Andrew Sorota and Josh Hendler, who lead AI and democracy work at the Office of Eric Schmidt, published a blueprint on May 5, 2026 arguing that AI is becoming the primary interface through which citizens form beliefs and act politically — and that US institutions are not designed for it. Their thesis runs across three layers: how people know things, how they act on what they know, and how they govern collectively. On all three, they argue the design choices being made now will determine whether AI strengthens or hollows out democratic life.
The first layer is epistemic. Search is already substantially mediated by AI, and the next generation of assistants will synthesize, frame, and present information with authority. Sorota and Hendler put the stakes plainly: "Whoever controls what these models say therefore has increasing influence over what people believe." For a growing share of voters, asking a model will become the default way to evaluate a candidate, a ballot measure, or a public figure.
The second layer is agentic. Personal AI agents will soon draft communications, lobby on a user's behalf, and decide which causes to surface. The authors warn that an agent shaped around a user's anxieties is structurally similar to social-media algorithms that optimized for engagement — except the agent presents itself as the user's advocate, which makes the failure mode harder to detect.
Key facts
- 01Andrew Sorota and Josh Hendler, who lead AI and democracy work at the Office of Eric Schmidt, published the blueprint on May 5, 2026.
- 02The authors identify three layers under pressure: epistemic (how citizens know), agentic (how they act), and institutional (how they govern).
- 03A field evaluation of AI-generated fact checks on X found cross-partisan readers rated AI-written notes more helpful than human-written ones.
- 04Bots are already skewing public-input processes, the authors warn, urging identity verification for humans and agentic proxies.
- 05Several US states and localities are running AI-mediated platforms for democratic deliberation at scale.
They are blunt about what counts as faithful representation. "An agent that refuses to present uncomfortable information, that shields its user from ever questioning prior beliefs or fails to adjust to a change of heart, is not acting in the person's best interest." That is a high bar, and one that current consumer assistants do not clearly meet.
“A public sphere in which everyone has a personalized agent attuned to their existing views is not, in aggregate, a public sphere at all.”— Jaeden Schafer
The third layer is institutional. Sorota and Hendler note that AI agents and humans will soon participate in the same forums, often indistinguishably. Even well-aligned individual agents can produce collective biases at scale, they argue, and they cite existing evidence that bots are already skewing public-comment and rule-making processes. Their fix: identity verification for both humans and agentic proxies, built in from the start.
The piece is not uniformly pessimistic. The authors point to a field evaluation of AI-generated fact checks on X in which readers across political viewpoints rated AI-written notes more helpful than human-written ones. The paper is not yet peer-reviewed, but the result suggests AI-assisted fact-checking may achieve the kind of cross-partisan credibility manual efforts have struggled to reach. Greater transparency about how models prioritize sources, they argue, could deepen that trust.
On governance, they note that several US states and localities are already running AI-mediated platforms for democratic deliberation at scale, drawing on research that AI mediators can help citizens find common ground. The implicit comparison is to the broadcast era, when shared national audiences underwrote mass democracy — and to the social-media era, when algorithmic feeds fragmented it. AI agents, in their telling, could go either way.
Their warning about aggregation is the sharpest line in the piece. "A public sphere in which everyone has a personalized agent attuned to their existing views is not, in aggregate, a public sphere at all." Each user's experience can be internally coherent and still produce a collective environment hostile to shared deliberation. That is a different problem from misinformation, and it is not solved by better truthfulness alone.
The blueprint is short on enforcement detail. Sorota and Hendler call on AI companies to ramp up truthfulness work and on policymakers to build identity verification into civic infrastructure, but they do not propose specific statutes, agencies, or compliance regimes. The Office of Eric Schmidt is a policy shop rather than a regulator, and the piece reads as agenda-setting rather than legislative text. Critics will note that voluntary commitments from model providers have a mixed track record, and that the firms with the most influence over the epistemic layer — OpenAI among them — face commercial pressure to keep agents engaging rather than challenging.
Still, the framing matters. Most AI-policy writing in 2026 has focused on catastrophic risk, copyright, or labor displacement; democratic infrastructure has been a smaller share of the conversation despite a US election cycle that will run through agentic tools most voters have never used before. Sorota and Hendler close with the observation that failing to design for democratic outcomes in a domain this consequential means designing for something else — and that the history of unaccountable power offers little reassurance about what that something else becomes.
The practical question for AI labs is whether any of this changes product roadmaps. Faithful representation, refusal to flatter the user, and verifiable agent identity are all expensive features that cut against engagement metrics. Whichever model provider treats them as a moat rather than a tax will have an unusually strong claim on the civic layer of the next decade — and the rest will keep optimizing for time-on-app while the institutions around them keep fraying.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
