AI Ethics Frameworks
AI ethics frameworks provide structured approaches to thinking about what responsible AI development and deployment looks like. Since 2016, over 160 sets of AI ethics principles have been published by governments, companies, academic institutions, and civil society organisations. While there is substantial convergence on core principles — fairness, accountability, transparency, safety, privacy — the principles themselves are often too abstract to drive specific technical or organisational decisions. Understanding both the frameworks and their limitations is essential for anyone working in AI governance.
Core AI Ethics Principles
Across the major frameworks, five principles appear consistently enough to be considered the de facto consensus foundation:
Fairness
AI systems should not produce unjust or discriminatory outcomes — across individuals or demographic groups. Requires proactive identification and mitigation of bias, not just the absence of deliberate discrimination.
Accountability
Someone must be responsible for the outcomes of an AI system — developers, deployers, or operators. Accountability requires that responsibility is assigned, not diffused, and that there are consequences for failures.
Transparency
Relevant information about AI systems — what they do, how they work, where they are used, and what they affect — should be available to appropriate parties. Transparency is contextual: what developers need to know differs from what affected individuals need.
Safety
AI systems should not cause harm to individuals, organisations, or society — including harms that arise from errors, misuse, or unanticipated interactions with deployment context.
Privacy
AI systems that process personal data must respect individuals' privacy rights — including not just compliance with data protection law but the spirit of privacy as a fundamental right.
Human autonomy
AI systems should support rather than undermine meaningful human agency — preserving the ability of individuals to understand and, where appropriate, contest AI-influenced decisions affecting them.
EU High-Level Expert Group: Seven Requirements for Trustworthy AI
The EU HLEG on AI (2019) produced the most influential ethics framework in policy terms, as it directly informed the EU AI Act. The seven requirements apply across the full AI lifecycle:
| Requirement | Key implication |
|---|---|
| 1. Human agency and oversight | AI should support human decision-making, not replace it where human judgment matters; meaningful human control must remain possible |
| 2. Technical robustness and safety | AI must be accurate, reliable, and secure; able to handle adversarial conditions and unexpected inputs without causing harm |
| 3. Privacy and data governance | Privacy must be respected throughout the data lifecycle; data quality, integrity, and lawful use must be ensured |
| 4. Transparency | Traceability and explainability of AI decisions; disclosure of AI use to affected individuals; no deception |
| 5. Diversity, non-discrimination, fairness | Avoid unfair bias across groups; ensure accessibility and usability for all relevant populations |
| 6. Societal and environmental wellbeing | Consider broader societal impacts and sustainability; AI should benefit all of humanity, not just users and deployers |
| 7. Accountability | Responsibility must be assigned; mechanisms for redress when harm occurs; auditability of AI systems by appropriate parties |
Industry AI Principles
Most major AI companies have published AI principles. These vary in specificity and enforceability:
- Google AI Principles (2018): Be socially beneficial; avoid creating or reinforcing unfair bias; be built and tested for safety; be accountable to people; incorporate privacy design; uphold standards of scientific excellence; be made available for uses that accord with these principles. Explicitly lists prohibited applications: weapons causing mass casualties, surveillance violating norms, undermining international law.
- Microsoft Responsible AI Principles: Fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Microsoft has operationalised these through Responsible AI Standard v2 — a detailed internal governance document that sets specific engineering requirements.
- Anthropic Constitutional AI: Uses an explicit constitution of values to guide RLHF training — AI systems are trained to evaluate and revise their own outputs against the constitution before responding. Published in Bai et al. (2022).
- OpenAI: Mission statement focuses on AGI benefit to all humanity; Preparedness Framework published 2023 defines risk tiers for catastrophic risk categories. More safety-process-oriented than a principles list.
IEEE Ethically Aligned Design
IEEE published Ethically Aligned Design (EAD) in 2019 — a detailed practitioner guide aimed at engineers building AI and autonomous systems. Key contributions beyond principles:
- Human Rights: AI systems should not infringe upon universally recognised human rights — a specific rights-based framing not present in most corporate principles
- Wellbeing: AI should prioritise wellbeing as a metric, not just economic efficiency or task performance
- Data agency: individuals should have authority over their own personal data used in AI systems
- The IEEE P7000 series of standards attempts to operationalise EAD principles into engineering practice — including P7001 (transparency), P7002 (data privacy), and P7003 (algorithmic bias)
From Principles to Practice: The Implementation Gap
The most significant critique of AI ethics frameworks is the gap between stated principles and operational practice — sometimes called "ethics washing":
Why principles fail to drive practice
- Principles are too abstract to inform specific engineering decisions
- No enforcement mechanism: principles are commitments, not obligations
- Conflicting principles (e.g., fairness and accuracy) with no guidance on how to trade them off
- Principles defined by organisations whose commercial interests conflict with their application
- Teams that see ethics as a compliance checkbox rather than a design input
What makes principles actionable
- Translate principles to specific testable requirements: "fairness" → "demographic parity gap ≤ 5pp across race groups"
- Embed ethics checkpoints in process gates — ethics review is required before model deployment, not after
- Assign ownership: name the person accountable for each principle in each AI system
- Independent oversight: external ethics board or regulator review, not just internal compliance
- Public reporting: publish metrics on ethics performance, not just principles
Checklist: Do You Understand This?
- Name the five core principles that appear across the major AI ethics frameworks.
- List the EU HLEG's seven requirements for trustworthy AI and explain what each means.
- What distinguishes Anthropic's Constitutional AI approach from a traditional AI principles document?
- What is the "implementation gap" in AI ethics, and what structural factors cause it?
- Describe three specific steps that turn a vague ethics principle into an actionable engineering requirement.