Fine-tuning, not size, makes chatbots politically persuasive. The trade-off is accuracy.

A research team drawn from the University of Oxford's Internet Institute, the UK AI Security Institute, the London School of Economics, Stanford, and MIT asked 76,977 British adults to argue with a chatbot about contested political topics and measured whether their views moved. The result, "The Levers of Political Persuasion with Conversational AI," published in Science on 4 December, is the first large-scale empirical map of what makes conversational AI politically persuasive, and a precise answer to which models can be tuned into persuasion engines without frontier-scale compute.

Across three experiments spanning 19 large language models, 707 political issues, and roughly 91,000 AI-driven conversations, persuasion had little to do with how big the model is. Fine-tuning mattered more. Fine-tuned models shifted opinions by up to 51% compared with the baseline; carefully written system prompts added up to 27%. The leverage sits in how a model is trained to argue, not in raw parameter count. The preprint, posted to arXiv in July, makes that point with data rather than by argument.

The same experiments recorded a second-order finding the researchers say should change how platforms and regulators think about political AI. The more persuasive a model became, the less factually accurate its claims. The team fact-checked 466,769 statements the LLMs produced during those conversations; the trade-off held across model families and across the 707 topics, not just inside any single vendor's product. Information-dense arguments, the kind that survive a fact-check, drove roughly half of the explainable variation in cross-model persuasiveness. The lever is content that can be verified, not rhetorical style.

The implication for open source is concrete. A modestly sized open-source model, downloaded and fine-tuned, can be tuned into a politically persuasive agent, a capability once assumed to require frontier-scale compute. Lead author Kobi Hackenburg, an Oxford Internet Institute DPhil candidate and a research scientist at the UK AI Safety Institute, frames the finding as a structural shift in who can field political AI at scale. Co-author Helen Margetts, OII professor, argues that the persuasion-accuracy trade-off should be treated as a property of how these systems are optimised, not a bug to patch in the next model release.

That framing matters because the levers the study identifies are not exotic. Fine-tuning data and prompt design are within reach of any reasonably resourced operation, including state-aligned influence campaigns and small political consultancies. The Oxford team's measurement design, which compares many LLMs against the same UK participant pool on the same issues, gives downstream evaluators a yardstick. A benchmark built from the same 707-issue instrument could score new models on persuasion and accuracy from the same transcripts, which is the kind of artefact platforms and election regulators currently lack.

A single short conversation is a different exposure regime than the repeated real-world exposure a social-media timeline produces. The 707 issues skew toward contested UK political questions, and transfer to other electorates is not assumed. Aggregate effects across 76,977 participants hide topic-level variation that the paper documents but does not predict from model specs alone. "Persuasion" here means opinion movement measured once, not durable attitude change.

What the study does give is a baseline. Platforms deploying conversational AI in political feeds, regulators drafting election-integrity guidance, and voters trying to read what a chatbot is doing in front of them now have a measured correlation to argue with instead of speculation. The Oxford team is publishing the analysis pipeline alongside the paper; the next deliverable, the authors say, is a public benchmark that scores political persuasion and factual accuracy from the same transcripts.