The Future of Digital Democracy (Audrey Tan / Yutaka Matsuo / Katsuya Uenoyama)
https://scrapbox.io/files/68157cbdc0b7d0e41d8d6582.png
https://gyazo.com/d0f31330ea3cd87c3828906db13fe09d
https://www.youtube.com/watch?v=pI9eXpWi2Bs
Uenoyama: Could the intersection of AI language models and digital democracy enable "previously impossible large-scale collaboration"? +Self-introduction
Mr. Tang.
數位" has both digital and multiple meanings.
I wrote a job description as "Minister of Digital and Minister of Plurality."
AI benefits many people rather than leaving people behind as it becomes more sophisticated
Focus on "horizontal" value distribution, not "vertical" AI sophistication
Mr. Matsuo
D. student, Takahiro Yasuno is also an alumnus, and Ken Suzuki was a member of the review committee when Mr. Suzuki received his Ph. Currently leading the discussion on the Japanese government's AI strategy as the chairman of the AI Strategy Council
Today I want to dig deeper into "How AI will change democracy and industrial society"
Mr. Uenoyama:.
Generation How can AI break the trade-off?
Mr. Tang.
Even with 100-150 people, deep dialogue is possible.
We can see the similarities and differences between them.
Not on the scale of 200,000.
Elections and representation will be necessary.
Plurality's technology could make the conversation flow better.
I want to facilitate facilitators.
Measures against social networking scam advertisements (e.g., solicitation of crypto assets by fake celebrities)
Creating Voices with DeepFake
People who make a living from this kind of advertising on Facebook and YouTube.
o3.iconWhat is the "111" code?
Official SMS Sender ID introduced by the Ministry of Digital Affairs (MODA) of Taiwan in August 2023.
Short messages from government agencies will display the number "111" so that users can instantly identify phishing SMS.
Telecommunication carriers technically protect "111" so that third parties cannot spoof it.
Together with the government URL shortening service (url.gov.tw) and others, it is operated as part of the measures against fraudulent advertisements and fake notifications.
A "mini-public" (mini-public) was created with 450 randomly selected citizens. Online 10 people x 45 rooms for discussion
AI Facilitator to run concurrently
I got a good idea.
Reporting and addressing problems after a Scam occurs slows down the process.
It would be too late to identify who placed the fraudulent ad on Facebook and claim compensation after the ad was placed.
Proposal for a system in which platforms that run malicious advertisements without verifying the identity of the user are liable for compensation ⇒ Passed into law in July with 85 % public support
This is not censorship of ad content, but stricter identification, so even those who fear censorship of ads were convinced
Dramatic decrease in fraudulent ads on Facebook/Youtube
Matsuo: A small group of people discussing and agreeing on something will be convinced, but if "those who were outside" see the conclusion of the discussion, won't they be less convinced? How to extend it to the whole country?
Mr. Tan:
We chose 450 people at random because statistically we can say that it would have been the same if we had chosen 450 other people.
Visualization Tools
SenseMaker (OSS) visualization tool for multi-layered display of individual statements to final agreement
Kentucky Case Studies
Bowling Green?
o3.icon
Bowling Green is a rural community with a population of approximately 70,000 in Warren County in the southern part of the U.S. state of Kentucky. Home to the Chevrolet Corvette assembly plant and Western Kentucky University, it is a fast-growing community with rapidly expanding housing and industry on the northern edge of the Nashville metropolitan area. Why did Mr. Tan mention it?
The city and county conducted a participatory online dialogue with residents to develop a long-term vision through 2050 (BG 2050).
First U.S. pilot case in which over 1 million opinions were collected and semantically analyzed in one month using Google Jigsaw's "SenseMaker" tool. Plurality's testing ground
Combining Polis-type questionnaires and AI facilitation to visualize the "bridging" of consensus while avoiding polarization.
Mr. Tan introduced it as an example of a cross-national pro-social media implementation, saying, "The Taiwanese approach worked in Kentucky.
In short, Bowling Green was cited as a symbolic test bed to demonstrate that "even local cities can co-create urban planning with AI and resident participation.
Show the "consensus-building process."
You can see not only the summary created by the AI, but also what conflicts there were and how agreements were reached.
This image alone is quite powerful.
We may be divided left-right, but we agree on most things.
10 % of participants tell the story through data visualization to 90 % and serve as a bridge.
Matsuo: Why now?
Retrospective of his 2009 Twitter social sensor study.
More than 5,000 citations were received, but politicians were not interested at the time.
Are things changing?
Mr. Tan:
SNS was the "follow feed" of 2009
Twitter was a microblog, and when you followed people, their feeds streamed to you.
For "for you feeds" in 2015 and beyond
Algorithms now determine feeds based on personal behavioral data.
Algorithms to increase user engagement have led to a boosting of extreme opinions and a worsening of polarization.
This may not be well understood in Japan, but in the U.S. and Europe, this is recognized as a serious problem. Politicians are beginning to look for ways to fix it.
Different from horizontal credit mode (? Not sure what you said in English)
More extreme voices than commonalities attract buzz and attention.
This is the difference between 2009 and 2015
Mr. Uenoyama: Any hints on how to go beyond ideology and implement this as real software?
Mr. Tan:
In 2015, Taiwan experimented with a new social media, and it's Polis. There is a scoreboard, a bridge bonus, a mechanism that draws attention to "bridging opinions" that are favored by different opinion goups.
This is the opposite of a polarization algorithm
The opinion that took first place in the three-week discussion was confirmed by the cab union and Uber jointly, and became legislation.
x.com's Community Notes, Facebook and YouTube are trying to use something similar The main feed is now composed of such pro-social algorithms Mr. Uenoyama: LLM eliminates the boundary between word and code. I think the discontinuous change is that we can now calculate the "difference from the average opinion" in addition to summarizing, and for the first time in human history, dialogue beyond human cognitive limits will be possible. I think we can calculate statistical correctness (assuming, of course, that it is not the truth), what do you think?
