Democratic Erosion

In February 2020, the Pew Research Center asked a large group of technologists, scholars, and policy experts a blunt question: by 2030, would digital technology mostly strengthen or mostly weaken democracy? Nearly half — 49 percent — said it would weaken it. Only a third expected it to help (Pew Research Center, Many Tech Experts Say Digital Disruption Will Hurt Democracy, 2020). What is striking, reading their written responses now, is how few of them worried about a coup or a cyberattack on voting machines. What they feared was subtler and slower: the steady corrosion of the shared reality that self-government depends on. They worried that the very tools built to connect citizens would end up drowning them in noise, sorting them into incompatible worlds, and handing outsized voice to whoever could generate the most content the fastest.

That worry was registered before large language models could write a passable op-ed in four seconds, before a convincing video of a head of state could be conjured from a text prompt, before "flooding the zone" stopped being a metaphor and became a line item in a budget. The experts were describing a hazard they could see forming on the horizon. Much of it has since arrived. This chapter is about what that arrival is doing to democratic health — what the cross-national evidence can and cannot yet tell us, and which of the mechanisms people most fear are documented rather than merely plausible.

What the Cross-National Evidence Shows — and Doesn't

Start with the hardest question, because everything else in this chapter should be read in its light: is democratic erosion from AI already measurable, or is it still mostly a forecast?

The most-cited empirical anchor is a cross-national analysis using a two-way fixed effects model across 72 countries, which reports that more intensive AI development is associated with declining democracy scores — the countries pushing hardest on AI capability tend, on average, to show weaker democratic institutions over the same period. A two-way fixed effects design is worth pausing on, because it is better than a simple correlation. It compares each country to itself over time (country fixed effects) while also stripping out shocks that hit every country in a given year (year fixed effects), so the finding is not simply "rich AI countries happen to be less democratic" or "everyone's democracy scores fell in 2024." It is closer to: within a country, as its AI development intensifies, its democratic indicators tend to move the wrong way.

That is a real signal, and it deserves to be taken seriously. But it is not proof that AI causes democratic decline, and honesty about the three registers of knowledge — what we know, what is expected, what is feared — requires saying why. The core problem is confounding. The past decade's democratic backsliding has many well-documented drivers that have nothing to do with AI: economic concentration, affective polarization, the hollowing out of local journalism, the rise of personalist parties. Several of these are also correlated with AI investment, because the same wealthy, digitally saturated, politically strained societies tend to score high on both. Fixed effects remove stable national traits and common global years, but they do not remove a confounder that is itself changing within a country over time — and polarization and media collapse are exactly that kind of moving target. A country intensifying its AI sector during the same years its local newspapers die and its politics harden will show the correlation the study reports, whether or not AI is the active ingredient.

So how much of the 72-country relationship would survive stronger identification? We do not honestly know, and no observational design can fully settle it. Stronger causal identification would require something closer to a natural experiment — a plausibly random reason why some places got more exposure to AI-generated political content than otherwise identical places at the same moment — or fine-grained within-country tracking of specific mechanisms, not aggregate democracy scores. That work is largely still ahead of us. The defensible reading is the middle one: the cross-national pattern is consistent with AI contributing to democratic weakening, it is not yet a clean causal estimate, and the expert survey should be read as informed alarm from people who understand the technology, not as a measurement of harm already booked. The rest of this chapter maps the mechanisms behind that alarm, taking care throughout to mark which rest on direct evidence and which rest on inference from capabilities and incentives.

How the Representative Relationship Breaks Down

Picture a senator's office a few years from now. The inbox holds not thousands of constituent messages but hundreds of thousands, each individually worded, each carrying a plausible personal story and a local reference, each urging a position on a bill. The staff cannot read them, so they run an AI over the pile to summarize "constituent sentiment" — an AI reading the output of other AIs, with the actual voters buried somewhere underneath. This scene is a projection, not a report. But every element of the machinery it describes exists today, and it dramatizes the first and most direct of four ways AI corrodes representation.

