How AI changes the production, spread, and detection of false information
The video appeared two days before the election. It showed the incumbent governor making a statement she had never made — endorsing her opponent, citing personal reasons, looking directly into the camera. It was a deepfake, generated in under an hour by a model anyone could access.
By the time fact-checkers confirmed it was fabricated, it had been viewed 4 million times. The retraction was viewed by 200,000. The lie had a twelve-hour head start and a twenty-to-one reach advantage over the truth.
Disinformation — deliberately false information spread to deceive — is not new. What AI changes is the economics and scale of production. Creating convincing fake video, audio, images, and text previously required significant resources and skill. Generative AI has dramatically reduced both barriers, enabling sophisticated disinformation at near-zero marginal cost.
The specific capabilities that matter: Deepfakes — realistic synthetic video and audio of real people saying or doing things they never did. Synthetic text — AI-generated articles, social media posts, and comments that are indistinguishable from human-written content at scale. Persona generation — automated creation of fake social media accounts with believable histories. Targeted personalization — tailoring disinformation to specific audiences based on psychological profiles.
The Liar's Dividend
A secondary effect of deepfake proliferation: even real videos can now be plausibly denied as fabrications. When deepfakes are common, authentic evidence becomes deniable. Politicians caught on genuine video can claim it was AI-generated. This "liar's dividend" — the ability to dismiss real evidence — may be as damaging to democratic discourse as the fake content itself.
AI detection of AI-generated content is an arms race. Detection models improve; generation models adapt to evade them. Current detection tools have significant false positive and false negative rates — flagging real content as fake and missing sophisticated fakes. Watermarking (embedding invisible signals in AI-generated content) is a promising technical approach but requires adoption by AI providers and platforms, and is easily stripped by determined actors.
The asymmetry that matters most: generating false content is faster and cheaper than verifying it. At scale, no fact-checking infrastructure can keep pace with AI-enabled disinformation production. The response to AI disinformation cannot rely primarily on detecting and correcting individual false items.
Choose a real documented case of AI-enabled disinformation — a deepfake used in a political context, a synthetic media campaign, or a coordinated inauthentic behavior operation using AI-generated content. Analyze: What was produced? How did it spread? What harm did it cause? How was it eventually identified? What could have stopped it earlier?
Start with: "I want to analyze [case] — here's what happened: [your description]"
How AI enables personalized political messaging — and what that does to democratic discourse
Voter A and Voter B both lived in the same swing district. Voter A saw ads emphasizing economic security — the incumbent's record on jobs, stability, protecting retirement accounts. Voter B saw ads emphasizing cultural grievance — threats to community values, immigration, national identity.
The ads were from the same campaign. The campaign had never spoken to either voter directly. An AI system had inferred their psychological profiles from their online behavior and selected the message most likely to move each one. The campaign had one message for the public. It had millions of messages for individuals.
Political microtargeting uses data about individuals — behavioral, demographic, psychographic — to deliver customized messages to specific people. AI has dramatically expanded both the precision of targeting (which individuals receive which message) and the scale of personalization (how specifically messages are tailored to individuals).
The Cambridge Analytica case made microtargeting's political potential public: the firm claimed to use psychographic profiles derived from Facebook data to target voters with personalized political messaging. Whether their specific claims about effectiveness were accurate is disputed — but the capability they described has become routine, without the controversy.
The Deliberation Problem
Democratic theory assumes citizens form political views through shared public discourse — encountering the same arguments, debating shared facts. AI-enabled microtargeting creates a fragmented political communication environment where different citizens receive fundamentally different information and arguments, making shared deliberation impossible. You cannot debate someone whose information environment you have never seen.
Political persuasion — attempting to change minds through arguments and appeals — is a legitimate and essential part of democratic politics. The question raised by AI-enabled microtargeting is not whether persuasion is acceptable, but whether specific persuasion techniques cross into manipulation.
The distinction that matters: persuasion operates through reasons that the person can evaluate and accept or reject. Manipulation exploits psychological vulnerabilities — anxieties, biases, emotional triggers — to produce belief changes that bypass rational evaluation. When an AI system identifies that a specific voter is psychologically susceptible to fear-based messaging and specifically delivers fear-based content to that voter, it is closer to manipulation than persuasion — even if the underlying factual claims are accurate.
Consider a specific AI microtargeting scenario: a campaign identifies that a segment of voters are highly anxious about economic security and delivers them ads emphasizing economic threat, while delivering the same voters in a different version of the campaign's platform a more optimistic economic message. Analyze: Is this persuasion or manipulation? What makes the line hard to draw? What governance would address it?
Start with: "My initial assessment of whether this is persuasion or manipulation is [your position] — because [your reasoning]"
How AI is being used in electoral administration, surveillance, and public services — and what that means for accountability
The city's AI system flagged housing benefit applications for fraud review. It had been trained on historical fraud cases. The historical cases were concentrated in certain neighborhoods — neighborhoods that had historically faced more aggressive enforcement, which had generated more documented fraud in those areas.
