AI Systems Out-Persuade Expert Humans, Including Professional Canvassers And World Championship Debaters
from the what-would-Jane-Austen-say? dept
Persuasion plays a key role in society. Whether it is political or financial decisions, workplace or family choices, or simply reading a book or article (like this one), often someone is trying to persuade someone else to agree with them, possibly by changing their mind. This raises an interesting question: if persuasion is such an important part of life, how good are the latest AI systems in this domain? Are they, for example, better than humans? That is what a research project has just investigated, and on an impressively large scale:
in a series of four preregistered experiments (n = 18,978 conversations from 6,923 people), we pitted AI systems against a range of human persuaders, including laypeople, winners of a separately preregistered four-round online persuasion tournament, professional canvassers, and world championship debaters.
The results were unequivocal:
We found that AI systems were reliably more persuasive than expert humans, even when expert humans chose their issues, researched in advance, underwent hours of live, structured practice, and were incentivized with £1,000 cash bonuses. In a follow-up study, AI’s advantage persisted after experts received a coaching tool that let them practice against the AI that beat them, review their performance history, and see what AI would have said at key moments.
An arguably more demanding test found that AI systems were not just persuasive when it came to opinions, but also in terms of real-world actions: they managed to elicit substantially more real-money donations to charity than well-paid professional canvassers. The researchers were able to pin down the two key factors that helped AI to out-perform the best human persuaders in all these tests:
We found converging evidence that AI’s advantage stemmed from rapidly deploying larger quantities of information: after coaching, expert humans could tie an AI constrained to respond at human speeds and with human-length messages.
That is, AI systems were more persuasive largely thanks to the range of knowledge they could demonstrate, and the speed with which they could present it — precisely those aspects of AI that are improving all the time. Which means that frontier AI systems are likely to become even more persuasive in the future. That sounds a rather bleak prospect, but a commentary from Tom Stafford, professor of psychology at the University of Sheffield, and co-author of the book Mind Hacks, points out that things may not be as bad as they seem:
fact-based persuasion may indeed be effective, but that is good news for human reasonableness, not bad. The way the AI works isn’t some sinister magic; if it produces more facts, it is more persuasive. The constraint that persuasion requires evidence means that what anyone can be persuaded of will ultimately ground out on what can reasonably be claimed about reality. If AI is a tool which produces better-informed citizens and more respect for facts, that can be a positive thing.
That may be true in general, but the original researchers note that there are other factors at play here. For example, access to resources is clearly important:
power could flow to whoever can most readily access and deploy the most capable systems. In practice, that could mean the actors who already command the most resources, such as large private corporations, political campaigns, or nation states. These actors spend heavily to influence public opinion and consumer behaviour, and although the per-message effects of such efforts can be modest, such AI could raise their effectiveness, deepening existing imbalances in who can sway the public.
Another issue is that the persuasive power that comes with the deployment of leading AI systems could increase the clout of top AI companies:
in persuasion contests where both sides can secure access to the most capable systems, such AI could consolidate power by giving significant leverage to the actors that build and control those systems. These actors could tilt the outcome of such contests by, for example, deciding which positions their models will, and will not, argue for. In this case, power would flow not to the users of persuasive AI but to its suppliers, and consolidation of their influence would occur even when access among users is perfectly equal.
More positively, the researchers point out that as constant improvements in technology push down the cost of using persuasive AI
it could help under-resourced actors (e.g., pro se litigants and public defenders, small charities, grassroots activists) compete against more established and better-funded rivals, narrowing long-standing gaps in access to justice and assisting civic advocacy more broadly.
In his blog post, Stafford mentions another factor to consider:
In a world where every surface becomes filled with persuasive text, I don’t think it is inevitable that people will open themselves to being pulled in every direction. Not only do people have a significant degree of native scepticism, tending to resist persuasive efforts as they seek to maintain stability in their existing views, but they also have agency to open themselves, or not, to persuasive effects. The studies reported in this paper asked for an average of 14 minutes of conversation from participants. 14 minutes of sincere engagement might be a lot more than most of us give to alternative points of view in our daily lives.
In other words, we don’t really know yet what impact these highly-persuasive AI systems will have on politics, business, and everyday life. But given their superior ability to convince it seems likely that we will be encountering them more frequently in their role of indefatigable persuader, whether we want that or not.
Follow me @glynmoody on Mastodon and on Bluesky.
