## Fostering More Inclusive Democracy with AI

## Source details

**Canonical URL:** [Fostering More Inclusive Democracy with AI](https://www.imf.org/-/media/files/publications/fandd/article/2023/december/pov-landemore.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/fandd/article/2023/december/pov-landemore.pdf.md)
- [Structured JSON version](/-/media/files/publications/fandd/article/2023/december/pov-landemore.pdf.json)

---

### Framing: risks and opportunities
- Public fears: AI may undermine democracy by taking away jobs, destabilizing the economy, widening the divide between the rich and the poor, concentrating power in a few tech companies, weakening government regulatory structures, and delegating human decision-making to machines (algocracy).
- Counterpoint: Societies currently have political ability and responsibility to shape AI development and the technological opportunity to harness AI to enhance democracy and collective governance rather than merely regulate AI.

### Participatory experiments in democracy
- OECD report: found close to 600 cases in which a random sample of citizens engages deeply with an issue and formulates informed policy recommendations.
- Mass-participation examples:
  - Participatory constitutional processes in Brazil, Kenya, Nicaragua, South Africa, and Uganda (1980s and 1990s); more recently Chile, Egypt, and Iceland using mass consultations and crowdsourcing.
  - France’s 2019 Great National Debate: 1.5 million participants.
  - EU Conference on the Future of Europe: 5 million people visited the website and 700,000 engaged in debate.
- Observation: Most such large campaigns have been low-tech and analog; politicians have sometimes ignored citizens’ input because of being overwhelmed by raw, multifaceted data or unsure of its meaning. Result: people were allowed to speak but were not always heard; deliberation was often superficial.

### Enhanced deliberation through AI
- Taiwan Province of China case:
  - pol.is platform allows elaborate opinion expression and voting on others’ opinions; maps opinion landscapes, identifies consensus, minority and dissenting opinions, and lobbyist groupings to reduce polarization.
  - pol.is evolved to integrate machine learning and an LLM that speaks for different opinion clusters; used to consult with residents, engaging 12 million people, or nearly half the population.
- Corporate experiments:
  - Meta Community Forums (2022): randomly selected users deliberated on climate content regulation.
  - December 2022 Meta experiment: 6,000 users from 32 countries in 19 languages discussed cyberbullying in the metaverse over several days; deliberations were facilitated on a proprietary Stanford University platform by (still basic) AI that assigned speaking times, helped decide topics, and advised when to pause topics.
- Scalability: AI facilitators are not yet proven superior to humans but will be much cheaper, which matters for scaling deep deliberative processes from thousands to millions of participants.

### Translation, summarization, and analysis
- Research finding: AI is 50 percent more accurate than human beings when it comes to summarization (evaluated by trained undergraduates comparing AI summaries and human coders’ summaries of deliberation transcripts).
- Applications envisioned:
  - Instantaneous translation among multilinguistic groups.
  - Summarization of collective deliberations.
  - AI as aid to human analysts, facilitators, and translators where human judgment remains necessary.
- Initiatives: OpenAI’s Democratic inputs to AI grant program subsidized the 10 most promising teams developing algorithms that serve human deliberation.

### Addressing risks and inclusion challenges
- Identified risks: data bias, privacy concerns, potential for surveillance, legal challenges, digital divide, exclusion of illiterate and techno-skeptical groups.
- Technological mitigations and supports:
  - Zero-knowledge protocols (ZKP) to verify or “prove” identity without collecting participant data (examples: text messaging authentication or blockchain); applicable to online voting, sharing sensitive information, or whistleblowing in deliberative contexts.
  - Generative AI to expand access to scarce knowledge and tutoring resources; serve as custom-tailored interlocutor to explain technical policy issues in individuals’ cognitive styles (including through images) and convert oral input into written input.
- Political, economic, legal, and social interventions will be required in addition to technological fixes.

### Institutional and policy implications
- Thesis: AI has the potential to produce a better, more inclusive democracy that equips governments with legitimacy and knowledge to oversee AI development.
- Claim: “AI regulation is likely to be better enforced and more effective in AI-empowered democracies.”
- Urgent needs:
  - Investment in AI tools that safely augment participatory and deliberative capacities of governments.
  - Attention to scaling, accessibility, and safeguards so democracy itself is not a casualty of the AI revolution.

*Source: Fostering More Inclusive Democracy with AI — Hélène Landemore, DECEMBER 2023, F&D*

---


_Source: https://www.imf.org/-/media/files/publications/fandd/article/2023/december/pov-landemore.pdf_