreshed
• Mitigations not tracked to closure
• EIA methodology not aligned with UNESCO Readiness Assessment

Per UNESCO AI: 11 policy action areas including ethical impact assessment + ethical governance + data policy + development + employment + culture + education + research + health + environment + gender + child.

• UNESCO AI evidence for UNESCOAI-4

• ethical impact + policy areas partial

Public awareness and understanding of AI technologies and the value of data should be promoted through open and accessible education, civic engagement, digital skills, AI ethics training, and media and information literacy.

• AI literacy training programme for staff
• User-facing materials explaining AI uses and risks
• Community engagement records
• Targeted training for high-risk roles

• AI training restricted to technical teams
• User materials available only in major markets and major languages
• No tracking of training completion or effectiveness

AI actors shall promote social justice and fairness and non-discrimination of any kind, including by avoiding bias in data and models and ensuring inclusive access to AI benefits.

• Bias and fairness testing methodology and results
• Documentation of fairness metric choices and tradeoffs
• Data sourcing and labelling fairness review
• Access analysis showing who benefits and who is excluded

• Single fairness metric chosen with no documented tradeoff analysis
• Bias testing performed on training data but not on operational outputs
• Access analysis not performed, leaving exclusion undetected

Governance of AI shall involve multiple stakeholders including governments, the private sector, civil society, academia, and the public, and shall be adaptive to evolving technological developments.

• AI governance body charter with multi-stakeholder representation
• Stakeholder engagement plan including civil society
• Periodic review of AI governance against new developments
• Open consultation records for material AI policies

• AI governance body is entirely internal
• Civil society engagement absent or one-off
• Governance documents not refreshed for new modalities

Member States should ensure that the ultimate responsibility and accountability lies with humans. AI systems should not displace human responsibility, especially for life-and-death decisions or other consequential outcomes.

• Decision rights catalogue specifying where humans must approve AI outputs
• Logs showing human review actions on consequential decisions
• Training for human reviewers including automation bias awareness
• Escalation paths when human and AI disagree

• Human-in-the-loop reduced to a rubber stamp under throughput pressure
• Reviewers not trained on automation bias
• No defined escalation path on disagreement

Privacy shall be respected, protected, and promoted throughout the AI lifecycle. Data protection frameworks shall be developed and applied in accordance with international law.

• Data protection impact assessments for AI systems
• Lawful basis register for training data sources
• Data subject rights handling procedure including for AI outputs
• Data minimisation and retention controls in AI pipelines

• Training data sources lack a documented lawful basis
• No process to handle subject rights against AI outputs (access, rectification, erasure)
• Retention controls absent in vector stores and feature caches

AI methods shall be appropriate and proportionate to the legitimate aim pursued. Risk assessments shall be conducted and harmful uses prevented, including by selecting less intrusive alternatives where they exist.

• AI risk assessment template covering proportionality
• Documented consideration of non-AI alternatives for sensitive contexts
• Approval gate that blocks disproportionate deployments
• Periodic re-assessment of operational systems

• Risk assessments completed only for new builds, not for repurposed systems
• Alternative analysis is informal and not recorded
• Approval gates exist on paper but not enforced in delivery

AI actors and Member States shall respect, protect, and promote human rights and fundamental freedoms and shall promote the protection of the environment and ecosystems, assuming their respective ethical and legal responsibility in accordance with applicable national and international law.

• RACI for AI lifecycle covering developer, deployer, operator, and assurance roles
• Auditability provisions in contracts with AI suppliers
• Records of responsibility allocation for incidents
• Independent audit reports

• Accountability assigned to AI itself or to a generic team
• Supplier contracts lack audit and remediation provisions
• Independent audits never commissioned

Safety and security risks shall be addressed and prevented throughout the lifecycle of AI systems. AI actors shall promote sustainable, privacy-protective, and rights-respecting data ecosystems, with proportionate security controls.

• Safety case for AI systems with safety-relevant outputs
• Threat model and security controls for AI pipelines
• Red team and adversarial testing reports
• Incident response procedure tailored to AI failures

• Safety case absent for systems whose failure has material consequences
• No documented adversarial testing for prompt injection or data extraction
• Generic incident response with no AI-specific runbooks

AI systems shall be assessed against sustainability impacts including the United Nations Sustainable Development Goals. Trade-offs between AI benefits and resource use shall be considered.

• Mapping of AI use cases to relevant SDG targets
• Documented tradeoff analyses between accuracy and resource use
• Sustainability KPIs for AI programmes
• Reporting cadence for sustainability outcomes

• SDG mapping treated as marketing rather than operational
• No tradeoff analysis between model size and value delivered
• Sustainability KPIs not reviewed by leadership

The transparency and explainability of AI systems are necessary preconditions to ensure rights and to allow contestability. The level of transparency shall be appropriate to the context and the impact of the system.

