Introduction

Generative AI has become an incredibly attractive and widespread tool for people across the world. Alongside its rapid growth, AI tools present a host of ethical challenges relating to consent, security, and privacy, among others. As Generative AI has been spearheaded primarily by large technology companies, these ethical challenges — especially as viewed from the vantage point of ordinary people — risk being overlooked for the sake of market competition and profit. What is needed, therefore, is a deeper understanding of and attention to how ordinary people perceive AI, including its costs and benefits.

The Meta Community Forum Results Analysis, authored by Samuel Chang, James S. Fishkin, Ricky Hernandez Marquez, Ayushi Kadakia, Alice Siu, and Robert Taylor, aims to address some of these challenges. A partnership between CDDRL’s Deliberative Democracy Lab and Meta, the forum enables participants to learn about and collectively reflect on AI. The impulse behind deliberative democracy is straightforward: people affected by some policy or program should have the right to communicate about its contents and to understand the reasons for its adoption. As Generative AI and the companies that produce it become increasingly powerful, democratic input becomes even more essential to ensure their accountability. 

Motivation & Takeaways

In October 2024, the third Meta Community Forum took place. Its importance derives from the advancements in Generative AI since October 2023, when the last round of deliberations was held. One such advancement is the move beyond AI chatbots to AI agents, which can solve more complex tasks and adapt in real-time to improve responses. A second advancement is that AI has become multimodal, moving beyond the generation of text and into images, video, and audio. These advancements raise new questions and challenges. As such, the third forum provided participants with the opportunity to deliberate on a range of policy proposals, organized around two key themes: how AI agents should interact with users and how they should provide proactive and personalized experiences for them.

To summarize some of the forum’s core findings: the majority of participants value transparency and consent in their interactions with AI agents as well as the security and privacy of their data. In turn, they are less comfortable with agents autonomously completing tasks if this is not transparent to them. Participants have a positive outlook on AI agents but want to have control over their interactions. Regarding the deliberations themselves, participants rated the forum highly and felt that it exposed them to alternative perspectives. The deliberators wanted to learn more about AI for themselves, which was evidenced by their increased use of these tools after the deliberations. Future reports will explore the reasoning and arguments that they used while deliberating.
 

Image

^{The participants of this Community Forum were representative samples of the general population from five countries - Turkey, Saudi Arabia, India, Nigeria, and South Africa. Participants from each country deliberated separately in English, Hindi, Turkish, or Arabic.}

Methodology & Data

The deliberations involved around 900 participants from five countries: India, Nigeria, Saudi Arabia, South Africa, and Turkey. Participants varied in terms of age, gender, education, and urbanicity. Because the deliberative groups were recruited independently, the forum can be seen as five independent deliberations. Deliberations alternated between small group discussions and ‘plenary sessions,’ where experts answered questions drawn from the small groups. There were around 1000 participants in the control group, who did pre- and post-surveys, but without deliberating. The participant sample was representative with respect to gender, while the treatment and control groups were balanced on demography as well as on their attitudes toward AI. Before deliberating on the proposals, participants were presented with background materials as well as a list of costs and benefits to consider.

In terms of the survey data, large majorities of participants had previously used AI. There was a statistically significant increase in these proportions after the forum. For example, in Turkey, usage rates increased from nearly 70% to 84%. In several countries, there were large increases in participants’ sense of AI’s positive benefits after deliberating, as well as a statistically significant increase in their interest. The deliberations changed participants’ opinions about a host of claims; for example, “people will feel less lonely with AI” and “more proactive [agents] are intrusive” lost approval whereas “AI agents’ capability to increase efficiency…is saving many companies a lot of time and resources” and “AI agents are helping people become more creative” gained approval. After deliberating, participants demonstrated an improved understanding of some factual aspects of AI, although the more technical aspects of this remain challenging. One example here is AI hallucinations, or rather, the generation of false or nonsensical outputs, usually because of flawed training data.
 

Image

Proposals

Participants deliberated on nineteen policy proposals. To summarize these briefly: In terms of whether and how AI remembers users’ past behaviors and preferences, participants preferred proposals that allowed users to make active choices, as opposed to this being a default setting or only being asked once. They also preferred being reminded about the ability of AI agents to personalize their experience, as well as agents being transparent with users about the tasks they complete. Participants preferred that users be educated on AI before using it, as well as being informed when AI is picking up on certain emotional cues and responding in “human-like” ways. They also preferred proposals whereby AI would ask clarifying questions before generating output. Finally, when it comes to agents helping users with real-life relationships, this was seen as more permissible when the other person was informed. Across the proposals, gender was neither a significant nor consistent determinant of how they were rated. Ultimately, the Meta Community Forum offers a model for how informed, public communication can shape AI and the ethical challenges it raises.

The results show dramatic opinion shifts and notable consensus-building across party lines. Most notably, dissatisfaction with American democracy dropped 21 points overall—from 75% to 54%—with Republicans, Democrats, and independents all showing significant improvement in democratic confidence at the end of the weekend.

"When Pennsylvanians were given the space for informed, civil conversation, they consistently depolarized on issues that dominate cable news narratives as hopeless partisan battles," said James Fishkin, Director of Stanford's Deliberative Democracy Lab. "This experiment proves that America's political divisions and opinions are not as intractable as they might seem. Voters, when presented with balanced information and the opportunity to listen to one another, emerged with considered judgments about what needed to be done as well as greater respect for those they disagree with. The results offer a look at what really matters to voters when they think in depth about the issues. In my view, it also offers an inspiring picture of how democracy could actually work better.”
 

Key findings:
 

• Immigration: Support for increasing visas for low-skilled workers doubled from 25% to 50%, with Democrats moving from 41% to 69% support and Republicans increasing from 9% to 30%. State-level DACA protections gained significant Republican backing, rising from 18% to 38%.
• Voting Rights: Support for broad voter enfranchisement jumped to 96% (up from 83%), with Republicans increasing their support by 22 points. Democrats increased their support for voter ID requirements, increasing from 48% to 57%.
• Election Integrity: An overwhelming majority of participants supported increases in election integrity, with 77% supporting random ballot audits, and 87% supporting criminal penalties for voter intimidation.
• Healthcare: Rural healthcare initiatives achieved near-unanimous support, with 94% backing loan forgiveness for healthcare workers in underserved areas and 88% supporting tax credits for rural facilities.
• Foreign policy: Support for providing military support to Taiwan in case of Chinese invasion doubled from 35% to 69%, with massive bipartisan increases among both Democrats (40-point jump) and Republicans (30-point jump).
• Education: While trade school subsidies gained overwhelming support (81%), free college tuition support dropped from 59% to 47% as participants weighed budget realities.
• Transformed relationships & Understanding: Perhaps most significantly, 91% of participants reported respecting opposing political viewpoints (up from 72%) following their experience at America in One Room: Pennsylvania and 90% expressed willingness to compromise with political opponents (up from 80%). As a whole, 97% of participants reported that A1R: PA was valuable in helping them clarify their positions on key public policy issues debated.

America in One Room: Pennsylvania is the fifth Deliberative Polling® event organized by Helena in collaboration with Stanford’s Deliberative Democracy Lab. Public Opinion Strategies conducted outreach, selected the representative samples, and administered the questionnaires.

Full results and executive summary are available below: