Ahead of the 2024 election cycles across dozens of democracies, the dominant expert prediction was that generative AI would meaningfully degrade the information environment. Two years and many elections later, the picture is more complicated than either the alarmists or the dismissers claimed.
What did not happen
The apocalyptic scenario — a single viral deepfake that swings a national election — did not occur in any well-documented case. Where AI-generated content did circulate, it was generally caught quickly, correctly identified as fake, and did not appear to move measurable public opinion. The infrastructure of fact-checking, platform detection, and rapid attribution, while imperfect, held up reasonably well against high-profile fakes.
What did happen
The subtler effects are harder to measure but appear to be real. AI-generated content is now used routinely for micro-targeted political advertising, giving campaigns the ability to produce personalized appeals at previously impossible scale. Non-obvious AI-assisted content — reworded talking points, AI-drafted opinion pieces, AI-generated commenter personas — has proliferated in ways that never generate a single dramatic incident but collectively shift the information environment.
Perhaps the most consequential shift is what researchers have started calling the 'liar's dividend': the increasing ability of any political actor to dismiss authentic damaging content as an AI fabrication. This is the mirror image of the deepfake panic — not fake content being believed, but real content being disbelieved.
- Micro-targeted persuasion at unprecedented scale is the dominant real-world use.
- The 'liar's dividend' — dismissing real content as fake — may be the most consequential effect.
- Single-incident deepfake attacks have been rare and largely unsuccessful.
“The threat to elections was never one perfect fake. It was a million imperfect ones — and, more corrosively, the plausibility of dismissing anything real as one of them.”
What the interventions taught us
Rapid-response fact-checking, platform provenance disclosure, and pre-bunking campaigns all showed measurable positive effects in the elections where they were deployed. The interventions that failed were ones that assumed audiences would seek out corrections; the ones that worked pushed corrections into the same channels where the original content had circulated.
For the next cycle
The policy conversation is shifting from 'ban deepfakes' toward 'require disclosure, invest in provenance, and prepare institutions for the liar's dividend.' None of these are complete solutions. All of them are more tractable than the counterfactual of restricting the underlying technology, which the past two years suggest is neither politically feasible nor especially effective.
Key Topics
Extended Knowledge
- The single-fake apocalyptic scenario has not materialized.
- Cumulative micro-targeted AI content is the dominant real-world electoral use.
- The 'liar's dividend' may be the most durable effect for democratic institutions.
Frequently Asked
Partly. The specific scenario of a single decisive fake did not occur. Broader concerns about the information environment have been at least partly validated.
Rapid attribution, provenance disclosure, and pre-bunking campaigns showed measurable positive effects.
The 'liar's dividend' — the erosion of trust in authentic evidence — may be harder to address than the fakes themselves.