• Disclosure when users interact with AI systems
• Model cards or system cards covering purpose, limits, and known risks
• Explanation interfaces appropriate to user expertise
• Contestability mechanism for affected persons

• AI interaction not disclosed in customer-facing channels
• Model cards held internally and not shared with deployers
• Contestability buried in support flow with no SLA

Per UNESCO Recommendation on the Ethics of AI: Proportionality and Do No Harm + Safety and Security + Fairness and Non-Discrimination.

• UNESCO AI evidence for UNESCOAI-1

• ethical impact + policy areas partial

Per UNESCO AI: Sustainability + Privacy and Data Protection + Human Oversight and Determination + Transparency and Explainability.

• UNESCO AI evidence for UNESCOAI-2

• ethical impact + policy areas partial

Per UNESCO AI: Responsibility and Accountability + Awareness and Literacy + Multi-Stakeholder Adaptive Governance.

• UNESCO AI evidence for UNESCOAI-3

• ethical impact + policy areas partial

AI systems must respect, protect, and promote human rights and fundamental freedoms as established in international law. AI systems should not be used to undermine human dignity.

• Algorithmic transparency disclosures
• AI risk register with human-rights and sustainability lenses
• Stakeholder consultation records for high-impact AI
• AI literacy and awareness training records

• Lack of multi-stakeholder governance forum
• EIA scope limited; missing high-risk use cases
• No traceable link between principles and operational controls
• Bias testing absent for protected attributes

AI actors should promote peaceful and just societies. AI must not be used to fragment, divide, or threaten international peace and stability.

• AI literacy and awareness training records
• AI risk register with human-rights and sustainability lenses
• AI ethics policy aligned to UNESCO values and principles
• Stakeholder consultation records for high-impact AI

• EIA scope limited; missing high-risk use cases
• Lack of multi-stakeholder governance forum
• Bias testing absent for protected attributes
• Environmental footprint of AI not measured

AI should promote social, cultural, and biological diversity. Development and use of AI should reflect diverse perspectives and avoid cultural homogenization.

• AI literacy and awareness training records
• AI risk register with human-rights and sustainability lenses
• AI ethics policy aligned to UNESCO values and principles
• Stakeholder consultation records for high-impact AI

• Environmental footprint of AI not measured
• Lack of multi-stakeholder governance forum
• EIA scope limited; missing high-risk use cases
• Bias testing absent for protected attributes

AI actors should reduce environmental impact of AI systems. AI should be leveraged to support environmental monitoring, protection, and restoration.

• Ethical Impact Assessment (EIA) reports for AI systems
• Algorithmic transparency disclosures
• Stakeholder consultation records for high-impact AI
• AI ethics policy aligned to UNESCO values and principles

• Bias testing absent for protected attributes
• No traceable link between principles and operational controls
• Environmental footprint of AI not measured
• Lack of multi-stakeholder governance forum

AI systems and their use shall respect, protect, and promote human rights, fundamental freedoms, and human dignity throughout the lifecycle. No AI system shall be used in a manner that violates international human rights law.

• Human rights impact assessment for AI systems
• Policy statement aligning AI use with international human rights instruments
• Mitigation register for identified human rights risks
• Channel for affected persons to raise concerns

• Human rights assessment treated as a one-off pre-launch exercise
• Mitigations recorded but not tracked to closure
• No accessible channel for affected persons outside customers

AI systems shall respect and promote diversity and inclusiveness, addressing gender equality, accessibility for persons with disabilities, and inclusion of marginalised and vulnerable groups.

• Inclusive design guidelines applied to AI systems
• Accessibility conformance reports (for example WCAG)
• Engagement records with marginalised user groups
• Bias and representation testing across protected characteristics

• Accessibility tested for the UI but not for AI outputs themselves
• Marginalised group engagement limited to internal staff networks
• Bias testing limited to a single protected attribute

AI actors shall promote environmental and ecosystem flourishing, including assessment of environmental impacts across the AI lifecycle and choices that reduce harm to ecosystems.

• Lifecycle environmental assessment for AI systems including training and inference
• Reporting on energy and water use of AI infrastructure
• Procurement criteria favouring lower-impact compute
• Targets for reducing environmental footprint

• Energy reporting limited to scope 2 only
• No assessment of model size choices against environmental impact
• Procurement decisions ignore embodied carbon

AI actors shall avoid contributing to the disruption of peaceful societies, including avoiding the use of AI for surveillance that undermines fundamental freedoms or for social scoring that violates rights.

• Use-case prohibition list including unlawful surveillance and rights-violating social scoring
• Review process for ambiguous or sensitive use cases
• Customer terms reflecting prohibited uses
• Records of declined engagements

• Prohibition list ignores second-line applications by customers
• No documented review when an internal team requests a sensitive use case
• Declined engagements not logged