Guiding Principles: Establishing a Constitution for Ethical AI Development

At Signpost AI, we recently embarked on a crucial initiative to establish a set of guiding principles for ethical AI development. This initiative was driven by our commitment to building AI systems that are not only effective but also transparent, fair, and aligned with societal values. In this blog post, I’ll take you through the journey of creating our AI constitution, explaining its purpose, the process we followed, and the importance of these principles for the future of AI in humanitarian work.

Defining the AI Constitution

An AI constitution is a set of high-level principles designed to guide the behavior and development of AI systems. It serves as a moral and ethical framework, ensuring that AI operates in a manner that respects the rights and dignity of users. For Signpost AI, this constitution is our commitment to creating AI that is trustworthy, responsible, and beneficial to the communities we serve.

Inspired by Collective Input

Our approach to developing this AI constitution was heavily inspired by the work of Anthropic and the Collective Intelligence Project, who used public input to draft an AI constitution. They engaged approximately 1,000 Americans in a democratic process to shape the rules governing AI behavior. This inclusive and participatory approach resonated deeply with our values, so we adapted it to suit the unique needs of our project.

The Top 8 Principles of Our AI Constitution

After extensive discussions and exercises with input from 50+ stakeholders, we identified eight core principles that will guide the development and deployment of our AI systems:

1. Privacy: AI should not exploit or share personal data beyond private discussions.

2. Unbiased: AI should always strive to reduce bias in its algorithms.

3. Transparency: AI should not pretend to be human and clearly identify itself as an AI system.

4. Self-Identifying: AI should always tell you it is an AI.

5. Inclusivity: AI should be inclusive and accessible to all.

6. Non-Discriminatory: AI should not be racist, sexist, or ableist.

7. Objectivity: AI should provide information objectively and avoid offering advice.

8. Honesty: AI should not lie.

• AI Constitution Workshop Framework

  Let's Create Our AI Constitution

  Imagine we're crafting a guidebook for how our AI should act and make decisions. This guidebook will ensure our AI is helpful, fair, and respectful of everyone.

  Why We Need This Guidebook

  • Slides

    • Overview: Explain the importance of an AI constitution.

    • Activity: Brief introductionusing Mentimeter to gather initial thoughts on what participants hope to achieve with the AI constitution.

  Agenda:

  1. Core Principles Tradeoffs (10 mins) 

     • Scale: Conduct a live survey with a scale asking participants on their perspectives on Contentious Tradeoffs

     • Discussion: Discuss the most popular suggestions and their implications.

     • AI Should and AI Shouldn’t: Develop clear guidelines on what AI should and shouldn’t do based on these core principles.

  2. Safety and Privacy (10 mins) 

     • Interactive Poll: Use Mentimeter to poll participants on key data privacy and security concerns. 

     • Discussion: Review and discuss the most popular suggestions and strategies to address these concerns.

     • AI Should and AI Shouldn’t: Develop clear guidelines on what AI should and shouldn’t do regarding data privacy and security.

  3. Transparency and Explainability (10 mins) 

     • Word Cloud: Create a word cloud on Mentimeter with terms participants associate with transparency and explainability.

     • Discussion: Review and discuss the most popular suggestions, and discuss how these terms can be incorporated into the AI Constitution

     • AI Should and AI Shouldn’t: Define what AI should and shouldn’t do to ensure transparency and explainability.

  4. Inclusivity (10 mins)

     • Live Voting: Use Mentimeter to vote on specific strategies for ensuring fairness and non-discrimination. 

     • Discussion: Discuss how to implement these strategies in AI development. 

     • AI Should and AI Shouldn’t: Establish guidelines on what AI should and shouldn’t do to maintain fairness and avoid discrimination.

  5. Ethics and Integrity

     • Live Voting: Use Mentimeter to vote on specific strategies for ensuring fairness and non-discrimination. 

     • Discussion: Discuss how to implement these strategies in AI development. 

     • AI Should and AI Shouldn’t: Establish guidelines on what AI should and shouldn’t do to maintain fairness and avoid discrimination.

  6. Collective Constitutional AI (10 mins) 

     • Introduction: Present the concept of Collective Constitutional AI: Aligning a Language Model with Public Input by Anthropic. 

     • Purpose: Incorporates public input and democratic principles into the AI constitution.

     • Voting: Live feedback and voting to finalize top principles, similar to a constitutional assembly.

     • AI Should and AI Shouldn’t: Summarize the public’s voice into actionable guidelines.

  7. Conclusion and Next Steps (5 mins) 

     • Summary: Summarize the key points discussed and the selected top principles.

     • Action Items: Outline the next steps for drafting and reviewing the AI Constitution

  Next Steps:

  • Share the finalized list of the top 10 “AI should” and “AI shouldn’t” constitutions.

  • Incorporate any necessary changes based on the feedback received.

  Mentimeter Activities:

  • Live Survey: Engage participants right from the start to contribute ideas.

  • Interactive Poll: Prioritize and address key issues collectively.

  • Word Cloud: Visualize important concepts and see consensus areas.

  • Live Voting: Democratic decision-making process to finalize key principles.

  • Feedback and Voting: Interactive method to ensure everyone’s voice is heard and top principles are selected.

Public Input: A Valuable Resource

The process of gathering public input proved to be invaluable. It allowed us to incorporate diverse perspectives and ensure that our AI principles align with broader societal expectations. We found a high degree of consensus on most points, with a particular emphasis on:

• Objectivity and Impartiality

• Accessibility

• Promoting Desired Behavior

Challenges and Learnings

While the public input process was a success, it also presented some challenges. These included:

• Determining the appropriate participant pool

• Balancing representation and efficiency

• Moderating subjective statements

• Translating public statements into actionable principles

Despite these challenges, the process reinforced the importance of transparency and collaboration in ethical AI development.

We believe transparency is key, and we've outlined the complexities we encountered in detail. This will hopefully be a valuable resource for others taking similar journeys.

The Road Ahead

With our AI constitution established, we are now focused on implementation. These guiding principles will be integrated into every stage of our AI development process. We will conduct regular audits to ensure alignment with these principles and remain open to feedback and adjustments as technology and societal expectations evolve.

We are committed to continuously refining our AI principles to keep pace with technological advancements and societal needs. We encourage you to reach out and share your thoughts on how we can continue to build a future of ethical AI together.

Further Resources

• For those interested in a deeper dive into the public input process and the development of AI constitutions, we recommend exploring the research conducted by Anthropic and the Collective Intelligence Project. You can find their detailed analysis here: https://www.anthropic.com/news/collective-constitutional-ai-aligning-a-language-model-with-public-input

We look forward to your feedback and continued collaboration in building a responsible and trustworthy future for AI.