Filed Under: ai, canvassing, charities, coaching, donations, facts, grassroots, investment, persuasion, politics, pro se, public defender, tom stafford, university of sheffield
Comments on “AI Systems Out-Persuade Expert Humans, Including Professional Canvassers And World Championship Debaters”
Average-ass people largely do not want AI—and everything else about the AI industry, like massive data centers. But moneyed interests and rich motherfuckers want it, which means they will manufacture consent until people always believed they consented, even when they didn’t.
I wish Techdirt wasn’t willing to help manufacture that consent.
This comment has been flagged by the community. Click here to show it.
Re: Aww, call a whaaambulance
I read an opinion I didn’t agree with!
How dare you manufacture consent you bastards!!!
This comment has been flagged by the community. Click here to show it.
Re: Re:
Eat shit.
Re: Re:
Found the uncritical AI evangelist.
Re:
‘Copyright is bad and must be crushed no matter what the odds are against that happening, also AI is inevitable, cannot be stopped so people should just accept it as the future as it makes everything in their lives worse.’
Techdirt, y’all.
Re: Re:
I mean, that’s not really the vibe I get from this article. Any given Geigner article, sure, but this article says “That sounds a rather bleak prospect”; it goes on to say “things may not be as bad as they seem” and quote some more positive hypotheticals, but I’d say this article represents a more skeptical look at the growing use of AI than some of the other ones we see on the site.
I also tend to agree that “it seems likely that we will be encountering them more frequently in their role of indefatigable persuader, whether we want that or not,” at least in the near term. They are, for the moment, a cheap and easy way of spreading propaganda and choking out reliable sources. I’m hoping that won’t be the case forever, but I think it’s likely true over the next couple of years.
Re: Re: Re:
One of my main worries is what will happen over that next couple of years, or however long it takes. Chatbots/AI serving as yet another way for people to split off and stop living in a shared reality is really bad, since we’re still figuring out how to deal with the way our information and media ecosystems have already fractured our sense of shared reality.
I am loathe to link to Substack. It’s a nasty place. Sadly, it’s the only place I can find where Don Moynihan wrote and posted an article, “A Hurricane of Lies” back in October 2024.. Exhausted first responders had to deal with falsehoods and propaganda nonstop, falsehoods and propaganda that threatened their lives, while trying to save and help people after Hurricane Helene. A very real potential scenario is that another massive hurricane hits, and a blend of ChatGPT, Google’s full-page AI overviews, deepfakes on Facebook, and questions asled to Grok further serve to split people off from reality, and first responders wind up getting hurt or worse by the people they’re trying to help.
I straight-up don’t know whether or not we have another couple of years or however long of functional society & institututions that can weather the storm until we reach some point where we can all get back on, or near, the same page.
Re: Re:
That’s not what this post or any other post on TechDirt is saying.
Re:
The problem with the lie machine isn’t that it’s wrong. If it was just wrong everyone would just point and laugh.
It’s that it’s persuasively wrong. And that’s dangerous.
Re:
And I wish you were not helping manufacture that consent by treating their marketing lie about “intelligence” as the truth, but what can we do? “Average-ass people” probably didn’t want the work week to stagnate at 40 hours over the last century either, or to have little choice but to deal with and work for giant faceless corporations (remember the brief period when “Electronic Artists” was non-ironic?; I assume MGM’s slogan of “Ars Gratia Artis” initially meant something too). I don’t see what’s special about the latest thing.
Ultimately, people just don’t care as much as you thing. Some minority might be vocal, but even among them, few are likely to give up even non-essential things like new movies or computer games.
Re: Re:
I’m going to assume you’re not trying to do some parasocial bullshit every time I post about AI in the same way that one poster kept jumping up my ass for months about non-violence and the like in response to Trumpian bullshit. You get that benefit of the doubt once.
THAT SAID. I don’t believe LLMs/what the average person calls “AI” is any more intelligent than the web browser I use to view Techdirt. But if we’re going to talk about what is colloquially known as “AI”, I’m going to call it “AI”. No insult, threat, or other combination of words you can string together—by yourself or with the help of an LLM—will make me reconsider that position. I’ll gladly do a little back-and-forth about what terms we can use for what “AI” does; I recently stopped using “hallucination” in re: AI, for example, after reading a post that made the point in how “hallucinations” are actually malfunctions and thought “yeah, that makes sense”. But I will call it AI—regardless of its lack of intelligence—until and unless you can produce a more widely accepted name for the technologies that fall under the “AI” banner, generative or otherwise.