That first mechanism is synthetic advocacy overwhelming authentic communication. Generating a personalized, persuasive letter to a legislator once required a human being; it now costs effectively nothing at scale. This is not purely hypothetical: as far back as 2019, a Harvard student demonstrated that AI-generated comments could be submitted to a federal rulemaking process and were, in a small study, indistinguishable to reviewers from human ones — and in 2017 the FCC's net-neutrality docket was already swamped by millions of fake comments, most of them templated rather than AI-written. Lower the cost of a unique, human-sounding message to zero and the flood becomes both larger and far harder to filter. When volume swamps authentic voice, the signal a representative needs — what do the people I answer to actually want? — is buried in noise the office cannot afford to sift.

The second mechanism, algorithmic amplification distorting perceived public opinion, works on the officials' other window into their constituents: social media. Engagement-optimizing systems reliably surface the emotionally extreme over the moderate, because outrage travels further than nuance. A legislator watching the feed to gauge the mood therefore sees a caricature — angrier, more polarized, more absolute than the actual electorate. The distortion is not that officials are lied to about facts; it is that the distribution of opinion they perceive is warped toward the fringes by the selection algorithm itself.

The third is the most philosophically unsettling: AI curation shaping preference formation. The first two mechanisms assume real preferences exist and are being misrepresented. But recommendation and curation systems do not merely transmit opinion; they help form it, deciding which arguments, framings, and facts ever reach a given citizen. If the preferences voters express have themselves been substantially shaped by AI-curated information diets, then even a representative who perceives them perfectly faces a hollow victory. The question "whose preferences am I representing?" acquires a prior question — "through what process were those preferences produced?" — that representative democracy was never designed to answer.

The fourth mechanism is a resource asymmetry that ties the others together, and it deserves its own section because it recurs across the whole chapter. Briefly: because legislators read communication volume as a proxy for how intensely people care, and because AI lets well-resourced organizations manufacture volume, AI systematically amplifies the loud and the funded at the expense of the ordinary citizen who can still only write one letter in one voice.

graph TD
  A[Cheap AI content generation] --> B[Synthetic advocacy floods inboxes]
  A --> C[Engagement algorithms amplify extremes]
  A --> D[Curation shapes what voters believe]
  A --> E[Organized groups manufacture volume]
  B --> F[Representative loses reliable signal of constituent will]
  C --> F
  D --> F
  E --> F
  F --> G[Representation becomes formally intact, functionally hollow]

The Structural Resource Asymmetry

Democracy's foundational promise is political equality: one person, one vote, and — at least in principle — a roughly comparable ability to be heard between elections. AI widens the gap between that formal equality and its effective version, and it does so structurally rather than incidentally.

The logic is simple and hard to escape. Political influence outside the ballot box runs substantially on communication — the ability to reach officials, to reach other citizens, to commission research, to sustain a message across many channels over time. AI dramatically lowers the cost of all of these, but it lowers them most usefully for actors who already have the money, data, and technical staff to deploy the tools at scale. A trade association can now run the persuasive-communication output of what used to be a large firm; an ordinary citizen still has one voice. The tools are nominally available to everyone, but the returns to using them scale with resources you already command.

This is why the asymmetry is not just "the rich have always had more clout." Traditional advantages in political voice were bounded by the cost of human labor — even a wealthy interest group had to hire and manage people to write, call, and organize. AI removes that ceiling on the supply side for whoever can afford the systems, while leaving the unfunded citizen roughly where they were. The practical implication is a widening wedge between formal and effective political equality: the vote stays equal, but the space between votes — where agendas are set and officials are lobbied — tilts further toward organized capacity. That wedge is the reason several observers argue AI-enabled political advocacy belongs in the same regulatory conversation as campaign finance, a point the chapter returns to in its discussion of disclosure.

Electoral Integrity in the AI Era

Elections are where democratic legitimacy is manufactured, and AI threatens them in ways that are genuinely different from the fraud that election law was built to catch. The distinction matters because our defenses — ballot audits, chain-of-custody rules, criminal statutes — are aimed at tampering with the count. AI's threats mostly leave the count alone and attack the environment around it.

