The Phantom Public

In the AI Story

The phantom public's AI governance consequences: decisions about AI—by legislatures, regulators, boards, school districts—are made whose consequences compound over decades. The EU AI Act, American executive orders, emerging frameworks crystallize pictures of what AI is, what it threatens, what governance is appropriate. Each picture was constructed while technology evolved faster than governance could track. Each will govern behavior long after technology has moved beyond the picture's frame. This is governance by pseudo-environment: regulating technology based on pictures already obsolete. The EU AI Act classifies AI systems by risk levels designed for 2023–2024 capabilities. The 2026 capabilities do not fit the scheme. The governance structure governs a picture rather than the technology.

The expertise asymmetry compounds the phantom public problem. People who understand AI most deeply—researchers, engineers, corporate leaders—have interests diverging from public interest in specific, identifiable ways. They are not malevolent—they are structurally positioned to see AI through the builder's pseudo-environment, which emphasizes capability and possibility while systematically underweighting costs falling on people outside the building. This is not conspiracy—it is fishbowl glass. The expert whose knowledge is necessary for governance is an expert whose expertise was formed inside a pseudo-environment structurally biased toward builders' interests rather than the built-upon's interests.

Lippmann proposed two partial solutions, both relevant to AI governance though neither sufficient. First: institutional mediation—bodies translating complex realities into forms non-experts can use for decision-making. Not making the public expert (impossible) but constructing better pictures. The AI ecosystem has begun constructing such bodies: advisory councils, ethics boards, government AI offices, academic policy centers. But most suffer from the bias Lippmann identified: populated by people whose expertise formed inside the builder's pseudo-environment, funded by institutions with stakes in pictures produced, accountable to governance structures lacking technical literacy to evaluate whether pictures are accurate or merely plausible. Second: lowering democratic expectations—designing governance to function when the phantom public is dissolved (its natural state) rather than only when momentarily materialized by crisis.

Origin

The Phantom Public (1925) was written as a correction to Public Opinion's residual optimism. Where the earlier book proposed intelligence bureaus and institutional reforms, the later book conceded these proposals collided with a harder truth: the engaged, informed public required by democratic theory does not exist in any operationally meaningful sense. The book's tone was bleaker, more resigned, and more honest. Critics called it elitist; defenders called it realistic. Lippmann called it structural observation—not what ought to be but what is, given the cognitive limits of individuals and the informational complexity of modern governance.

The concept's 21st-century vindication came through empirical political science: voter ignorance studies (Ilya Somin), rational ignorance theory (Anthony Downs), and the documentation that most citizens know almost nothing about most issues they vote on. Lippmann was not arguing citizens are stupid—he was arguing the world is too complex for any individual to be adequately informed about everything governance requires. The AI moment has made this argument impossible to avoid: the technology is more complex than nuclear policy, more pervasive than trade agreements, evolving faster than any previous subject of democratic deliberation.

Key Ideas

Public as phantom. The continuously engaged, adequately informed citizenry that democratic theory requires does not exist—it materializes in crisis, reacts, dissolves. Governance happens in the intervals, conducted by small numbers who remain attentive.

Governance by pseudo-environment. Democratic institutions regulate not reality but pictures of reality constructed during periods when reality was evolving faster than pictures could track—producing regulations obsolete before implementation.

Expertise asymmetry. Those who understand AI most deeply inhabit a builder's pseudo-environment structurally biased toward capability over cost—expert intermediaries necessary for governance are the same people whose interests diverge from public interest.

Lowered expectations. If the public cannot be fully informed, design governance to protect rather than inform—accountability mechanisms constraining experts when public cannot evaluate, checks operating independently of public attention.

Permanent, not solvable. The phantom public problem is a structural feature of democratic life under complexity, not a temporary condition better institutions will eliminate. The discipline is designing for the world that exists.