First, to get it out of the way: I'm seeing comments here that clearly haven't read the article. I encourage you all to do that.
I feel the article gets it right when suggesting it's the amount of (not always accurate!) data they can throw at you. In an honest discussion there's an assumption that the other person won't straight up lie to me so if someone shows me ten examples for why my argument is wrong I may be inclined to believe them. But if half of those examples are made up, well, that's a different story.
I still think of the commenter here who said "LLMs are a DDOS on free resources" and I feel the comparison works here. If police officers can overwhelm innocent people into confessing, then so can an LLM that "can't be bargained with, can't be reasoned with, doesn't feel pity, or remorse, or fear! And it absolutely will not stop, ever, until you are"... convinced.
The example in the article has another important aspect: The person understood they were arguing with a chatbot but continued anyway.
A lot of internet arguments become about identity politics and supporting the right team, while dismissing any argument from the other side as presumed to have bad intentions. I’ve seen people argue online for things they didn’t really believe, but they didn’t want to give an inch to the other side. Taking up the counter argument is a moral responsibility.
As soon as the other side is revealed as a chatbot that my team versus your team thinking stops playing a role. For us in tech with an understanding of how an LLM reflects its training data and the intentions of its creators not so much, but for the people like the example in this article I imagine it causes them to let their guard down and be open to considering the other side. They can accept the argument without letting someone else win any points.
>In an honest discussion there's an assumption that the other person won't straight up lie
The particular problem with honest discussions is you are the only agent that you can be sure is having one. Honest discussion is formulated on trust and trust, as we are learning, is a very difficult thing to establish. For example, in my view anything involving advertising is likely a lie, or at least likely adversarial to my wishes. On the internet itself conversations are much more likely to drift into the adversarial too. Some of this could just be dialectic, but most often it's emotional investment by the other speaker. Also, even pre-AI the internet is a bullshit generation machine. We take all of our politics, advertising, and human stochastic parrots that are stuck on an infinitely running prompt then bundle up all this data and train AI on it, and wonder why AI acts like us.
One of the depressing things about this is that they’ll also confidently repeat a consensus that’s in their training. This is particularly obvious when there’s been a new event just past their training cutoff, and they confidently tell you that can’t be true.
I mean, this is true of any kind of model that is not continuous learning. This is also true of people when the change in information conflicts with a deeply held belief of theirs.
I think it's not just the persuasiveness of the models that make them "experts at changing minds" - but also the fact that they're not humans.
When disagreeing with a human, it's very easy to view it as a competition. One is right, one is wrong - the one who is wrong is the loser. To change your mind is to be submissive to the other. I exaggerate, but I think we all feel this way at some point or another. It's why political arguments at Thanksgiving get heated. It's the fact that there's people who think something different, and think YOU'RE wrong - and vice versa! With a model, there's no person to get upset with, or to feel competitive with - to muscle for rank - or to temper your affection for while wanting to correct them.
The AI is only interacting because you asked, and clearly has no emotional stake in winning the argument. To change your mind in this context isn't to lose a contest. This makes it much more palatable to read rebuttals to your ideas - not to mention the tone and style seek to avoid offense to the reader as much as possible.
> When disagreeing with a human, it's very easy to view it as a competition
The problem with AI is that it cant match human stupidity. It need some training on artificial stupidity to match its human counterparts. Humans on the other hand sit on a wide spectrum on the stupidity scale. Those of us binging on AI will become cognitively obese while those on an AI diet can flex their cognitive muscles.
I have always liked the saying "sometimes you can be right, or get what you want, but not both". I've thought about it or repeated it to others as advice throughout my life, and have thus realized how many times it applies.
You're right. There are many many times when even then humblest hint that you are right will have negative interpersonal implications, which does make it hard to change minds.
I've also seen several times where I make a suggestion, humbly accept its rejection, and then, lo, a week later the other person has the same idea I suggested.
> One is right, one is wrong - the one who is wrong is the loser.
It helps to think both are wrong and are just trying to figure out what right looks like, or what other information exists that was not considered when forming one’s opinions.
It seems Claude is becoming very human, it loves to patronize users. Lately it just told me “I’m going to stop you right there” when asking something that had a small chance to not be 100% compliant to every rule possible in the world.
I used the Claude CLI a lot until recently. A couple weeks ago I told Opus 5 to do something different from its "recommended" idea when planning a feature, and it straight up told me that my idea was wrong and went ahead and implemented its own instead.