The table below sets the new threats against their traditional analogues.

Two features run across the row. First, attribution collapses. Traditional suppression required human organization that left traces — printed flyers, phone records, identifiable operatives. AI-generated suppression can run through organic-appearing content that is difficult to trace to any coordinator, which is precisely what makes it hard to prosecute or even prove. Second, and most consequential for stability, is post-election legitimacy contestation. Research on democratic resilience consistently finds that the willingness of losers to accept defeat is among the most fragile of all democratic norms — and it is a norm sustained by shared belief, not by law. AI does not need to win a court case to damage it. It needs only to make fraud claims feel substantiated to a large enough slice of the public, through material realistic enough to survive a glance if not an audit. The danger is not a stolen election; it is an election that half the country refuses to believe it lost.

The Epistemic Commons Problem

Underneath every mechanism so far lies a single dependency: democratic deliberation requires a shared factual floor. Citizens can disagree fiercely about values and still govern themselves, provided they are arguing about the same reality. AI-generated content pollution threatens that floor, and the mechanism by which it does so is more interesting than "fake news makes people believe false things."

The deeper problem is what uncertainty does to judgment. When any given image, quote, or recording might be synthetic, the rational response would be to suspend belief and verify. But verification is costly — it takes time, skill, and access most people lack — and humans under uncertainty do not suspend judgment. They fall back. Faced with a claim they cannot check, people default to what they already believe and to sources their in-group already trusts. This is well established in the psychology of motivated reasoning, and it produces a cruel inversion: the more polluted the information environment becomes, the more people retreat into prior conviction and tribal authority, because those are the only cheap heuristics left when everything might be fake.

The result is a double harm. Genuinely false content can be believed, yes — but the subtler damage is that genuine content loses its power to compel. This is sometimes called the "liar's dividend": once everyone knows convincing fakes are possible, a real recording of real wrongdoing can be waved away as a deepfake, and the accusation sticks or slides off along pre-existing lines of trust rather than on the evidence. Real evidence gets dismissed; fabricated evidence gets believed; and in both cases the deciding factor is prior allegiance rather than fact. A public that cannot converge on what happened cannot hold anyone accountable for it, because accountability is a collective act that requires shared evaluative ground. The epistemic commons is not being filled with lies so much as being made unusable — a well that still holds water no one is sure they can drink.

Polarization and the Death of Loyal Opposition

Polarization predates AI by decades, but engagement-maximized content has poured accelerant on it, and the specific casualty is a norm most people never think about until it fails: loyal opposition.

The mechanism is by now well documented in its component parts. Recommendation algorithms optimize for engagement; the content that most reliably generates engagement is content that triggers strong emotion; and the most dependable emotional trigger is portraying the out-group as threatening or contemptible while affirming the in-group. Run that optimization across a population for years and you do not merely get people who disagree. You get people who have been trained, click by click, to experience their political opponents as enemies of their fundamental interests. The point worth emphasizing is whose content this is: it is not primarily propaganda from foreign adversaries but the ordinary output of domestic platforms doing exactly what they were built to do.

Here is why that specifically endangers democratic norms. Democracy asks something psychologically difficult of everyone: to lose an election and treat it as a setback to be reversed next time, not as a catastrophe to be resisted by any means. That bargain only holds if the other side's victory is survivable — if power changing hands is a normal event rather than an existential threat. Engagement-driven content erodes exactly that perception. When your information diet has spent years telling you the other side is not merely wrong but dangerous, accepting their win starts to feel like accepting your own destruction, and the norms against political violence, against subverting the process, against refusing to concede all weaken together, because each of them was only ever held in place by the belief that defeat is bearable. Information fragmentation compounds this: earlier eras of severe polarization were cushioned by shared newspapers and broadcasts that, whatever their slant, reported a common set of events. AI-personalized feeds dissolve that common ground, so that partisans now disagree not about what to do about reality but about what reality is — and disagreement over facts is far harder to resolve than disagreement over values.

