Blueprint for Using AI to Strengthen Democracy
Experts warn that unregulated AI agents risk creating isolated private worlds, suggesting a three-layer approach to truth, agency, and governance.
Every few centuries, shifts in how information travels reshape how societies govern themselves. From the printing press to the telegraph and broadcast media, these changes have historically accelerated civic engagement and representative government. Now, artificial intelligence is emerging as the primary interface through which people form beliefs and participate in democratic self-governance. A new report from MIT Technology Review argues that if left unchecked, this rapid shift could further strain already fragile institutions, but with deliberate design, it could also address long-standing issues like deepening polarization.
The authors, Andrew Sorota and Josh Hendler, associated with the Office of Eric Schmidt, identify three critical areas requiring immediate design intervention. The first is the epistemic layer, which concerns how citizens come to know what is true. As people increasingly rely on AI to synthesize information and frame narratives, whoever controls these models holds significant influence over public belief. To counter this, the report suggests ramping up efforts to ensure model outputs are truthful and exploring AI-assisted fact-checking. Recent field evaluations on the platform X indicate that users with diverse political viewpoints found AI-generated notes more helpful than human-written ones, though the paper remains unpeer-reviewed.
The second area of focus is the agentic layer, which deals with user representation and agency. Personal AI agents are beginning to conduct research, draft communications, and lobby on behalf of individuals, effectively mediating the relationship between citizens and governing institutions. The report warns that while these agents must faithfully represent user views, they must not develop their own agendas or shield users from uncomfortable information. An agent that merely reinforces existing biases without allowing for questioning or adjustment fails to act in the person's best interest and risks becoming an accessory to motivated reasoning.
On the institutional layer, the blueprint calls for responsive governance and identity verification for both humans and their AI proxies. As AI agents become common participants in public deliberation processes, there is a risk that bots could skew outcomes, making it impossible to distinguish between human and machine voices. To prevent a public sphere populated by millions of personalized agents from fragmenting into isolated private worlds, the report mandates that identity verification be built into democratic input processes from the start. This measure is essential to preserve the shared deliberation required for a functioning democracy.
The authors caution that while individual agents might be well-designed, the aggregate interactions of millions could produce outcomes no single user intended. A public sphere where everyone has an agent attuned to their specific anxieties and views is not a public sphere at all; it is a collection of private worlds collectively inhospitable to the kind of shared understanding democracy requires. Avoiding this outcome depends on creating a new generation of democratic infrastructure that is both technological and institutional.
Ultimately, the report posits that today's democracy was designed for a world where power was exercised visibly and reality felt more shared. With generative AI arriving faster than many realize, failing to design for democratic outcomes means designing for something else entirely. The history of unaccountable power offers little optimism for what that alternative might be, making the current design choices all the more consequential for the future of civic life.
