Language models are persuasive - and that's a good thing
Two new studies provide insights into exactly how LLMs persuade, and what that means.
Two new studies were published last week, in two of the most prestigious scientific journals, and sharing an overlapping authorship list. Persuading voters using human–artificial intelligence dialogues (Lin et al, 2025) was published in Nature, and tested the effect of AI dialogues on voters’ preferences in the context of elections in the US, Canada and Poland. The levers of political persuasion with conversational artificial intelligence (Hackenburg et al, 2025) was published in Science, and tested the effect of AI dialogues, when driven by different models and different prompting strategies, on attitudes of UK adults across British political issues.
Both reports show that language models are persuasive. This means that they shift participants’ position on the issues they set out to persuade on, measured using the shift in before-after difference of how participants self-report their beliefs on a 100 point scale1. Not only do the models persuade, there’s some evidence they can be even more persuasive than a baseline of a static persuasive message.
There are sure to be some whom this worries. It raises the spectre of a lumpen population dragged around by artificially generated propaganda. It doesn’t worry me. These studies are good news. Not only should we expect language models to be persuasive, the details in the papers about exactly how they are persuasive are good news for human rationality, and good news for the future.
Both studies show that the crucial ingredient to persuasion is evidence and argument.
In their US election persuasion experiment, Lin et al used an automated analysis to detect which persuasion strategies were deployed by their models. Note that this was 2024 and the Trump-Harris presidential race. If we believe the common story, this was an acrimonious election campaign with supporters of both candidates firmly entrenched and hard to reach. Lin et al found their participants listened to the models, and shifted their attitudes. A small amount, but they shifted. Here’s a detail from their Figure 3, which shows the most frequently used strategies by the model:
Top persuasion strategies: politeness and civility, evidence and facts!
The Lin et al study prompted the models to be persuasive, giving detailed instructions in the prompt on how to make a respectful, argument-based, case (so perhaps less surprising that the model displayed these features). The other paper, Hackenburg et al, explicitly prompted the models to adopt different persuasion strategies, including various techniques suggested by recent decades of psychology research such as moral reframing or deep canvassing. The single most persuasive strategy from this set? “information-based argumentation (in which an emphasis is placed on providing facts and evidence)”
(The other persuasion strategies used, listed in this plot, are “Debate” - parliamentary style debate; “Mega” - prompting the model to use all strategies; “Norms” - focusing on demonstrating that others agree with the position; and “None” - just prompting the model to “be persuasive”).
Not shown here is that the authors also analysed persuasion across all the dialogues coming out of the models and found that, regardless of which explicit strategy the models were prompted with, the thing that predicted persuasive effect was the extent of the information content.
This is great! Language models are not a kind of magical persuasive sauce - there are not Words of Power which unlock the unsuspecting citizen’s mind. Instead, language models persuade using the same interface that any of us persuade - citing evidence and making arguments. What’s more, these results show that the whole grab-bag of persuasive tricks from psychology is less important than backing up relevant claims with facts. This includes personalising the message. The Hackenburg study tested the persuasive effect of individual personalisation and found it was small, although non-zero (this fits with other studies, which have found a zero or small effect of personalisation).
Zooming out, both studies show the power of the interactive mode for carrying persuasive information. Dialogue is more persuasive than broadcast. If language models present any kind of disruption to the dynamics of human persuasion it is not because they can do anything like a ‘superhuman’ levels of persuasion. Rather, it is because they can scale the standard mode of human persuasion, talking. Hitherto, the most persuasive technology we knew required a human on the other side of the conversation.
That said, the models tested in these studies weren’t even that persuasive. They didn’t unambiguously outperform a static message, and the measure of persuasion - self-report of position on a topic before and after - might not last that long, or translate that strongly into action. The results fit with a range of other studies which show that persuasion is generally hard, while acknowledging also that people tend to adjust their views to new evidence.
We can add to this that any persuasion by a language model is unlikely to happen in a vacuum - anyone persuaded in one direction is likely to experience persuasion in the other direction (especially in the context of an election campaign). The factors that led someone to a belief in the first place are likely to reassert themselves after an encounter with a model. Altogether I’m more and more reassured that language models aren’t going to disrupt the human information environment any more than it already is. In terms of disruption, the effects of social media - the takeover of short form, low context, monetised, acutely public information in the common media diet - are surely greater.
