AI and Democracy
People welcome and trust AI, often more than their institutions, but autonomous AI that makes decisions for people is the outlier.
Public assessments track closely to how much independent power a system holds. Advisory AI is embraced; AI that acts on its own is not.
- Benefits vs. Risks: 49.7% say the benefits of AI chatbots outweigh the risks, a more favorable verdict than social media apps receive; but 53.8% say risks outweigh benefits for AI that acts in the real world without supervision, and 46.8% say the same for AI that outperforms humans on most economically valuable work.
- Calibrated Trust: At 55.9%, chatbots are trusted well above elected representatives (32.6%) and large corporations (35.5%), but remain below doctors (86.7%) and public research institutions (70.9%), suggesting trust that is high for advice and low where autonomy is involved.
- Autonomy Stays Rare: 56.9% have never had an AI system complete a real-world action for them without supervision, even as daily use becomes routine.
Help, Not Authority
The welcome is for assistance, and the line hardens as the stakes rise
Given a choice, people prefer an AI helper to a human helper across every domain tested, while refusing to let anyone, human or AI, decide for them.
- AI Help Leads Everywhere: The AI-assisted model leads the human one on local services (46.7% vs. 35.8%), spending local taxes (47.8% vs. 28.0%), permits and licenses (45.4% vs. 27.2%), and online content (40.9% vs. 22.2%), by margins of roughly 11 to 20 points.
- Seen as Fairer: Asked who would weigh their interests more fairly on a decision affecting daily life, 39.8% chose "an AI system, consistent and treats everyone by the same rules" over the people who currently make those decisions (30.8%).
- Contested Decisions Reverse the Pattern: For a contested local consultation, 59.4% prefer a longer in-person meeting where people debate together over a fast AI-driven answer (33.0%), and open-ended responses are emphatic that humans should decide community matters.
The Limits of Delegation
Human review expands acceptance, but never for coercive or foundational decisions
When a human reviews and approves every decision, people accept AI in many domains. That tolerance stops sharply at the exercise of coercive authority.
- Review Raises Acceptance: With a human approving every decision, 58.2% accept AI involvement in content removal, 49.5% in public-sector hiring, and 46.5% in welfare decisions.
- The Floor Holds: Acceptance stays low for asylum decisions (34.4%), deciding the rules everyone lives under (28.9%), and criminal sentencing (24.4%); 19.6% reject AI in all of these even with review.
- Speak With, Not For: 79.5% would use AI to summarize long civic documents and 75.6% to translate them, but only 26.6% would let it represent their arguments and 9.5% would have it attend a meeting in their place.
A Public That Feels Unheard
The dominant barrier to participation is futility, not access or complexity
AI meets a population that already doubts its own civic voice, and the constraint it runs into is motivational rather than logistical.
- Futility First: 48.7% name "I don't think my view will change anything" as what most stops them from participating, far ahead of not knowing decisions are happening (26.0%) or not understanding the details (25.7%).
- Thin Participation: 25.9% have never acted to make their views known in the past year, and 38.5% only once or twice; in the venue that matters most to them, 45.8% say their opinion matters "a little" or "not at all."
- Government Looms Largest: 56.4% rank national government first among those who make the most important decisions affecting their community, and 40.1% say equal treatment is what would most earn their trust, ahead of efficiency (25.1%).
AI Cuts Both Ways in Civic Life
A personalized report raises willingness to participate, but contested decisions send people back to each other
The same public wary of AI in governance responds strongly to one concrete application, while insisting that genuine disagreement be worked out among humans.
- The Report Effect: 77.2% say they would be more likely to take part in a community decision if an AI sent them a personalized report on how it would affect their household, a large effect that runs against the general wariness about AI in civic settings.
- Verify It Themselves: Judging whether online information is false, 74.9% chose an option where they check it themselves (a search tool or an AI assistant that helps them verify) over an official body or official AI making the determination (21.4%).
- Responsibility Lands on Humans: When an AI wrongly denies a benefit, people assign responsibility to shared parties (34.4%) or the approving official (34.0%) far more often than the AI company (6.1%), and the pattern is identical whether the error harmed or benefited them.
Trust in the Method Itself
People would trust an AI summary of public views, and very few want that summary to decide anything
This round tested the Global Dialogues method on its own terms, and the result both validates and bounds it.
- Trusted to Summarize: 77.4% would trust an AI summary of public views on a contested issue at least somewhat, and only 4.1% would not trust it at all.
- Useful, Not Authoritative: 38.0% say AI should be used to find common ground whenever useful (11.4% say never), but only 13.5% would make that common ground the basis for new policy; the preferred roles are informative only (46.2%), one input among several (43.2%), and a starting point for human debate (41.7%).
- Capability Beats Provenance: With equal capability, 46.0% prefer a public, community-built AI over a commercial one (32.9%), but the preference fully reverses if the public AI is less capable, with only 23.6% trusting it more and 43.6% trusting it less.
Where the World Agrees
The strongest cross-group consensus is about the conditions for trusting AI
Disagreement in this round is modest. Where the public converges hardest is on transparency, and the agreement holds across every demographic segment measured.
- A Consensus Ceiling: Every one of the top bridging responses in the round came from a single question, "What would the AI need to do for you to trust it to summarise public views fairly?"
- Transparency Is the Price: The most broadly endorsed response, calling for an open and trusted source of data, was agreed with by at least 67% of every one of 79 valid segments, followed by showing its process (0.66) and providing verifiable sources (0.65).
- Humans Decide: Explaining who should handle a public consultation, the most bridging response was that humans, as the ones affected, should make the decision rather than an AI that is not affected at all, agreed with by at least 45% of every group.
Conclusion
The findings from this Global Dialogue describe a public defined by conditional trust. People have folded AI into daily life and trust it, in many cases more than the institutions meant to represent them, but that trust is calibrated with care. It is granted freely for help, understanding, and consistency, and withheld firmly from autonomy, coercive authority, and contested judgment. This is a population that already feels unheard in civic life and is open to AI that lowers the barriers to participation, provided humans keep the final say and the process stays transparent. The clearest mandate in the data is not for AI decision-makers but for AI that makes democratic participation more legible, more personal, and more worth the effort, while leaving the deciding to people.
Methodology
This Global Dialogue surveyed approximately 1,100 participants (per-question n ranging roughly 1,000 to 1,060) on the role of AI in civic and democratic life, drawing on both closed-ended items and open-ended responses.
Cross-group consensus was measured using bridging scores. For each response to an open-ended question, we measured agreement within every demographic segment of the sample (every age group, gender, region, religion, and attitude-toward-AI group) and took the bridging score to be the agreement rate in the least supportive segment. A bridging score of 0.67 therefore means that in every group measured, at least 67% of people agreed, with most groups agreeing at higher rates. It is a deliberately conservative measure: a single dissenting group pulls the score down, so high-scoring statements represent genuine cross-group consensus rather than majority opinion masking pockets of rejection.