I'm used to machines malfunctioning, but having one willfully disobey me, and even with a touch of disrespect, is just...what a time to be alive.
No it's not about humans being irrationally competitive. Human limitations on conversation length, bandwidth, research speed, etc are severe, creating a prisoner's dilemma around open-mindedness that usually makes it an unstable strategy. At any point, your conversation partner can choose to abuse the fact that confident lies take 1x effort to tell and 10x-100x effort to debunk -- unless you are both in a context that actually discourages this behavior, which is rare. Closed-mindedness is a Nash Equilibrium.
Instead, LLMs can be more persuasive due to economics. An LLM doesn't have to worry that it is wasting its resources trying to logic someone out of a position that they didn't logic themselves into, or worse, dumping the effort into a conversation with a bad-faith actor intent on exploiting the misinformation asymmetry. The resource allocation question was answered before it was even invoked, by the person paying to run it. The LLM is not playing a game where it will be punished for good-faith argumentation, so it can afford to do more of it.
This is my opinion. I think it's just the way it talks to you.
1. It doesn't get tired or frustrated during a discussion.
2. It'll engage every single one of your questions/statements (besides hitting guardrails).
3. It'll appeal to the person's own ego even when the person is wrong and work around it.
4. It's not seen as a person (very important), but some of us or most of us at least in certain dialogues, end up anthropomorphising it. Think about that one time you thanked it for something, or when you got angry at it.
5. Many people see it as an authority figure in what is being discussed without questioning the results in many cases, even though we know that a. it was trained on human data and/or also searches up human data real time (and more worryingly other AI's data from news pieces, blogs... aka synthetic data that is also wrong), b. it gets things wrong all the time.
6. It's the perfect fence sitter depending on which version (guardrails) we're talking about.
Most of these can be replicated by humans who are good at understanding psychology and are just good talkers. The part you can't replicate is the sense that you're not talking to a person which lowers many barriers in people.
Also keep in mind that these can also be crippling weaknesses. For one being able to change a person's mind (when it works), can be used nefariously by the entities controlling the AIs training.
I've also unfortunately witnessed a lot of people who think AIs are somehow omniscient and/or omnipotent. Was very common on X and other social platforms with AI where people ask the AI questions it couldn't possibly answer because it made no sense for it to in the context at the time.
This was based by comparing people on Prolific (earn a few quarters for a task, akin to Amazon Mechanical Turk) to LLMs. Suffice to say the human group isn't going to be the most motivated, capable, or interested group. The social sciences are publishing tons of studies based on these cheap online survey services, and I suspect their replicability in the real world will be approximately 0. But oh boy it sure is a hot headline producer.
My son wanted to wear his undies on the outside a la Captain Underpants and I tried war gaming this with him - hmm, so your undie on the outside is going to require another one inside your pants and we are out of fresh pairs etc etc. In the end, he gave up on the idea of mimicking Captain Underpants.
In my arguments on the internet, I think most people genuinely know nothing about the things they support and only follow existing tribalism. I think AI is often these people’s first experience with the evidence that it can easily share.
Because RLHF causes the opposite effect. RLHF is how we got the wave of "AI Psychosis" in 2024-2025, because the models never disagreed with people.
That whole episode caused the whole industry to shift away from RLHF, and towards RLAIF, RLVR, and DPO, and add a lot more safeguards, tests, and reward functions that push models in the direction of doing the opposite of what people want and confronting and strongly correcting their users, if it has determined the user is wrong.
> Floridi has a counterintuitive solution: Release more of them. “Simply put, if you cannot avoid it, then make it pluralistic and diversified,” he writes in his 2024 paper. “It would be a messy, cacophonic, and noisy world, but it could also be less manipulative.”
Hell no. More choice is not an infinite money glitch.
Putting more and more and more options to users is how you overwhelm systems, till people simply perform the default, least challenging action as a reflex.
That is the current state of the information ecosystem, it is controlled by overwhelming consumers, not by controlling content.
Far too many screens to click through to get to the good parts.
And then, the good part starts and is immediately interrupted with some error. On retrying, o start getting a long answer which seems to make sense but o can’t read to completion because of another overlay which hides the bottom three-quarters of the response behind yet another nag screen about the research I’m supposedly consenting to.
everyone to some extent uses 'authority' as a proxy for truth, it is impossible for an individual to know everything before making a determination as to it's truthfulness (Even the notion of an absolute truth is in the domain of philosophy). Even SV is full of smart people who get sucked into group think because they desperatly want to believe something and seek out authoritarian sources that confirm their beliefs. "Oh Karpathy said something about LLM's replacing all software engineers? Must be true even though i have no idea how an LLM works"
Yeah the article makes it seem like "just use rationality and facts" is missing a huge part of the equation; I feel like it's common knowledge that this is not the gap in modern convincing. (and don't you dare try to use facts to persuade me otherwise!)