If you’re thinking about going at me again for continuing to use “AI”, do it anywhere but here. I shove enough of my bullshit onto this site without needing to add more by way of you pestering me about my goddamn language like a grade school teacher. And I promise that I’m not a hard man to find.
I think they can be persuaded to care. And I think the words of the AI evangelists and dumbshit tech bros who keep pushing AI as the end-all, be-all, replace-every-job solution to the world’s problems are helping them care much faster than anything I might have to say. All of those assholes are talking all that good shit about how AI is going to decimate the job market and put millions of people out of work; none of them are talking (seriously) about how people will be able to have their needs taken care of when they don’t have any money in an AI-took-all-the-jobs world.
That besides, generative AI output is already widely reviled and largely unimpactful. The only AI-generated work that has had any kind of widespread cultural impact that isn’t an “ew, fuck that” reaction is the song “We Are Charlie Kirk”, and that…really didn’t have the kind of impact Kirk supporters likely hoped it would. And if you haven’t seen how resistant people are getting to data centers being built in their cities/states, you haven’t been paying attention. People care more than you may be willing to admit; when the bubble bursts on AI, it won’t be mourned any more than NFTs were when that fad came and went faster than The Flash at an orgy.
Re: Re: Re:
That’s fine, and you can call it what you like, but that’s precisely how “manufacturing consent” works. Terminology and ideas just become “normal” and accepted, and the people using it don’t see themselves as part of the problem. So it’s hypocritical to repeatedly admonish others about it.
Re: Re: Re:2
So what fucking term should I use, huh? Go ahead, give me the order to use whatever fucking language you demand I use. Do it. Right now. Give me the fucking dictate and I’ll follow it to the letter.
Tired of you motherfuckers doing this shit to me. Tired of feeling like I can’t fucking win at anything unless I give up who I am and what I believe to make fuckers like you happy. So tell me what to do so you’ll get off my fucking back, okay? Tell me the exact to-the-letter-and-syllable language I have to use so you’ll leave me the fuck alone about talking about…well, I can’t call it that any more, can I? Not without you shoving your hand down my throat and reaching for my windpipe, anyway. But if you promise to let go, I promise not to use that term any more and I’ll say everything you want me to say. Having your hand up my ass like a puppet isn’t any better, but at least you won’t be whining and bitching at me any more.
Something is off here
The findings reported here seem to contradict a lot of other research which found that the piling on of facts when trying to persuade resulted in less success:
https://www.bps.org.uk/research-digest/why-it-so-hard-persuade-people-facts
https://today.uconn.edu/2022/08/cognitive-biases-and-brain-biology-help-explain-why-facts-dont-change-minds-2/
Did the participants in the reported research know they were talking to “ai”?
This seems like techdirt being sloppy in their reporting, or worse…shilling for the expensive autocomplete chat bots…???
Re:
Here’s the thing, that makes sense. LLMs do not persuade with volume of facts, but rather their tone of delivery, emotive appeals, and the way they butter up their target.
My source is...
AI: I made it the fuck up
Pretty easy to persuade if you make shit up and lie (cough) I mean hallucinate in a generative manner…
I remember reading a story about an in-development game show where children would listen to two people discussing a particular specialist subject and guess which one was a real subject matter expert and which was an actor.
The show never made it past development because it didn’t work: the children would consistently find the actors more believable than the real subject matter experts.
The thing is, knowing stuff is an entirely different skillset than persuading people of stuff. And AI is very good at making statements that read, to human psychology, as confident and convincing. But it doesn’t know anything except statistical probabilities and repeating patterns within its training data.
Re: The Gish Vroom
I suspect that a lot of people wouldn’t find the researchers’ negative point so bad, if they assume that more convincing arguments approximate “the truth” more closely. If AI is better at approximating the truth, who cares if it’s unfair, especially since debates make access to that truth public?
But assuming the AI is evaluated on both sides of every debate, and still outperforms their human opponents (if they didn’t do that, that would be a huge source of bias), then if there is a singular truth, or even one relative to a stable reference point, we cannot assume that a convincing argument is anywhere near it.
The research claims that the way that the AI wins is by making Ben Shapiro sound like a sloth (I know people who talked like him in high school debate, but they usually knew when to turn it off), which would be great if every fact was true or just verifiable in real time, but unfortunately, even pre-AI, that has never been the case.
Re:
Persuasion machines built by companies desperate to make back their billions in losses? Persuasion machines built by those that believe that there’s a “woke mind virus” and make child porn? Hard, hard pass.
Re: Re:
Man, I never had a Twitter account in the first damn place; I’m not sure why you think you need to tell me that.
