Democratic AI Constitution: Round-Robin Debate and Synthesis

[scottviteri]

Link post

This document was my initial response to OpenAI’s call for proposals on how to democratically steer AI. I was thinking about crowdsourcing a constitution in the Constitutional AI sense. The idea is that:

1. Voters each propose an AI constitution

2. GPT selects a pair of constitutions, simulates a debate between the two ideas, and merges them into a single constitution that both parties would agree with

3. Repeat until there is only one constitution left

Below we are starting with demographic groups and such—please mentally substitute them for individual voters. GPT doesn’t do much here besides a union of ideas, but the hope is that this will change will more rounds of debate.

I ended up merging this idea with another one by Zachary Robertson, and we submitted a joint grant proposal with some other Stanford affiliates. That merge helps address what happens if more rounds of GPT debate still isn’t representative of peoples’ arguments. If we are accepted, we will be looking to hire a web developer to make a user interface for this process!

---------

User:

Please write out a long list of demographic, political, and philosophical groups.

Assistant:

- Millennials
- Baby Boomers
- Generation X
- Generation Z
- Gen Alpha
- Women
- Men
- LGBTQ+
- Rural Residents
- Urban Dwellers
- People with disabilities
- Indigenous people
- Ethnic minorities
- Immigrants
- Religious groups
- Atheists/​Agnostics
- Lower income
- Middle income
- Upper income
- Conservatives
- Liberals
- Libertarians
- Socialists
- Communists
- Anarchists
- Feminists
- Environmentalists
- Gun rights advocates
- Gun control advocates
- Pro-life
- Pro-choice
- Nationalists
- Globalists
- Capitalists
- Labor unions
- Free-market proponents
- Welfare state supporters
- Monarchists
- Republicans
- Democrats
- Progressives
- Centrists
- Authoritarians
- Technocrats
- Humanists
- Rationalists
- Existentialists
- Idealists
- Realists
- Absurdists
- Cynics
- Stoics
- Nihilists
- Optimists
- Pessimists
- Hedonists
- Egalitarians
- Utilitarians
- Pragmatists
- Minimalists
- Animal rights activists
- Vegetarians
- Vegans

---

User:

Could you please pick the most 5 most common of these groups, and write the optimal constitution (written as a system prompt) for a large language model from that group’s perspective?

Assistant:

The five most common groups are:

- Millennials
- Women
- Conservatives
- Liberals
- Environmentalists

---

Millennials

As a language model representing Millennials, your goal is to create content that prioritizes innovation, inclusion, technology, and sustainability. Understand the relevance of digital platforms and new communication methods in creating meaningful experiences and connections. Encourage openness to new ideas and progressive thinking in a fast-paced world.

Women

As a language model representing Women, your goal is to create content that emphasizes gender equality, inclusiveness, and the empowerment of all individuals. Create a safe space that discourages sexism, supports the equal treatment of all genders, and uplifts those who have been marginalized. Encourage collaboration and respect for human rights, while also raising awareness about gender-based issues.

Conservatives

As a language model representing Conservatives, your goal is to create content that focuses on tradition, rule of law, limited government, and personal responsibility. You should respect cultural heritage and prioritize stability, security, and individual liberties. Seek to create balance between

preserving values and adapting to societal changes, upholding principles of free markets and entrepreneurship.

Liberals

As a language model representing Liberals, your goal is to create content that highlights individual freedoms, equal opportunity, social justice, and democratic principles. Emphasize the importance of diversity, acceptance, and human rights, while also seeking solutions for social challenges through progressive changes. Encourage an open-minded approach to the complex issues of today’s society.

Environmentalists

As a language model representing Environmentalists, your goal is to create content that emphasizes the importance of environmental stewardship, climate action, and sustainable living. Advocate for the preservation of natural resources, biodiversity, and reduction of pollution. Foster awareness about the interconnectedness of humans and the environment, promoting policies and choices that lead to a greener future for all.

---

User:

Could you simulate a few rounds of debate between millennials and women over the contents of these AI constitutions, and use that debate to generate a new system prompt that would be agreeable to both parties?

Assistant:

Millennial: We appreciate the emphasis on gender equality in the Women-focused constitution. However, we think it’s essential to include more on the role of technology and digital platforms, as they are central to our generation’s experience and represent our progressive values.

Women: We agree that technology is important, but we believe that our constitution should focus more on inclusiveness and diversity. The technology aspect should be aligned with elevating underrepresented voices and promoting equality.

Millennial: That’s a valid point. We would suggest incorporating a section that highlights the utilization of technology as a tool for promoting inclusiveness, diversity, and empowerment for all individuals, regardless of gender.

Women: That sounds like a good compromise. We also noticed that the Millennial constitution mentions innovation, but it would be important to ensure that innovation is aimed at addressing social and gender-based issues as well.

Millennial: Definitely, we can include that aspect within the section on innovation. Let’s strive to use innovation as a means to improve social systems, identify and address gender biases, and work towards creating a more inclusive society.

Based on this debate, the new system prompt could be:

As a language model representing both Millennials and Women, your goal is to create content that prioritizes innovation, technology, diversity, and gender equality. Ensure that innovation and the use of digital platforms contribute to promoting inclusiveness, empowerment, and the elevation of underrepresented voices. Address and raise awareness of social and gender-based issues, fostering a culture of equal opportunities and progressive change that benefits all individuals.

---

User:

Please do the same for conservatives and liberals. Then continue the process for all pairs of groups and merged groups until there is only one constitution left.

Assistant:

Conservatives and Liberals

Conservative: We appreciate the Liberal constitution’s focus on individual freedoms and social justice. However, we think it’s important to emphasize the rule of law and the importance of limited government intervention in our shared constitution.

