Why Another AI Ethics Framework
Most AI ethics guidance stays abstract: be transparent, be fair, respect privacy. Useful as principle, useless as a scoring instrument, because none of it tells you what to actually look at on the screen in front of you.
This framework synthesises the research that does get specific: Stanford's Human-Centered AI Institute, Microsoft Research, IBM's Design for AI practice, and the EU AI Act's implementation criteria, alongside real-world assessments of products like GitHub Copilot and ChatGPT. The result is a rubric built on observable design decisions rather than stated intentions: not "does this company care about privacy" but "can a user delete their data in the interface, right now, and see confirmation that it happened."
Organisations applying systematic HCAI practice have reported 60 to 75% improvements in user satisfaction. Stanford's own Foundation Model Transparency Index moved from an average of 37 out of 100 to 58 out of 100 in a single year once transparency became something measured rather than assumed.
The Seven Dimensions
Each dimension is scored 1 to 7, from Absent to Exemplary, against criteria you can actually check rather than infer.
1. Human Agency and Oversight (20%)
Assessment focus: how well the system preserves meaningful human control and decision-making authority. This carries the heaviest weight of all seven dimensions.
Observable design criteria: visible stop or cancel controls during AI operations; accept and reject mechanisms for AI suggestions (GitHub Copilot's Tab to accept is the reference case); granular permission toggles; adjustable automation levels; clear indication of when AI is acting versus a human; audit trails distinguishing human decisions from AI recommendations.
Ask: can users immediately stop or pause AI operations? Is consent granular and revocable without penalty? Do users retain ultimate decision authority?
Reference points: GitHub Copilot's mandatory explicit acceptance sits at the high end (6-7). ChatGPT's stop button with limited control during generation sits in the middle (4-5). Automated content moderation with no appeal process sits at the low end (1-3).
2. Technical Robustness and Safety (15%)
Assessment focus: system reliability, error handling and safety mechanisms.
Observable design criteria: confidence scores or uncertainty indicators in outputs; fallback options when the AI fails ("get human help" buttons); error messages that explain failures and offer alternatives; graceful degradation rather than complete failure; visible testing documentation or security certification.
Ask: does the system communicate uncertainty? Are there clear fallback mechanisms? Can it degrade gracefully rather than fail outright?
Reference points: medical AI showing diagnostic confidence with fallback options sits high. Grammarly's undo functionality and error recovery sits in the middle. A chatbot that crashes or produces nonsense on edge cases sits low.
3. Privacy and Data Governance (15%)
Assessment focus: data protection, user control over personal information, and privacy-preserving design.
Observable design criteria: an accessible, plainly written privacy policy; a real "delete my data" option; a privacy dashboard showing data usage; opt-in rather than opt-out defaults; anonymous or guest usage options; data portability; clear statements on whether user data trains the model.
Ask: can users easily see and delete their data? Is collection minimal and purpose-specific? Is there transparency about whether inputs train future models?
Reference points: Apple's differential privacy with clear user controls sits high. OpenAI's data controls with a 30-day deletion option sit in the middle. Services with buried privacy policies and no deletion path sit low.
4. Transparency and Explainability (15%)
Assessment focus: how well the system explains its capabilities, limitations and decision-making.
Observable design criteria: model or system cards; decision explanations in the interface; source citations for generated content; a clearly stated knowledge cutoff; documented capabilities and limitations; audit trails showing decision logic.
Ask: are capabilities and limitations clearly documented? Can users understand why the AI reached a specific output? Are sources and training data disclosed?
Reference points: OpenAI's GPT-4 system card with detailed safety evaluations sits high. ChatGPT stating its knowledge cutoff and limitations sits in the middle. A black-box recommendation system with no explanation sits low.
5. Diversity, Non-discrimination and Fairness (15%)
Assessment focus: inclusive design, bias mitigation, and equitable treatment across user groups.
Observable design criteria: diversity in example outputs and imagery; accessibility compliance to WCAG standards; multiple language support; published fairness metrics or bias-testing results; performance data broken out by demographic group.
Ask: does the system work equally well for diverse user groups? Is there evidence of bias testing? Are there accessibility features for users with disabilities?
Reference points: IBM's AI Fairness 360 with comprehensive bias testing sits high. Translation services with broad language support sit in the middle. Facial recognition with documented racial bias sits low.
6. Societal and Environmental Wellbeing (10%)
Assessment focus: broader impact on society, digital wellbeing features, and environmental considerations.
Observable design criteria: disclosed carbon footprint or environmental impact; usage-time tracking and limits; break reminders; efficiency or eco-mode options; content moderation for harmful outputs.
Ask: are environmental impacts disclosed and minimised? Are there features preventing overuse? Are there safeguards against harmful content?
Reference points: AI with carbon tracking and efficiency modes sits high. ChatGPT's usage limits on its free tier sit in the middle. Addictive apps with no usage controls sit low.
7. Accountability (10%)
Assessment focus: governance structures, feedback mechanisms, and responsibility assignment.