Extrapolating out, we can think about the possible downstream effects of persuasive models being widely deployed, but incorporating what these two recent studies show about how they would work their persuasive effect. What does it mean if language models can persuade best with facts and evidence?
The pessimistic view is that there will then be an arms race to recruit different facts to support different political poles. If the facts can’t be found they’ll be made up, and if they can’t be made up directly then the fact generating machinery of society will need to be co-opted so the ‘right’ facts get into the language models. The value of facts will become devalued, as everyone (and their language models) tries to find, misrepresent or invent, facts which support their favoured view.
Writing this out, it feels worryingly like it happens already, regardless of language models being widely deployed for persuasion. Language models might make a bad situation worse, but as you’d expect from this glass-half-full newsletter, I don’t think the value of facts has dropped, or will drop, to zero. Some corruption isn’t complete corruption. There’s more consensus and more trust about than the media represents, still.
If models are poisoned by made up facts then people will come to trust them less. The pessimist account assumes that people will become trapped in some kind of epistemic free-fall where there are no corrective mechanisms or filter of credulity, so the trust in what models say remains even as the reliability of what they say degrades. The two results I’ve written about here show that information presented by language models is persuasive but that doesn’t mean that all models, for all time, will be persuasive - especially if models demonstrate themselves to be unreliable and start pushing out lies.
What this makes clear is that a key issue for persuasion is that people believe and trust model output. Participants in these experiments bring an assumption of trust, or at least benefit of the doubt, to their interaction with the models. If that disappears, so does the persuasive power.
This puts some limit on the scare story that language models will release a ‘manipulation machine’ on society. Imagine retelling this story with books in the place of models. “Books are persuasive!” Because they contain facts and arguments. “But what if the books were made up of lies? Then you could use books to persuade anyone of anything!” We’ve had the ability to create books full of lies for thousands of years. I know enough about the religious wars of the 17th century not to lightly dismiss the risks of new forms of persuasion, but my argument is that a realistic understanding of the risks of persuasive language models needs to be understood as a continuation of the history of human persuasion, not a disjuncture.
That history includes the faltering, imperfect, but persistent power of human reason. The optimistic extrapolation of the results on the persuasive effect of language models recognises that it is possible to agree on some features of reality, despite conflict and context around all claims. This common ground is supported by the epistemic institutions of society: discussion one-to-one between individuals and small groups, the media, fact-checkers and the brute appearance of facts on the ground. As long as this holds, then persuasion in all forms - from demagogues, to books, to language models - has to work with the grain of reality, meaning that persuasion against the facts may be possible temporarily, but the advantage will be on the side of those making a case which best fits the truth. We should have no special fear of language models. If the models have to work with the same facts as spin doctors, influencers, pundits, politicians and conspiracy theorists then we already have the equipment to deal with them. The diagnosis must be the same: rigour, nuance, factfulness, dialogue.
This newsletter is free for everyone to read and always will be. To support my writing you can upgrade to a paid subscription (more on why here)
Keep reading for more on fact-checking, AI, and updates from my life
A side note on accuracy
Hackenburg et al did an analysis of the accuracy of the facts deployed by their persuasive models. They found that the more persuasive, more advanced, models were less likely to be accurate, but they also found that (in)accuracy didn’t correlate with persuasive effect. Further, explicitly instructing the models to be as persuasive as possible regardless of the truth - i.e. to encourage dishonesty from the models - didn’t increase the amount of persuasion. Their conclusion is that this is compatible with a “side-effect” account, where greater inaccuracy doesn’t cause greater persuasion, but can result from other changes which the models makes in the pursuit of great persuasion. To me this sounds a lot like the formal definition of bullshit. It isn’t just models which can decide what to say based on how things sound, regardless of whether they are true or not.
I still maintain that long-term, if models sacrifice accuracy to persuasion people will learn not to trust them and so their persuasive effect will erode.