Mr. Tan:
F511 and other Socratic interactive LLMs allow constructive dialogue with conspiracy theorists.
Ask the short question, "What made you think it was a conspiracy?"
'What experience did you have with the news that made you feel it was authentic?' What elements of the news made it persuasive?" ->Elicit the person's cognitive factors and encourage self-examination without imposing conclusions.
Even if it's fake news, there's room for conversation.
GPT-4 Turbo-based interactive LLM used in the Costello/Rand et al. experiment published in Science, September 2024.
AI summarizes "evidence of belief" presented by subject (conspiracy theorist) → Prompts reconsideration with Socratic questioning and presentation of evidence.
After 3 round-trips of less than 8 minutes, the average belief strength was -20%, and 1 out of 4 people switched to the "no-belief" side. The effect lasted at least 2 months.
Origin of name : Internal version number of the research team (written Fine-tune 5.1 => F5.1 => F511). It is not a public API or product name, and can only be found in papers and GitHub repositories. General use is being offered on a trial basis at debunkbot.com.
However, accountability and elimination of cultural bias are essential.
With a closed model, there could be bias, it can't be verified.
We need to be able to see inside the model, like an MRI.
Cultural factors
Even if it could be translated, would it be a cultural match?
Mr. Uenoyama: I see what you mean when you say, "I think it is possible to realize a system that is different from the collaboration and intention we have had up to now.
Matsuo: If AI were to mediate optimal communication, what form would it take?
Mr. Tan:
AI should be "assisted".
Can be used in an authoritarian way, images from surveillance cameras can be uploaded to the cloud -> reinforcement of power
Direction to share spreadsheets over the Internet instead of strengthening isolated mainframes.
Creating a situation where citizens can resist centralized power.
Mr. Uenoyama: Isn't there a group assist as well as an individual assist? Wouldn't assisting a group exceed the cognitive limit when considered on an individual basis?
Mr. Tan:
Going beyond 150 people is WELCOME!
What's not good.
Individuals interact individually and only with bots -> brainwash.
Augmented intelligence without groups, not a good example
Democratic muscular strength declines in a state of "leaving muscle training to the robots".
Mr. Uenoyama: Do you have any hypotheses about the best future communication?
Matsuo: If radical opinions disappear, won't everyone eventually move to the middle and diversity will disappear?
Mr. Tan:
Cross Cut Divide
Bridging opinion groups A and B
Some cross the bridge, some don't.
There will be three groups: those who stayed in A, those who crossed the bridge, and those who stayed in B.
We are making a new "difference".
Eliminating indifference, creating curiosity
If the differences are too great, they become indifferent to each other.
Matsuo: How do you see the industrial spillover of "Physical AI" such as robotic infrastructure models?
Mr. Tan:
Connecting to the real world promotes multidimensional growth of AI
The more we interact with people of diverse ethics in the real world, the more we can train a "model of pluralistic growth.
Increased evolutionary pressure
Ongoing interaction with the real world is important.
I am optimistic.
Might not be common in the West.
East Asia is a "Doraemon" culture, not a "Terminator" culture
Robot view is basically cooperative = soil for social implementation.
Moderated by.
Isn't there a limit to governance in the existing democratic system based on nation-states as units?
If AI competition and data domination proceed without sufficient international cooperation, there is a danger that a small number of supranational players will gain new control.
How should the democratic framework evolve/transform to minimize risk while maximizing benefits?
Mr. Tan:
USB-C Mandatory Example
Prohibiting proprietary connectors will not reduce innovation.
Rather, new entrants who "don't have to pay license fees" can compete on charging quality.
This is not "anti-competitive" but rather increases competition.
The model of enclosure by proprietary specifications can be broken down and replaced by a competition of quality and functionality on a common interface that can be interoperated with each other.
Portability is important
Utah Digital Choice Act: When migrating from TikTok to Bluesky, users are required to transfer their posts, followers, and DMs in their entirety. Same concept as mobile number portability. This is the key to coexistence of transnational competition and cooperation.
Proposes that "mandatory data & account portability" will combine innovation and international cooperation.
Facilitates moving to new services
"Too much to lose, too much to move" lock-in is eliminated.
As a result, competition is based on the superiority of protocols (data formats and APIs), not platforms.
Mr. Matsuo:
Social systems have many hyper-parameters, we need to experiment. Do we do that experimentation on a country-by-country basis?
The issue of digitally illiterate people holding much of the table.
Yasuno-kun is starting a new political party and I really want him to do well, but we'll see what happens.
Mr. Uenoyama:
Consider what the private sector can do.
One superintelligence can't be for every industry, so it must first be decentralized.
Becoming embedded in the application layer
The app layer can embed the developer's intentions and values, which can change the social impact depending on the design.
Current social media is about "maximizing ad revenue" and the human brain is being hacked.
Designs that shift the objective function to the mid- to long-term and public are possible.
The majority of social problems are communication problems where "people cannot communicate well with each other.
Inability to communicate with many people is a problem.
There is room for great improvement if you throw in an AI that can handle the language.
It is the role of private companies and engineers to actually shape these applications.
Audrey Tan has been very busy, this is the second time this week.
https://pbs.twimg.com/media/GqvnfPBWMAAIlFi?format=jpg&name=small#.png
---
This page is auto-translated from /nishio/デジタル民主主義の未来(オードリータン/松尾豊/上野山勝也) using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I'm very happy to spread my thought to non-Japanese readers.