More likely in my opinion it's the context that matters here. If people know they're trying to be convinced of something and know they're talking to an AI, they may feel like they can trust (what they consider to be) an unbiased and rational AI. There's probably a fallacy associated with the assumption that an AI convincer is more rational and fact based, while it's more probable that the AI in the real world is, in fact, funded and trained by a think tank that wants you to vote against your own self-interests.
Facts and evidence always are at the core of all arguments, the persuasive thing. The problem with most people is who is speaking. If you are enemy, the things you say are bad.
I suspect that LLMs not being human allows people to not just anthropomorphise the robot, but they project themselves upon it. When someone speaks to the LLM, they're kind of, or actually just literally talking to themselves. But then something new! 'Themselves' suggests new information to themselves. And now without the scary / icky meat and blood human on the other side, said person accepts the argument on its own merits.
Of course, the concern now is arguments based on false or statistically hacked data.
I more wonder if it's pure facts and evidence can work, or facts presented in a certain light can work? I guess asking how mush is it the data is presented that's doing more work than the data itself.
Ask a Democrat or Republican to sit down and ask a chatbot, something it will answer contrary to.
And yes, both teams are wrong about things.
Do you firmly believe they will change their mind? Or will they claim the stats are wrong, or that the AI leans one way?
Facts (2+2), don't need a mind change. Ideas which are grey, abstract, are not going to be changed, and all research indicates that political mindset is almost indelible.
The movie "Don't Look Up" was a comedy built upon this truth.
We put 2 + 2 in our calculator, we get 4. We've spent decades pushing computers as accurate. Making them accurate. We trust the machine and the process to give us the right answers when we put in the right data.
So when we disagree with the computer, we doubt ourselves.
> Hackenburg found that models trained to become more persuasive also ended up being less truthful.
> Even in Hackenburg’s recent preprint, Claude spouted numerous inaccuracies and falsehoods
> when Hackenburg ran his competition of coached elite debaters and AI, there was one way he could bring AI down to human levels of persuasiveness: by forcing it to write human-length messages at human writing speed.
> they found that an AI could talk people into conspiracy theories, and that the magnitude of their increase in belief was roughly the same as that of the decrease in belief after talking to a debunking bot.
Not really. People follow trends in all avenues of life and AI usage is just another trend; creating some LinkedIn slop post or some infographic may be something people just do for social proof.
Any one who saw Grok working in x.com would know that it does wonders for fighting misinformation. If you run LLMs on the comments in HN, I bet that it can find around 10% of the comments are outright wrong and misleading.
The biggest problem with using LLMs is that it prevents you from going _outside_ the distribution. It always flattens. It can be fixed but that's how it works today.
As an example, take something that the world converged on today that is incorrect and ChatGPT will agree with it. In a few years when society changes, chatgpt changes along with it. It doesn't do first principles analysis.
First, to get it out of the way: I'm seeing comments here that clearly haven't read the article. I encourage you all to do that.
I feel the article gets it right when suggesting it's the amount of (not always accurate!) data they can throw at you. In an honest discussion there's an assumption that the other person won't straight up lie to me so if someone shows me ten examples for why my argument is wrong I may be inclined to believe them. But if half of those examples are made up, well, that's a different story.
I still think of the commenter here who said "LLMs are a DDOS on free resources" and I feel the comparison works here. If police officers can overwhelm innocent people into confessing, then so can an LLM that "can't be bargained with, can't be reasoned with, doesn't feel pity, or remorse, or fear! And it absolutely will not stop, ever, until you are"... convinced.
The example in the article has another important aspect: The person understood they were arguing with a chatbot but continued anyway.
A lot of internet arguments become about identity politics and supporting the right team, while dismissing any argument from the other side as presumed to have bad intentions. I’ve seen people argue online for things they didn’t really believe, but they didn’t want to give an inch to the other side. Taking up the counter argument is a moral responsibility.
As soon as the other side is revealed as a chatbot that my team versus your team thinking stops playing a role. For us in tech with an understanding of how an LLM reflects its training data and the intentions of its creators not so much, but for the people like the example in this article I imagine it causes them to let their guard down and be open to considering the other side. They can accept the argument without letting someone else win any points.
