---
tags: vTaiwan
---
# Democratic Input to AI Demo Day

## Some participants info
- Alice Siu: https://stanforddeliberate.org/
- Lama Ahmad; Open AI risk manager

## 1. Case law of AI principle / Policy

- high level of statement about principle v.s. low level repository
- case: potential input to the system
- judment: determination of appropriateness

### Three phase process
- seeding a case repository
- ideal ones should be
- comprehensive
- domain expert / expert workshop
- ? how to choose the expert?
- ? are they able to deal with AI issue?
- clear
- controversial
- refine the case repository
- engage the public and stakeholders
- judge / preturb
- focus on clear / controversial ones
- ? can it fit into different context?
- Turning cases into precedents

## 2. Deliveraive Process using Collective Dialogue

- background: applying the expertise on making cease fire things on AI
- quality / representitive
- collective dialogue
- many polis response
- prompt them
- educate
- how ? what kind of material
- facilitate deliberation
- ? do participants need to have concensus?
- elicit view > bridging responses > policy initiative
- generate demographic
- policy design / intial policy
- expert refinement
- ? why let the expert to participate into the process after the collective dialogue?
- policy refinement
- making it representive
- Human right consistency
- Automated pipeline to translate result to regulation
- Takeaways:
- Collective dialogues permit deliberation at scale
- data driven Democratic policy generation
- scalable deliberative process -> quality AI policy represnting informed public consesus
- AI-enabled tools compressed execution timelines
- Next steps
- Objective
- tackle more contentious issues
- scale to global representiveness
- integrate into existing AI
- Colin
- how to balance the representiveness and expertise is hard thing
- Q2
- how to test? in person
- in person: deliberation part can be/ eliciting part still need to get online
- education?
- exploration / testing / imagine how to teach in the classroom
- hyperthesis: representive things first, then quality

## 3. Aligned by Energize AI

- two question: how to get input global / formning the consensus
- what if we just use google doc?
- some sort of memo
- too messy to get the input
- how do we design the platform to do about the thing?
- Three Principles:
- Simple
- Scalable
- Transparent
- https://app.energize.ai
- human readable / machine readable constitution

Steps:

- propose a guideline
- AI refine the guidelines
- avoid partisan language
- how to turn it into meaningful output / consensus?
- ? can it fit into different context?
- in cooperation of community note / bridging the algorithm
- ? is there any adjustment for minority/ cultural-disadvantaged group or just simple by numbers?
- Q: How to get the number?
- conbination of a deep row to the guidelines
- ID / users
- invited guest
- How are you thinking about transparency and provenance of resulting principles in the public eye? E.g. answering concerns like "I did not vote for these principles!" or "if you change the AI model, would the results be substantially different"?
- fully open / open source

## 4. Democratic fine-tuning

- generating a fine-tuned model / not policy
- hard to make one fine-tuning model under fractional idealogy
- Solution: Align model with values not preferences
- Value is contextual
- Spotting the ideological commitment to underlying value
- People are connected with values
- three phase:
- Articulation part
- when chatting, an user can generate the cards
- Link the value together
- participants will show a story for the context
- Convergence
- the moral graph can be input to ChatGPT to do fine-tuning
- wiser and wiser model of ChatGPT
- ? how to make sure the result is representitive?
- Q:is the moral graph becomeing a cycle ?
- it is possible
- Q: how to deal with the biased opinion?
- Q: representitive stakeholders
- different job
- Q: How to know if the process goes well with the chat and something like that
- the most difficult way is the talk to the chatbot

## 5. Deliberation at Scale

- a web app for deliberation on scale
- Challenge:
- participating the delibration should be as simple as saying hello
- funnel: safe > informed > deliberation > consensus
- FInding " that's right"
- Mapping
- how to test the matrix
- online have been through, and offline is planning
- informed people
- still exploring and divergence on whether to offer the material for that
- when it comes to polarized issues, hard to get consensus, how to deal with

## 6. Proportial Opinion Summaries

- opinion is diversed
- summarization setting
- social choice theory
- LLM can make us do more
- generative: find out the unified statement
- discriminative: making prediction
- process
- statement generating
- participants ranks the statement
- Q: is the process scalable?
- some challenges will happen, maybe it will be expensive

## 7. Bridging the recursive public

## 8. AI dialogue
- rappler:
- the background:
- global south
-

- every layers matters
- inclusion
- diversity
- integrity

- Three parts
- Offline FGD
- for more nuance to happen
- Online FGD
- Particitory survey
- aiDiologue introduced in AI social good summit

- Policy refinement
- policy1: human oversight / liability
- policy2:

- challenge:
- language
- knowledge on AI
- quantifying the coherence / contradiction

- ? Offline FGDs / Online FGDs
- online may make ppl speak something they don't want to speak out face to face

- https://aidialogue.rappler.com/

## 9. Ubintu AI
- Problem: AI can amplify value extration
- technology don't need to extract value, but deign and aligned the value.
- genrative economy for equitable and inclusive model training
- people design and help align the value can grant themselves rewards
-

## 10. Inclusive AI
- inspired by decentralized autonomous organizations
- quadratic voting
-