A "Humans First" Approach to AI

Human-Centered AI Frameworks & Approaches

---

Value Sensitive Design (VSD)

• Developed by: Batya Friedman and Peter Kahn at the University of Washington
• Core approach: Systematically accounts for human values throughout the design process
• Methodology: Uses conceptual, empirical, and technical investigations to identify and address stakeholder values
• Key contribution: Provides a structured methodology to incorporate human values like privacy, autonomy, and trust directly into technical systems
• Resource: Value Sensitive Design: Shaping Technology with Moral Imagination

Center for Humane Technology (CHT)

• Founded by: Tristan Harris (former Google Design Ethicist) and colleagues
• Core mission: Realigning technology with humanity’s best interests
• Focus areas: Reducing digital addiction, combating misinformation, and promoting human-centered business models
• Notable work: The Social Dilemma documentary, Humane Tech design principles
• Key insight: Identifies attention extraction as the root problem of many tech harms
• Resource: CHT website and resources

Advancing Humans with AI (MIT Media Lab)

• Core approach: Focuses on AI systems that enhance human capabilities rather than replace them
• Research areas: Human-AI collaboration, AI literacy, augmented cognition
• Philosophy: Views AI as a tool for expanding human potential rather than a substitute
• Projects: Includes work on AI education, creative collaboration, and cognitive enhancement
• Resource: AHA at MIT

Responsible AI (Microsoft)

• Six guiding principles: Fairness, reliability & safety, privacy & security, inclusiveness, transparency, and accountability
• Implementation: Comprehensive governance and public policy framework
• Practical tools: Fairlearn (fairness assessment), InterpretML (model interpretability), AI Fairness Checklist, and the Responsible AI Dashboard
• Impact assessment: Includes structured approaches to evaluate potential harms before deployment
• Resource: Responsible AI at Microsoft

Human-Centered AI (Stanford HAI)

• Founded: 2019 as an interdisciplinary institute
• Mission: Advance AI research, education, policy and practice to improve human condition
• Core philosophy: Human-centered AI technologies should enhance human capabilities, not replace them
• Key areas: Augmenting human capabilities, addressing societal impact of AI, guiding AI’s development with human values
• Notable work: Research on AI ethics, policy recommendations, educational programs
• Resource: Stanford HAI

Ethically Aligned Design (IEEE)

• Scope: Comprehensive framework for ethical considerations in autonomous and intelligent systems
• Development process: Created through global, multidisciplinary collaboration
• Key principles: Human rights, wellbeing, data agency, effectiveness, transparency, accountability
• Implementation: Provides specific recommendations for standards bodies, policymakers, and engineers
• Technical focus: Includes detailed technical approaches for embedding ethics in AI systems
• Resource: IEEE Ethically Aligned Design

Montreal Declaration for Responsible AI

• Origin: Developed at the University of Montreal through collaborative deliberation
• Structure: 10 principles - wellbeing, autonomy, privacy, solidarity, democracy, equity, diversity, prudence, responsibility, sustainability
• Distinguishing feature: Created through public participation and consultation
• Implementation: Includes self-assessment tools and governance recommendations
• Goal: Guide digital transition so everyone benefits equitably from AI advancement
• Resource: Montreal Declaration

Additional Human-Centered Methodological Approaches

Participatory Design for AI

• Directly involves end users throughout the AI development process
• Emphasizes co-creation rather than designing “for” users
• Particularly valuable for AI systems serving marginalized or underrepresented communities
• Helps identify potential harms that developers might not anticipate

Algorithmic Impact Assessments

• Similar to environmental impact assessments
• Structured evaluation of potential societal impacts before deployment
• Often includes public disclosure requirements
• Increasingly being adopted in public sector AI governance
• Example: Canada’s Algorithmic Impact Assessment Tool

Consequence Scanning

• Developed by Doteveryone (UK think tank)
• Structured workshop approach for development teams
• Asks three key questions:
  1. What are the intended and unintended consequences of this product or service?
  2. What are the positive consequences we want to focus on?
  3. What are the negative consequences we need to mitigate?
• Integrated into regular development cycles, not just at the end

Ethics by Design

• Integrates ethical reasoning throughout the entire development lifecycle
• Uses tools like ethics canvas, value proposition canvas
• Incorporates ethics-focused design patterns and best practices
• Emphasizes proactive rather than reactive ethical considerations

Human-in-the-Loop Systems

• Ensures humans maintain meaningful control and oversight in AI systems
• Particularly important in high-risk domains (healthcare, justice, etc.)
• Different models: human review, approval, oversight, or collaboration
• Recognizes that full automation isn’t always the goal
• Preserves human agency and accountability

Common Principles Across Human-Centered AI Approaches

1. Transparency & Explainability

   • AI systems should be understandable to those affected by them
   • Decisions should be explainable in human terms

2. Inclusive Design Processes

   • Diverse stakeholders should participate in development
   • Systems should work for people of all backgrounds and abilities

3. Continuous Assessment

   • Ongoing evaluation of impacts rather than one-time assessments
   • Iterative improvement based on real-world effects

4. Augmentation Over Automation

   • Focus on enhancing human capabilities rather than replacing humans
   • Preserve meaningful human agency and decision-making

5. Accountability Structures

   • Clear lines of responsibility for AI outcomes
   • Mechanisms for redress when harms occur

6. Contextual Deployment

   • Recognition that no single approach works in all contexts
   • Adaptation to specific cultural, social, and domain needs

---

Lots more…ask an AI to expand on this!

slides