A look at how deepfakes, with recent software advances that easily create AI-generated fake audio and video, could be misused to undermine elections
PLENTY OF PEOPLE are following the final days of the midterm election campaigns. Yale law researcher Rebecca Crootof has a special interest—a small wager. Tweets: @nxthompson and @wired Tweets: Nicholas Thompson / @nxthompson : Deepfakes are coming to politics. Which will be a bummer because we'll get manipulated by things that are fake—and also by dishonest politicians denying videos that are true. @tsimonite http://www.wired.com/... @wired : “Deepfake” describes any fake audio or video created using machine learning. This raises concerns about recent software advances—and evidence that you don't need an AI PhD to use them. http://www.wired.com/...
Context & Ripple Effects
This 2018 Wired piece is the early-warning entry in an arc the corpus has since filled in: Nicholas Thompson's wager that deepfakes would reach politics captured the moment when machine-generated audio and video became cheap enough to weaponize against elections. What followed validated both halves of his bet — not only fakes fooling voters, but the liar's dividend, where convincing fakes give cover for dismissing real footage as fake.
By late 2023 the pattern had gone global at scale: pro-government outlets in Bangladesh were promoting AI-generated disinformation built with cheap tools ahead of the January 2024 vote, and US state legislators began racing to require disclosure of AI in political ads as deepfakes warped elections overseas. The 2018 question — could this undermine elections? — has become the operating assumption.
First-order effects
- Voters and campaigns face a degraded information environment immediately: any audio or video clip in a contested race now carries a plausibility discount, since machine-learning tools can fabricate convincing candidates' voices and faces at near-zero cost.
- Nicholas Thompson's specific warning lands on both sides at once — campaigns can be hit by fabricated clips, while politicians caught in genuine scandals gain a ready-made denial ('that's a deepfake').
Second-order effects
- Platforms and newsrooms are pushed into verification arms races; DARPA's detection research reported by CNN shows the countermeasure side, but detectors' accuracy problems mean every takedown or debunk decision becomes contentious.
- Legislators respond where platforms can't keep up: state-level disclosure laws for AI in political ads create a patchwork of penalties that campaigns must navigate, and cheap generation tools push the same playbook to lower-resourced democracies like Bangladesh.
Third-order effects
- If the pattern holds, the structural shift is from 'seeing is believing' to provenance-based trust — elections increasingly depend on authentication infrastructure (watermarking, capture-time verification) rather than raw human perception, with regulation fragmenting across states and countries.
- The same dual-use tooling splits into two markets: legitimate commercial synthesis like Synthesia's corporate training videos funds capability advances that continuously lower the cost of political misuse — a commercialization loop regulators have to chase.
The trend: Synthetic media is moving from novelty to standing election-infrastructure threat, with creation costs falling faster than detection and disclosure rules can standardize.