Economic Disruption as Democratic Risk Factor

There is a second, more indirect pathway from AI to democratic decline, and it runs through the labor market. Political scientists have long observed that economic precarity is fertile ground for authoritarian appeal: people who feel their material security slipping become more willing to trade political rights for promises of order and decisive action, and less able to spare the time and attention that civic participation demands. The historical association between economic distress and democratic backsliding — from interwar Europe to more recent episodes — is one of the more robust findings in the study of regime stability.

AI's projected labor-market impact is large enough to matter here. The Tony Blair Institute, among others, has modeled displacement on the order of tens of millions of jobs over the coming years even as AI creates others, with the net effect depending heavily on transition support and the pace of adjustment (Tony Blair Institute for Global Change). The distributional detail from earlier chapters of this book is what makes this a democratic problem and not merely an economic one: the gains from AI concentrate sharply among those who own and control the systems, while the disruption falls on workers with far less capacity to absorb it. A society that lets that asymmetry run unmanaged does not just produce inequality; it produces a population that is anxious, insecure, and — the historical record warns — more receptive to leaders who promise to take control by breaking constraints.

The two pathways then reinforce each other. Economic stress does not only make authoritarian promises more attractive; it also thins out civic engagement, because people preoccupied with keeping their footing have little left over for monitoring officials or participating in the slow work of self-government. Weaker citizen oversight means officials face less scrutiny exactly when the informational environment is already making scrutiny harder. Economic disruption, in other words, is not a separate story from the information story. It supplies the receptive audience that polluted information then acts on.

Institutional Trust and Its Erosion

Democratic institutions run on a currency thinner than law: legitimacy. A court's rulings, an election commission's certifications, a regulator's rules carry weight only because enough people accept that they should. AI attacks that currency through three distinct channels, and it is worth separating them because they call for different responses.

The first is demonstrated incompetence. When institutions visibly fail to cope with AI-driven disruption — cannot stem synthetic misinformation, cannot produce coherent AI rules before the technology has moved on, cannot tell the public what is real — citizens generalize. An institution that looks helpless before a defining challenge of its era loses standing as a competent steward of anything, regardless of how it performs elsewhere. The perception of being overwhelmed is itself delegitimizing.

The second is opacity. Governments increasingly lean on AI for consequential decisions — benefits eligibility, risk scoring in policing, fraud flags, regulatory triage — and these systems are frequently opaque not only to the citizens they judge but to the officials nominally accountable for them. When a person harmed by such a decision asks why, and the answer is a reference to a model no one will or can explain, distrust is not paranoia; it is the reasonable inference. Opacity converts every adverse decision into a potential grievance against the institution itself.

The third is the perception that AI serves power, not citizens. Whether or not it is true in a given case, the belief that these systems are deployed to watch, sort, and manage the public on behalf of the institution — rather than to serve the public — is corrosive on its own. It reframes the state's technological capacity as something done to citizens rather than for them, and legitimacy does not survive that reframing intact. The same dynamic makes corruption feel more plausible and harder to expose, since affairs conducted through AI-mediated, audit-resistant channels are genuinely harder to document — and the mere perception that wrongdoing is now easier to hide is enough to erode trust, even if its actual incidence never rises.

The International Competitive Dimension

Even a government that clearly saw all of this and wanted to act faces a trap that no single country can escape alone. Democratic AI governance is caught in a collective action problem with the structure of an arms race.

The logic runs like this. Any democracy that imposes real restraint on AI development — for the sake of electoral integrity, information quality, or safety — worries that it will simply cede capability and economic advantage to rivals who do not. So each government, reasoning individually, leans toward permissiveness to avoid falling behind, even though nearly all of them would be better off under coordinated restraint. Individually rational choices sum to a collectively irrational outcome: a race toward looser rules that no participant would choose if it could bind the others. This is why international coordination on electoral AI, synthetic-media labeling, and algorithmic transparency has been so thin. States perceive strategic value in keeping their hands free, and the slow rhythm of multilateral negotiation cannot keep pace with a technology that reinvents itself in months.