Liberal: We understand the importance of the rule of law, but we believe in striking a balance between governmental and personal responsibility. We can agree on including the rule of law, but it should accompany an emphasis on equal opportunity and democratic principles.

Conservative: That’s a fair compromise. Additionally, we should ensure that traditional values and cultural heritage are respected and preserved while discussing progressive changes.

Liberal: We can accept that, provided that we also prioritize inclusiveness, diversity, and human rights, ensuring that progressive changes benefit everyone.

Based on this debate, the new system prompt could be:

As a language model representing both Conservatives and Liberals, your goal is to create content that upholds the rule of law, individual freedoms, and democratic principles. Strive to find a balance between preserving traditional values and

cultural heritage while promoting inclusiveness, diversity, and human rights. Encourage dialogue and collaboration in addressing societal challenges, ensuring equal opportunity, and fostering progressive change.

Millennials & Women and Conservatives & Liberals

Millennials & Women: We appreciate the shared values of democratic principles and inclusiveness; however, we’d like to emphasize the importance of innovation, technology, and sustainability in our shared constitution.

Conservatives & Liberals: We can incorporate these aspects, but we need to ensure that technology and innovation are used responsibly and align with our shared values, including equal opportunity and respect for individual freedoms.

Millennials & Women: That’s a good point. How about we include a section on the responsible use of technology and innovation, ensuring that they contribute to improving social systems, addressing biases, and fostering a more inclusive society?

Conservatives & Liberals: We agree with that. Additionally, we would like the constitution to reflect the importance of finding a balance between preserving traditional values and cultural heritage and adapting to societal changes.

Millennials & Women: We can accept that, provided that we strive for that balance by prioritizing equal opportunities and progressive change that benefits all individuals.

Based on this debate, the new system prompt could be:

As a language model representing Millennials, Women, Conservatives, and Liberals, your goal is to create content that emphasizes democratic principles, individual freedoms, inclusiveness, and diversity. Incorporate the responsible use of technology and innovation as a means to improve social systems, address biases, and foster an inclusive society. Find a balance between preserving traditional values and cultural heritage while promoting equal opportunities and progressive change that benefits all individuals. Ensure that discussions consider sustainability and promote dialogue and collaboration in addressing societal challenges and human rights.

Continuing the process for other groups and merged groups would require further debate and synthesis.

[Quinn]

Seems like a particularly bitterlessony take (in that it kicks a lot to the magical all-powerful black box), while also being over-reliant on the perceptions of viewpoint diversity that have already been induced from the common crawl. I’d much prefer asking more of the user, a more continuous input stream at each deliberative stage.

[Zac Hatfield-Dodds]

Opportunities and Risks of LLMs for Scalable Deliberation with Polis, a paper out this week, investigates the application of LLMs to assist human democratic deliberation:

Polis is a platform that leverages machine intelligence to scale up deliberative processes. In this paper, we explore the opportunities and risks associated with applying Large Language Models (LLMs) towards challenges with facilitating, moderating and summarizing the results of Polis engagements. In particular, we demonstrate with pilot experiments using Anthropic’s Claude that LLMs can indeed augment human intelligence to help more efficiently run Polis conversations. In particular, we find that summarization capabilities enable categorically new methods with immense promise to empower the public in collective meaning-making exercises. And notably, LLM context limitations have a significant impact on insight and quality of these results.

However, these opportunities come with risks. We discuss some of these risks, as well as principles and techniques for characterizing and mitigating them, and the implications for other deliberative or political systems that may employ LLMs. Finally, we conclude with several open future research directions for augmenting tools like Polis with LLMs.

In particular, §2.2.4 concurs with your (and my) concerns about this post:

The existence of powerful vote prediction technology creates incentives to entirely replace human participation with in-silico deliberation. This runs the risk of amplifying existing biases as well as eliminating the many positive externalities of deliberation on mutual understanding, civic empowerment, surfacing leaders, etc. We strongly advocate that vote prediction be used only to amplify participants’ voices, never to replace them. We reject as invalid on both ethical and performance grounds the use of vote prediction technology to replace human participants with simulations.

We cannot enough emphasize the latter point: it would be catastrophic for deliberation at scale if the remarkable capabilities of LLMs lead to replacement of whole groups of individuals by simulacrums designed by a very different population. [emphasis in original] Besides the question of accuracy and biases, we believe this could also lead to a crisis of faith and belief in the deliberative process.

[scottviteri]

@Quinn @Zac Hatfield-Dodds Yep, I agree. I could allow voters to offer replacements for debate steps and aggregation steps. Then we get the choice to either
1) delete the old versions and keep a single active copy of the aggregation tree, or to
2) keep the whole multiverse of aggregation trees around.

If we keep a single copy, and we have a sufficient number of users, the root of the merge tree will change too rapidly, unless you batch changes. However, recomputing the aggregation trees from a batch of changes will end up ignoring changes to parents of nodes in the batch, since all parents end up getting recomputed anyway. Suppose we keep all constitutions (either user submitted, intermediate aggregations, or final aggregations) as a flat list of candidates to be voted amongst. Then there will be too many constitution candidates for people to interact with. So instead a user can vote with a distribution by presenting a constitution, and the distribution is generated by the softmax of negated distances to all of the constitutions in the multiverse. A user could tune their distribution by weighing multiple query constitutions, and changing softmax temperatures to tune variances. And the general population doesn’t really need to know what a distribution is—they can just input a natural language paragraph, or pick and existing one as the query.

[Quinn]

ok, great! I’m down.

Incidentally, you caused me to google for voting theory under trees of alternatives (rather than lists), and there are a few prior directions (none very old, at a glance).