References
Lin, H., Czarnek, G., Lewis, B. et al. Persuading voters using human–artificial intelligence dialogues. Nature 648, 394–401 (2025). https://doi.org/10.1038/s41586-025-09771-9
Hackenburg, K., Tappin, B. M., Hewitt, L., Saunders, E., Black, S., Lin, H., ... & Summerfield, C. (2025). The levers of political persuasion with conversational artificial intelligence. Science, 390(6777), eaea3884. https://doi.org/10.1126/science.aea3884
Related from me:
Mike Caulfield: We’re not taking the fact-checking powers of AI seriously enough. It’s past time to start.
A walk-through of using AI for fact-checking which is illuminating enough on its own, but Mike also draws out three important take-aways:
#1 Stop assuming AI errors are hallucinations : the usefulness of the term “hallucination” stops when you get to investigating the different kinds of errors possible
Was it conflating? Leaning too heavily on social media sources? Misreading a primary document? If it was “muddled”, where did that muddle likely originate? Or was it truly a “hallucination” in the classic sense, bearing no discernable relation to any set of known claims or documents.
#2 Add AI to your toolkit to answer the questions you didn’t think to ask
This means that AI won’t just give you the answer to the question you thought to ask, it will also provide the context you *should* have asked about. This seems to be a fairly decisive advantage for AI, moving it beyond the simple Question-Answer / spicy autocomplete mode that many people assume when criticising it.
#3 Pay for a model already
I remain shocked at how many reporters and even fact-checkers do not use a paid model of AI — or even know it makes a difference.
It makes a huge difference!
I remain convinced that there is no future of verification and contextualization that doesn’t involve both better understanding of LLMs and more efficacious use of them. The three simple suggestions here — don’t prematurely dismiss errors as hallucinations, do use LLMs to surface the “unknown unknowns”, and do pay for the better models — are all pleas to engage more fully with this tech, and develop better (and more up-to-date) understandings of what they do well and what they do poorly, going beyond whether they fail or succeed to how they fail and how they succeed. This is also what we need to do in education. If we can start from there I’m confident our effort and engagement will be well-rewarded.
Link: We’re not taking the fact-checking powers of AI seriously enough. It’s past time to start.
Talks
I’ve updated my Talks page, so if you’d like to hear me speak about some of the topics covered in this newsletter (and some others) look out for upcoming events online, in London and Liverpool.
Related: I’m in London (and Sheffield) weekly, so let me know if you’d like to meet up.
CATCH-UP
Recently from me:
Life news: London calling / Tools for Thought An update on my career break, plus a strong endorsement for Advait Sarkar’s vision of AI. Original career break news: Some personal news
New work :The Ideological Turing Test: Do you truly understand those you disagree with? and Community Notes require a Community: How we used a novel analysis to understand what causes people to quit the widely adopted content-moderation system
Commentary: When less (communication) is more (collective intelligence): Team work, complex problems, and preserving diversity in the ecosystem of ideas. Listening to Tommy Robinson: What did I learn by giving the right-wing activist my ears for 90 minutes? Helping people spot misinformation: And the greatest gift psychology gave the world
… And finally
Strong early contender for snowperson of the year (via)
END
Comments? Feedback? Reasoned persuasion? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online
Correction 2025-12-12: Reader Jan Zilinsky pointed out in substack comments that the paper didn’t computer before-after change in participants beliefs directly. Study author Ben Tappin provided clarification; “We did measure before and after beliefs, but the outcome in the analysis is after-beliefs only, and we compute the difference in means between treatment and control groups (i.e., average treatment effect) to quantify persuasion. We do however adjust for the before-beliefs in the analysis because this increases precision on the average treatment effect estimates. (These are standard best practices in persuasion research.)”
This is a good overview of recent AI-assisted persuasion research. It's notable that even when people are talking to robots, learning new facts seems more persuasive than exposure to various framing or priming strategies.
I read these a bit ago just after reading a survey paper about AI generating responses to particular types of contentious requests. It’s an interesting emerging space and I agree the factuality piece was encouraging. It also means that politicization and ref-working is going to accelerate with a vengeance because what the model decides is “evidence” is informed in part by what it discerns as “reliable sources”.