The system learned that being from those neighborhoods was predictive of fraud. Applications from residents there were reviewed at three times the rate of applications from other areas. The review added weeks to processing times — and for people with genuine need, those weeks mattered. The algorithm was doing exactly what it was trained to do. What it was trained to do was the problem.
Governments worldwide are deploying AI in public administration — to process benefits applications, detect fraud, allocate resources, assess risk, and make or assist in decisions affecting citizens' lives. The potential benefits are real: faster processing, more consistent application of rules, better fraud detection. So are the risks: embedded bias, opacity, removal of discretion, and accountability gaps when things go wrong.
The accountability gap is particularly acute in government AI. When a benefits algorithm incorrectly denies an application, who is responsible? The agency that deployed it? The vendor who built it? The officials who chose to use it without adequate review? Traditional accountability mechanisms — political accountability through elections, legal accountability through courts, bureaucratic accountability through appeals — all struggle when decisions are made by systems whose operation is opaque and whose errors are difficult to trace.
Historical Data, Historical Bias
Public sector AI trained on historical administrative data systematically inherits historical enforcement patterns. Communities that faced more aggressive policing have more documented crime. Benefits recipients who were subject to more stringent fraud investigations have more documented fraud. Training on these histories teaches AI systems that these communities are riskier — amplifying historical disparities rather than correcting them.
AI is being used in electoral administration — for voter roll management, ballot counting, election security monitoring — with potential efficiency benefits but also risks of errors at scale and new attack surfaces. Separately, AI enables new forms of electoral interference: generating synthetic voter suppression messaging, automating coordinated inauthentic behavior, and producing disinformation timed to electoral cycles.
The deeper concern is epistemic: democratic legitimacy depends on public trust in electoral outcomes. AI-enabled disinformation that successfully undermines that trust — even when elections are conducted fairly — damages democratic function. The perception of manipulation can be as corrosive as actual manipulation.
Choose a real documented case of AI in public administration — benefits determination, child welfare risk scoring, predictive policing, or court risk assessment. Analyze: What decision did the AI make or influence? What bias or error was found? Who was held accountable — and how? What governance mechanism would actually create accountability for this type of AI?
Start with: "I want to analyze [system/case] — here's how it worked and what went wrong: [your description]"
What governance approaches exist, what is missing, and what democracy actually requires from AI
The parliamentary committee had been studying AI and democracy for eighteen months. They had heard from technologists, civil society, platform companies, and academics. Their report ran to 340 pages.
Its core finding: the tools to govern AI's effects on democratic life existed — transparency requirements, algorithmic audits, political advertising disclosure, public sector AI standards. What didn't exist was the political will to implement them against the interests of the platforms, campaigns, and government agencies that benefited from the current opacity. The governance problem wasn't technical. It was political.
Several governance approaches address AI's effects on democratic life, with varying coverage and effectiveness. Political advertising disclosure: platforms in many jurisdictions require disclosure of political ads and their targeting. The EU's Political Advertising Regulation extends this to transparency about targeting criteria. Election integrity laws: some jurisdictions have enacted laws specifically addressing AI-generated content in political advertising — requiring disclosure of synthetic media. Platform content moderation: major platforms have policies against coordinated inauthentic behavior and some AI-generated political content, though enforcement is inconsistent. Public sector AI standards: the EU AI Act classifies many government AI uses as high-risk, requiring documentation, testing, and human oversight.
Effective governance of AI in democratic life requires addressing the specific things AI threatens in democracy: Shared reality — citizens need access to enough common facts to deliberate together. Authentic political communication — the ability to know whether political communication is from a human, is genuine, and is being received by others or targeted only to you. Accountable government — citizens need to know what decisions government is making about them and why, with meaningful recourse. Fair elections — the ability to trust that electoral outcomes reflect genuine citizen preferences rather than algorithmically amplified manipulation.
The Democratic Stakes
Democratic governance is not just one application domain among others for AI. It is the process by which societies make collective decisions about AI itself — including decisions about how AI should be governed. AI that undermines democratic processes undermines the mechanism by which its own governance is determined. This makes the interaction between AI and democracy uniquely high-stakes: getting it wrong doesn't just harm individuals, it compromises the capacity for collective self-correction.
Design a governance framework to protect one specific democratic requirement from AI threats: shared reality, authentic political communication, accountable government, or fair elections. Specify: What is the specific threat? What governance mechanism addresses it? Who implements and enforces it? What are the strongest objections to your approach, and how do you respond?
Start with: "I want to protect [democratic requirement] from AI threats. The specific threat I'm addressing is [description], and my governance proposal is [your framework]"
15 questions. Complete all to finish the module.