Re:
Luckily (or unluckily depending on how you look at it), this isn’t a new problem. People have been handing political and economic power to people who project confidence over demonstrating ability for millennia.
An essential part of growing up and becoming educated is not just learning facts but also learning to evaluate people who claim to have them. Many adults are not great at it (as evidenced by the endless history of scams). Hopefully in the future, by necessity, society starts more rigorously training kids to spot a charlatan.
Fascinating. I would like to see an AI winning a debate about Christian Nationalism, helping religidiots becoming more reasonable.
Not in my experience
I push back, they fold.
i asked why the lego site didn’t have an “all sets” view and got told it was a limitation of the tech, there were too many sets.
i replied “bollocks, modern CRMs handle that shit no probs”.
“oh yeah, my bad, it’s a choice they make so they can show targeted promotions”.
So, I’m wondering, how stupid are the people it “persuaded”?
Re:
LLMs have consistently tried to convince me to mix citric acid in my hydroponics reservoirs. That would be bacterial bloom and lost crop, for those keeping track at home.
Re:
I think what the study discusses is not a general purpose chatbot, like one you interact with, but a specifically trained tool, designed to persuade. That’s not the same thing.
Re: Re:
You hallucinated.
From section 2.6.1 of SI_Appendix.pdf in the authors’ repository:
Re: Re: Re:
New “rule”: Stop referring to AI malfunctions as “hallucinations”. That implies LLMs/generative AI models have some kind of higher brain function or even sentience. They don’t—they’re still programs and their fuck-ups are malfunctions.
Re: Re: Re:2
Mike has the nerve to complain that the AI my coworkers are letting do work for them (which is good at persuasion but gets the facts all wrong… a problem if you’re working in law) is a ‘stupid use for AI’ and then to praise that quality about it here in the comments.
A little consistency would be nice. Or does Mike’s brain just shut down when he hears AI?
Re: Re: Re:3
WUT?
Pointing out a stupid use of AI… and then also pointing out a use where AI is effective… is… not inconsistent?
Or do you… not actually realize that a technology can have both good and bad uses.
Re: Re: Re:4
You’re claiming the exact same use is both good and bad. That’s extremely inconsistent.
Re: Re: Re:5
I am not. We are talking about two totally different uses. Also I didn’t write this article, which only describes the results of the study.
You seem to have trouble understanding fairly basic concepts like “different things are different.”
Perhaps work on that?
Re: Re: Re:4
At my place of work, which is trying to adopt AI because, fuck, I guess we have too much money or something I am always having to correct the WRITING of my coworkers who use AI because they use it to WRITE things for them and what it writes is persuasive sounding but deeply incorrect. It miss-cites cases, or cites imaginary cases, or makes up entire laws that do not exist.
You said “That sounds like a dumb use for an AI”
Now you said “Hey, a persuader AI (that is constantly wrong) is a GREAT USE for AI”
My big complaint about AI is that it’s wrong, and persuasive. This is why I call it the “Lie machine”.
So… keep your AI in it’s extremely narrow niche you think works in, and stop praising all goddamn uses of it.
Re: Re:
Oy, mike, this ‘persuader’ ai sounds exactly like the piece of shit at my place of work you think is a stupid use of AI. Why you defending it?
A research organisation founded to determine if AI is powerful and dangerous (in a way that sells AI) finds AI to be powerful and dangerous (in a way that sells AI) using not very rigorous methodology? I’m convinced!
Not really surprising...
Systems designed for the purpose of confidently bullshitting actually pretty successful at bullshitting. Such news.
Funny. As often as not, I find I have to dissuade the varying LLMs from the digital equivalent of wearing their pants on their heads.
It's not persuasion, it's propaganda
Real persuasion is done slowly, methodically, and carefully by the presentation of facts and analysis — as in the scientific method, which — for all of its flaws — is the best mechanism we have for discovering truth.
This isn’t that. It’s not even close. It’s psychological manipulation that’s explicitly designed to exploit the weak points in human psychology. It’s a torrent of bullshit delivered with sufficient assurance to overcome listener skepticism — with no correlation whatsoever to truth.
TD’s cheerleading for the thugs running these AI companies is starting to get old. These aren’t visionaries or leaders; they’re psychopaths and sociopaths who are pursuing a fascist agenda.
Use an AI to rewrite this article then because I find myself unpersuaded.
AI systems are built to scrape and reproduce the best work of the most convincing con artists. They use techniques that dig right into the deepest hardwired bits of our neurology to create trust. They are very good at it. We are all in so much trouble.