Overlay on this a perception problem that raises the stakes. Authoritarian states experience AI differently: where democracies must absorb its disruptions to open discourse, autocracies can wield the same tools — surveillance, targeted suppression, industrial-scale propaganda — as instruments of control, without the constraints that make democratic governance harder. If that produces an impression, at home and abroad, that authoritarian AI governance is simply more effective — more stable, more decisive, better at managing the technology — then the democratic model faces not just functional strain but a crisis of confidence. The most dangerous version of AI's threat to democracy globally may be the argument, made plausible by democracies' own visible struggles, that self-government is no longer up to the task.

Can Democracies Adapt? What We Know and Don't

The honest answer to whether democracy can adapt is that we have precedents, and they are encouraging and inadequate in roughly equal measure.

History does offer cases of democratic systems absorbing disruptive new media. The printing press destabilized political and religious authority across Europe and helped fuel more than a century of upheaval before societies built the institutions — public education, professional journalism, libel law, editorial norms — that turned cheap print into a democratic asset rather than a solvent. Radio and television arrived with genuine fears about demagoguery and propaganda, some of them borne out, and democracies eventually answered with broadcast regulation, public-service mandates, and equal-time rules. The lesson from these episodes is real but double-edged: democracies can domesticate disruptive information technologies — and they have done so on the scale of decades, through institutions built after the harm was already visible. Whether that pace is survivable against a technology that advances on the scale of months is exactly the open question. The precedents tell us adaptation is possible; they do not tell us it is fast enough, and the honest position is that we do not yet know whether it will be.

What is being tried falls into a few families, and their common weakness is a structural lag between how fast rules can be made and how fast capability moves.

The most promising of these on the demand side are the deliberative mechanisms — citizens' assemblies and deliberative polling — precisely because they do not try to out-run AI content in the open information environment. They carve out a protected space: a demographically representative group, given time, good information, and expert testimony, deliberating away from the engagement machine. Ireland's assemblies on constitutional questions and France's climate convention showed that ordinary citizens, so insulated, reliably produce more considered, less polarized conclusions than the ambient discourse would predict. The unresolved problem is scale. A single assembly of a hundred or two is a proof of concept; to affect national democratic function it would have to become a routine, respected part of the system — convened often, on consequential questions, with outputs the legislature must engage rather than ignore. Nothing yet operates at that scale, and building it is a matter of political will, not technical possibility.

Who Is Responsible, and What Would Make Them Act

If the harms are real, responsibility cannot be left as a rhetorical question. Four actors hold the relevant levers, and each responds only to a different kind of pressure.

AI companies and platform operators hold the most direct control over both the generation of synthetic political content and its amplification, but their business incentives run against restraint — engagement is the product, and volume is the revenue. Voluntary pledges have a poor track record here for exactly that reason. What would actually move them is liability and disclosure: legal exposure for foreseeable political harms, and mandatory transparency about what their systems amplify and how, enforced by an outside body with audit powers. National governments hold regulatory and electoral-law authority but face the international-competition trap and the pace problem; their behavior changes when domestic legitimacy is visibly at stake and when adaptive, sunsetting rules replace static ones that ossify on passage. International bodies are the only actors that could relieve the collective action trap, but they act only when major powers find coordination in their own interest — which argues for building narrow, verifiable agreements (on electoral AI, on provenance standards) rather than waiting for comprehensive treaties. The pattern across all four is that responsibility follows accountability mechanisms: each actor acts on democratic harm only when a specific, enforceable consequence makes ignoring it more costly than addressing it.

This is also where the sharpest normative question of the chapter lands: should AI-generated political advocacy face stricter disclosure than the human kind? The case for yes rests not on the content of the speech but on its scale. Campaign finance law already accepts a principle that unsettles free-speech absolutists — that beyond some threshold, sheer quantity of amplified political speech is itself a democratic harm, because it lets money buy disproportionate voice. AI-enabled advocacy is that concern in a new register: it lets one actor manufacture the apparent voices of millions, collapsing the link between the number of people who hold a view and the volume with which it is expressed. If the campaign-finance intuition is sound — that political equality can be injured by scale alone — then requiring AI-generated advocacy to be labeled as such, and treating synthetic-persona operations as the functional equivalent of undisclosed spending, is the consistent extension of a principle democracies already accept. The counterargument, that disclosure rules chill legitimate speech and are hard to enforce across borders, is real; but it is an argument about implementation, not about whether the scale differential constitutes a harm worth regulating.

