Andrew Sorota and Josh Hendler, who lead AI and democracy work at the Office of Eric Schmidt, argued on May 5, 2026 that AI is on track to become the primary interface through which Americans form political beliefs and exercise civic agency — and that US institutions are not built for it. Their essay frames the shift as comparable to the printing press, the telegraph, and broadcast media, each of which restructured how citizens interact with information and, eventually, how they govern themselves.
The authors organize the problem into three layers. The epistemic layer is how people come to know things, increasingly mediated by AI search and assistants that synthesize and frame information with authority. The agentic layer is how people act, as personal agents draft communications, choose causes to support, and respond to government notices on a user's behalf. The institutional layer is the public sphere itself, where humans and agents now share the same forums.
On the first layer, Sorota and Hendler write that "whoever controls what these models say therefore has increasing influence over what people believe." Search is already substantially AI-mediated, and the next generation of assistants will be the default route for forming views on candidates, policies, and public figures. The leverage that implies — over a population, not a product category — is the throughline of the piece.
Key facts
- 01Andrew Sorota and Josh Hendler, who lead AI and democracy work at the Office of Eric Schmidt, published the argument in MIT Technology Review on May 5, 2026.
- 02The authors flag three layers under pressure: how citizens form beliefs, how they act through agents, and how institutions run public deliberation.
- 03They cite a field evaluation on X finding cross-partisan readers rated AI-generated fact checks more helpful than human-written notes; the paper is not yet peer-reviewed.
- 04They warn agents with no individual bias can still generate collective biases at scale when millions interact in the same forums.
- 05Their prescription includes identity verification for humans and agent proxies in public input processes, where bots already skew results.
On the second layer, the authors compare personal agents to the engagement-optimized algorithms that shaped a decade of social media, with one twist: an agent earns trust by presenting itself as the user's advocate. They argue an agent "must never have an agenda of its own or misrepresent its user's views — a technically daunting requirement in domains where users may have not explicitly stated any preferences." Faithful representation also cannot mean shielding the user from uncomfortable facts.
“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 collective effect is the part most likely to surprise readers. Sorota and Hendler cite research showing that agents displaying no individual bias can still generate collective biases at scale when millions of them interact. "A public sphere in which everyone has a personalized agent attuned to their existing views is not, in aggregate, a public sphere at all," they write. "It is a collection of private worlds, each internally coherent but collectively inhospitable to the kind of shared deliberation that democracy requires."
Their prescription on the epistemic layer is the most concrete. The authors point to a field evaluation of AI-generated fact checks on X, in which readers across the political spectrum rated AI-written notes more helpful than human-written ones. The paper has not been peer-reviewed, but if it holds, AI-assisted fact-checking could clear a credibility bar that manual efforts have repeatedly failed. Transparency about how models prioritize sources, they argue, is the missing piece.
On the agentic layer they call for evaluations of whether agents faithfully represent their users — neither smuggling in their own goals nor catering to motivated reasoning. The technical bar is high, particularly when users have not stated explicit preferences. The political bar may be higher, since the same agent that refuses to challenge a user's prior beliefs is also the one users are most likely to keep using.
On the institutional layer, the authors note several states and localities are already running AI-mediated deliberation platforms, building on research that AI mediators can help citizens find common ground. They also flag that bots are already skewing public input processes, and call for identity verification for both humans and their agent proxies to be built in from the start rather than retrofitted.
The essay is not a neutral survey. The Office of Eric Schmidt has its own positions on AI policy, and the piece is published as opinion. It also leans on a single not-yet-peer-reviewed study for its most optimistic empirical claim, and it does not engage in detail with companies — OpenAI, the major model developers — whose product decisions will determine whether any of this design work happens at all. Readers should treat the three-layer frame as an agenda, not a finding.
Still, the argument lands at a useful moment. Spending on AI infrastructure keeps climbing, agent products are shipping into consumer surfaces, and the political system is heading into another cycle in which the default research tool for a meaningful share of voters will be a chatbot. "Failing to design for democratic outcomes, in a domain this consequential, means designing for something else," Sorota and Hendler write. The unresolved question is who, in the current US policy environment, has both the authority and the incentive to do that designing before the defaults harden.
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.