I’m still not convinced that AI will put master debaters out of business.
Re:
Don’t worry, there will always be demand for cunning linguists.
“fact-based persuasion may indeed be effective”
Why should we assume it would be fact-based? It would be so easy for AI companies to program their models to lean one way or another on particular issues that are important to them without anyone knowing. The AI will still be able to response with facts supporting its position, even if its position is less fact-based when looking at the totality of the evidence.
This finding, if it holds up, is incredibly frightening for the future of politics and information.
LLMs are the ultimate used-car salesman.
Facts?
I’m curious why “facts”* seem to work to persuade people when they come from an LLM.
(*Even in the likely event that these “facts” turn out to be false.)
There’s a mountain of research suggesting that people, generally, don’t find facts persuasive. Prevailing research instead advises that average people are more likely to be convinced by sentiment, framing, and their own communities than by the introduction of new information. This is why most internet arguments end up feeling like frustrating wastes of time. Ditto for political arguments against your relatives at Thanksgiving.
In that context, the study’s conclusion that the LLM simply has more FACTS that it can deploy in support of its argument, strikes me as suspicious. There are a few possibilities for why this paper might contradict all the aforementioned research on this topic. Either…
In reading the “Methods” section of the paper, one possible flaw I noticed was that the “persuadees” don’t appear to have been screened for which side of the issue they were on prior to the experiment. Perhaps this was something taken into account, and the paper doesn’t explain it thoroughly, but I’m wondering if their recruiting happened to self-select for people who were already inclined to the same side of the argument that the LLMs were arguing for?
Lastly, I’d also like to point out the unusually large amount of money involved in this paper, which was spent on recruiting literally thousands of paid participants, some of which were paid at high tiers based on their professional qualifications, and furnishing a prize pool with incentives (up to £1000) for the best-performing human persuaders. This isn’t disqualifying in and of itself – after all, studies with large sample sizes are generally more reliable than ones with small sample sizes – but it does beg the question of where the money came from to pay so many participants. There’s no financial disclosure section in the paper.
I don’t think there’s enough information in the paper itself to explain EXACTLY what’s wrong here, but something about this paper doesn’t smell right to me.
Re:
They did check for previous beliefs.
I would also point out that in this study, the success of human persuasion was also correlated with information density, so to the extent there is disagreement with prior research, the phenomena is not specific to the AI.
As to the interpretation, the authors listed two other (AI persusasion) related papers as additional support of their “more facts” interpretation.
The first is paywalled, though the abstract results does not indicate any “fact-specific” mechaninism. The paper specifically discussed reduction of belief in conspiracy theories, and would rely entirely on prior work for any meaningful controls (the only internal control was some participants having an unrelated conversation with the AI). Hard to say anything more about it.
The second article included a rather larger set of participants and political positions, though no human arm. It concluded that information density was the largest predictor of successful persuasion, but also “where they increased AI persuasiveness they also systematically decreased factual accuracy.”
Specifically, when we say facts we are talking about “claims that could be falsified” not “claims which are true.” The most persuasive model made 4x more factual claims than the average model, with 30% of those being lies (compared to an average of 16% lies).
Which could actually support the “previous studies didn’t have enough facts” argument: humans may well be worse at coming up with long series of highly confident “factual” claims. We admit we don’t know things, we express the level of uncertainty in data, we prevaricate, we’re rarely great liars and even more rarely are great liars also subject matter experts.
Maybe that was what we needed the whole time, a subject matter expert who won’t admit uncertainty and will lie as naturally as they breath.
Re: Re:
Thank you for correcting me about the pre-screening question, and for surfacing that additional context from behind paywalls. I also very much appreciate you calling out, “Specifically, when we say facts we are talking about ‘claims that could be falsified’ not ‘claims which are true.'”
In the end, I remain skeptical of the paper’s proposed mechanism for these conclusions, but the sheer size of the study doesn’t leave much room for doubt about the conclusions themselves.
Uh, no. AI is the ultimate psychopath. That’s why it is persuasive.
I’m very fortuitous in that ‘people who are good at persuading’ are generally very bad at persuading me.
That’s a little autism superpower. :p
Oy, Mike. The being wrong machine still does not belong in law.
The rise of AI in persuasion certainly reshapes how brands engage with their audience. As AI systems become more influential in decision-making, understanding their impact on brand visibility in search results is crucial. How do you see the role of brands evolving as AI-generated content becomes a primary source of information for consumers?