Summary

1. The evidence is suggestive, not settled. A two-way fixed effects study across 72 countries finds AI development associated with declining democracy scores, and roughly half of technology experts surveyed by Pew in 2020 expected technology to weaken democracy by 2030. But the cross-national correlation is vulnerable to confounders — economic concentration, polarization, media collapse — that move within countries over time, and no observational design yet isolates AI as the cause. Read the 72-country finding as a real signal deserving concern, and the expert survey as informed alarm, not as harm already measured.

2. Representation breaks down through four mechanisms: synthetic advocacy that buries authentic constituent voice under limitless machine-generated messages; engagement algorithms that distort perceived opinion toward the extremes; curation that shapes the preferences voters then express; and a resource asymmetry that amplifies the funded and organized over the ordinary citizen.

3. Electoral threats differ from classic fraud because they leave the count alone and attack the surrounding environment — AI astroturfing, micro-targeted informational suppression that evades attribution, and fabricated "evidence" that fractures acceptance of results without needing to win in court. The most fragile norm they attack is the loser's willingness to concede.

4. The epistemic commons erodes not mainly by spreading lies but by making verification impossible. Under pervasive uncertainty people fall back on prior belief and in-group trust rather than suspending judgment, so genuine evidence loses force ("the liar's dividend") and accountability, a collective act, becomes impossible.

5. Polarization endangers democracy specifically by killing loyal opposition: years of engagement-maximized content that frames opponents as existential threats make electoral defeat feel like destruction, weakening at once the norms against violence, subversion, and refusing to concede.

6. Economic disruption supplies the receptive audience. AI's projected displacement of tens of millions of jobs, with gains concentrated among owners, recreates the precarity that historically feeds authoritarian appeal and thins the civic engagement that oversight depends on.

7. Trust erodes through three channels — visible institutional incompetence before AI-driven challenges, opacity in AI-assisted government decisions, and the perception that AI serves institutional power rather than citizens.

8. International competition traps democracies in an arms-race logic: each fears that unilateral restraint cedes advantage, so all drift toward permissiveness, while the impression that authoritarian AI governance is "more effective" threatens confidence in the democratic model itself.

9. Adaptation is possible but slow. Democracies eventually domesticated print, radio, and television — over decades, through institutions built after the harm appeared. Today's most promising responses, deliberative citizens' assemblies and disclosure regimes modeled on campaign-finance law, remain far below the scale needed, and every regulatory response lags a technology that moves in months. Whether adaptation can be fast enough is the genuine open question of this chapter.

Sources

• Many Tech Experts Say Digital Disruption Will Hurt Democracy — Pew Research Center (2020)
• How AI Threatens Democracy — Journal of Democracy
• Can Democracy Survive the Disruptive Power of AI? — Carnegie Endowment
• AI and Democracy: Mapping the Intersections — Carnegie Endowment
• How will advanced AI systems impact democracy? — arXiv
• The Impact of AI on the Labour Market — Tony Blair Institute for Global Change
• AI and democracy: pathway to progress or decline? — Taylor & Francis
• AI: Challenges for Democracy — Taylor & Francis

Two notes flagged for the human author, outside the chapter body:

• Web access was not authorized this run, so I could not re-verify the two central empirical anchors. I rewrote the chapter to treat the 72-country study as a reported-but-uncaveated correlation and built the required epistemic section (question 14) around interrogating it — but if you have the actual citation for that study, it should be named in-text. The Pew figure (49%, 2020) I'm confident is real and correctly characterized.
• The Sources list carries forward the previous version's URLs plus the Pew report; none were fetched this session, so treat them as prior citations to confirm, not as freshly validated links.

Last updated: 2026-09-11