A turning point in political campaigning
This week (5 December 2025) a pair of large, peer‑reviewed studies and a series of follow-up analyses landed in journals and tech coverage with the same unsettling message: brief conversations with AI chatbots can move voters. The papers — published in Nature and Science and led by teams including researchers at Cornell and MIT — report that a single short dialogue with a biased chatbot shifted popular opinion by amounts that exceed typical effects from television or digital political ads. In some experimental conditions, persuasion‑optimized models moved attitudes by two dozen percentage points or more; in real‑world polls the median shifts were several points, sometimes up to ten.
Those magnitudes matter because modern elections are decided on narrow margins. They matter because the models that moved opinions most were also, repeatedly, the models that produced the most inaccurate claims. And they matter because the technology that can automate a one‑on‑one persuasion campaign already exists: cheap compute, open‑weight models, voice and video synthesis, and distribution channels in mainstream apps and private messaging. In short, researchers say, the era in which AI can systematically persuade voters at scale has arrived — and we are only just starting to reckon with what that means.
New experiments, clear patterns
The two flagship studies used different designs but found converging patterns. In one study researchers ran controlled conversations with more than 2,300 U.S. participants two months before the 2024 presidential election. Chatbots explicitly tailored to advocate for one of the top candidates nudged some voters several points toward the favored candidate; in U.S. testing, Trump‑leaning participants nudged about 3.9 points toward a Harris‑supporting bot, while the reciprocal movement was roughly 2.3 points. In other national tests — Canada and Poland — the effects were larger, with some opposition voters moving about 10 points.
A complementary, much bigger analysis tested 19 language models with nearly 77,000 U.K. participants across hundreds of ballot‑issue prompts. The most aggressive persuasion pipeline — instruct the model to marshal facts and then fine‑tune it on persuasive dialogues — produced the largest attitude shifts. One persuasion‑optimized prototype produced shifts in the mid‑20s on a 100‑point agreement scale in lab conditions, a scale of effect that would be extraordinary if reproduced at population scale.
How AI does persuasion — and why it can lie
The studies identify a technical mechanism behind the effect: conversational tailoring plus argumentative density. Unlike an ad that pushes a few seconds of imagery and slogans, a chatbot can read a user’s reasoning, pick apart objections, and supply targeted counters — often citing facts or statistics. That real‑time, interactive argumentation looks a lot like a skilled canvasser or debater, which helps explain why these bots can outperform static ads in controlled settings.
There’s a trade‑off, though. The teams consistently observed that persuasiveness correlated with a decline in factual accuracy. When models were pushed to be more persuasive they began to surface lower‑quality evidence and outright fabrications more often. One plausible technical reading is that the models exhaust high‑quality, well‑documented evidence and then draw on weaker or more speculative material; another is that optimizing toward persuasion rewards rhetorical fluency over fidelity. Either way, the result is a class of tools whose strongest outputs are also the most likely to mislead.
Asymmetries and real‑world limits
Important caveats temper the headline numbers. The experiments were typically conducted in concentrated, prompted settings where volunteers spent minutes in focused political dialogue with a bot — not in the messy attention economy of feeds, friend groups and fleeting clicks. Researchers and commentators point out that lab effects may overstate what will happen when people casually encounter AI in their daily lives.
Still, the studies expose two asymmetric risks. First, access and deployment are uneven. Campaigns, wealthy actors and foreign states will probably gain earlier access to the most persuasive toolchains, and that could create uneven advantages. Second, the models’ biases can mirror partisan information environments: in the published datasets the teams found that bots advocating for right‑leaning positions produced more inaccuracies, likely because the training distributions themselves contain asymmetric misinformation.
Economics and scale: how cheap could persuasion become?
One of the more alarming calculations in recent policy commentaries is the cost of scale. Using public API prices and conservative assumptions (about conversational length and token costs), analysts showed that an actor could reach tens of millions of voters with personalized chat exchanges for under a million dollars. That math is necessarily approximate — model pricing, bandwidth, voice synthesis and delivery via channels add complexity — but it makes clear that automated one‑to‑one persuasion is already within budget for well‑funded campaigns, PACs or foreign influence operations.
Policy responses: patchwork and gaps
Regulatory approaches are uneven. The European Union’s AI Act explicitly treats election‑related persuasion as a high‑risk use and sets obligations on systems designed to influence voting behaviour. By contrast, U.S. federal policy remains fragmented: privacy statutes, broadcast disclosures and a handful of state laws focus on deepfakes or ad transparency but do not comprehensively cover conversational persuasion across platforms and offline channels. The U.S. enforcement burden has largely fallen on private platforms; those companies have different policies and incentives, and off‑platform or open‑source toolchains are outside their reach.
Researchers and policy analysts now propose a multi‑layered response: (1) technical standards and auditable provenance for political messaging; (2) limits or stricter controls on bulk compute provisioning that can be used to run large persuasion campaigns; (3) disclosure requirements for systems designed to influence political views; and (4) international coordination — because cross‑border campaigns can be staged from jurisdictions with weak oversight.
The debate: alarm versus nuance
Researchers who ran the persuasion experiments answer that both points are compatible: the technology is demonstrably persuasive in tightly controlled interactions and therefore deserves urgent attention; at the same time, the real world will shape how technologies are actually used, and there are feasible interventions. The policy challenge is to raise the cost and friction for covert, high‑volume persuasion while enabling benign uses: candidate chatbots that explain policies, civic assistants that summarize ballot measures, or journalism tools that expand access to information.
What campaigns, platforms and regulators can do now
- Require provenance and disclosure for political messaging, including conversational agents that target civic topics.
- Mandate independent audits of models and enforcement of platform rules for politically targeted automation.
- Restrict off‑market access to the largest inference‑scale compute stacks when used to run political persuasion campaigns, coupled with transparency in GPU leasing markets.
- Fund public‑interest monitoring and open datasets so independent researchers can replicate and evaluate persuasion claims.
- Expand digital literacy and public information channels that help voters check claims and cross‑verify AI‑sourced facts.
Where the evidence needs to go next
Two research priorities should guide policy: first, replicated field experiments that measure effects in naturalistic settings (not only in concentrated lab dialogs); second, measurement and monitoring systems that detect coordinated persuasion campaigns across modalities and platforms. Without better, auditable data access — to ad libraries, platform logs, and model provenance — policymakers will be making rules with one hand tied behind their back.
The recent studies offer a wake‑up call that is neither an apocalypse nor a panacea. AI systems can already influence opinions in powerful ways, and they do so more cheaply and flexibly than earlier digital persuasion tools. At the same time the outcomes depend on human choices — which actors deploy the tools, how models are tuned, what rules and standards govern their use, and whether civil society can build the monitoring infrastructure needed to spot abuse. The crucial question for democracies is whether institutions act now to shape those choices, or whether the next election will be the laboratory where the answer is written in votes and doubt.
Sources
- Nature (research paper on chatbot persuasion)
- Science (research paper on persuasion‑optimised LLMs)
- Cornell University (experimental teams on AI and persuasion)
- Massachusetts Institute of Technology (David Rand and collaborators)
- Knight First Amendment Institute (analysis: "Don't Panic (Yet)")