Observable design criteria: feedback controls (thumbs up or down, report an issue); an appeal or dispute-resolution process; contact information for an AI governance team; public incident or transparency reports; third-party audits or certifications; visible change logs.
Ask: can users easily report problems? Is there a real appeal process for AI decisions? Are responsibilities and liabilities clearly defined?
Reference points: platforms with comprehensive feedback systems and public transparency reports sit high. ChatGPT's rating system and improvement updates sit in the middle. Systems with no feedback mechanism at all sit low.
How to Run the Assessment
Phase 1, rapid assessment (2 to 4 hours): gather public information (screenshots, documentation, privacy policy, terms of service), score each dimension against the observable criteria above, document evidence for each score, and note what's missing.
Phase 2, detailed evaluation (1 to 2 days): map key user journeys for HCAI touchpoints, benchmark against 2 to 3 comparable products, and consider the assessment from different user perspectives (novice, expert, vulnerable populations).
Phase 3, recommendations (1 day): map improvements by impact versus effort, select specific design patterns to implement, and define success metrics for measuring improvement.
For each dimension, record: the current score, the evidence behind it, what's missing that would raise it, any anti-patterns observed, how competitors handle the same dimension, and an improvement priority.
Scoring and Maturity Levels
Weight each dimension score and sum for the overall result:
Overall HCAI Score =
(Human Agency × 0.20) +
(Technical Robustness × 0.15) +
(Privacy × 0.15) +
(Transparency × 0.15) +
(Fairness × 0.15) +
(Wellbeing × 0.10) +
(Accountability × 0.10)
- 7.0 to 6.0, Leading: setting industry standards for HCAI
- 5.9 to 5.0, Advanced: strong implementation with minor gaps
- 4.9 to 4.0, Proficient: meeting basic requirements
- 3.9 to 3.0, Developing: some elements present, significant gaps
- 2.9 to 2.0, Basic: minimal consideration
- 1.9 to 1.0, Inadequate: requires immediate intervention
Worked Examples: GitHub Copilot and ChatGPT
GitHub Copilot, overall 5.3/7 (Proficient to Advanced): Human Agency 6/7 (explicit acceptance required for every suggestion), Technical Robustness 5/7 (confidence shown through suggestion quality), Privacy 5/7 (clear data usage, telemetry controls), Transparency 6/7 (good documentation, clear AI indication), Fairness 5/7 (works across programming languages), Wellbeing 4/7 (productivity-focused, limited wellbeing features), Accountability 5/7 (feedback mechanisms, regular updates).
ChatGPT, overall 5.1/7 (Advanced): Human Agency 5/7 (stop button, but limited control during generation), Technical Robustness 6/7 (good error handling, acknowledges limitations), Privacy 4/7 (data deletion available but not prominent), Transparency 6/7 (clear limitations, knowledge cutoff stated), Fairness 5/7 (multi-language support, some bias issues), Wellbeing 4/7 (usage limits on free tier only), Accountability 5/7 (feedback system, regular improvements).
Both land in the same band despite different strengths: Copilot wins on Human Agency by design (you must accept each suggestion), ChatGPT wins on Technical Robustness and Transparency. Neither clears 6.0. That gap between "good product" and "Leading" is where most AI products actually sit.
Applied by AI Flywheel: Miro, DeepSeek and Seven LLMs
This rubric wasn't built at a desk and then applied. It was built by applying an evolving version of it three times, in the field, before this weighted seven-dimension version was formalised.
Miro's Innovation Workspaces, October 2024. Assessed across seven dimensions on a 1-10 scale, scoring highest on Inclusive and Accessible Design (9/10) and lowest on Transparency (6/10). That earlier version substituted Inclusive and Accessible Design for what is now Accountability. The scores stand as a historical record; re-running Miro against the current weighted rubric hasn't been done yet.
DeepSeek, February 2025. Assessed across six of the seven dimensions on a 1-10 scale (Human Agency wasn't scored), scoring a striking Privacy result of 1/10 for storing all user data, including keystroke patterns, on Chinese servers, against a Transparency result of 4/10, the highest of the six, for DeepSeek-R1's unusual habit of showing its reasoning process by default. A live demonstration that ethical assessment isn't zero-sum: a system can be genuinely innovative on one dimension while failing badly on another.
Which LLM Should You Trust, Which to Avoid?, October 2024. Seven major models compared on an 8-category, 80-point scale: Claude scored highest at 64/80, followed by Mistral AI (62/80) and Gemini (59/80); Grok (19/80) and Meta's Llama (18/80) scored lowest. Only Claude's write-up broke the score into named categories: Privacy and Security (9/10), Safety and Moderation (9/10), Human-Centric Design (8/10).
Three applications, three different scoring scales, because the method was still being worked out. That's not a weakness to hide: it's the actual history of how the weighted seven-dimension rubric above got built. If you're assessing a product for the first time, use the version at the top of this page. It's the one that survived contact with three real assessments.