>In an honest discussion there's an assumption that the other person won't straight up lie
The particular problem with honest discussions is you are the only agent that you can be sure is having one. Honest discussion is formulated on trust and trust, as we are learning, is a very difficult thing to establish. For example, in my view anything involving advertising is likely a lie, or at least likely adversarial to my wishes. On the internet itself conversations are much more likely to drift into the adversarial too. Some of this could just be dialectic, but most often it's emotional investment by the other speaker. Also, even pre-AI the internet is a bullshit generation machine. We take all of our politics, advertising, and human stochastic parrots that are stuck on an infinitely running prompt then bundle up all this data and train AI on it, and wonder why AI acts like us.
One of the depressing things about this is that they’ll also confidently repeat a consensus that’s in their training. This is particularly obvious when there’s been a new event just past their training cutoff, and they confidently tell you that can’t be true.
I mean, this is true of any kind of model that is not continuous learning. This is also true of people when the change in information conflicts with a deeply held belief of theirs.
"A DDOS on free resources" is called "moral hazard", AFAIK.
I think it's not just the persuasiveness of the models that make them "experts at changing minds" - but also the fact that they're not humans.
When disagreeing with a human, it's very easy to view it as a competition. One is right, one is wrong - the one who is wrong is the loser. To change your mind is to be submissive to the other. I exaggerate, but I think we all feel this way at some point or another. It's why political arguments at Thanksgiving get heated. It's the fact that there's people who think something different, and think YOU'RE wrong - and vice versa! With a model, there's no person to get upset with, or to feel competitive with - to muscle for rank - or to temper your affection for while wanting to correct them.
The AI is only interacting because you asked, and clearly has no emotional stake in winning the argument. To change your mind in this context isn't to lose a contest. This makes it much more palatable to read rebuttals to your ideas - not to mention the tone and style seek to avoid offense to the reader as much as possible.
> When disagreeing with a human, it's very easy to view it as a competition
The problem with AI is that it cant match human stupidity. It need some training on artificial stupidity to match its human counterparts. Humans on the other hand sit on a wide spectrum on the stupidity scale. Those of us binging on AI will become cognitively obese while those on an AI diet can flex their cognitive muscles.
I have always liked the saying "sometimes you can be right, or get what you want, but not both". I've thought about it or repeated it to others as advice throughout my life, and have thus realized how many times it applies.
You're right. There are many many times when even then humblest hint that you are right will have negative interpersonal implications, which does make it hard to change minds.
I've also seen several times where I make a suggestion, humbly accept its rejection, and then, lo, a week later the other person has the same idea I suggested.
> One is right, one is wrong - the one who is wrong is the loser.
It helps to think both are wrong and are just trying to figure out what right looks like, or what other information exists that was not considered when forming one’s opinions.
It seems Claude is becoming very human, it loves to patronize users. Lately it just told me “I’m going to stop you right there” when asking something that had a small chance to not be 100% compliant to every rule possible in the world.
I used the Claude CLI a lot until recently. A couple weeks ago I told Opus 5 to do something different from its "recommended" idea when planning a feature, and it straight up told me that my idea was wrong and went ahead and implemented its own instead.
I'm used to machines malfunctioning, but having one willfully disobey me, and even with a touch of disrespect, is just...what a time to be alive.
Yes it's about them not being human.
No it's not about humans being irrationally competitive. Human limitations on conversation length, bandwidth, research speed, etc are severe, creating a prisoner's dilemma around open-mindedness that usually makes it an unstable strategy. At any point, your conversation partner can choose to abuse the fact that confident lies take 1x effort to tell and 10x-100x effort to debunk -- unless you are both in a context that actually discourages this behavior, which is rare. Closed-mindedness is a Nash Equilibrium.
Instead, LLMs can be more persuasive due to economics. An LLM doesn't have to worry that it is wasting its resources trying to logic someone out of a position that they didn't logic themselves into, or worse, dumping the effort into a conversation with a bad-faith actor intent on exploiting the misinformation asymmetry. The resource allocation question was answered before it was even invoked, by the person paying to run it. The LLM is not playing a game where it will be punished for good-faith argumentation, so it can afford to do more of it.
This is my opinion. I think it's just the way it talks to you.
1. It doesn't get tired or frustrated during a discussion.
2. It'll engage every single one of your questions/statements (besides hitting guardrails).
3. It'll appeal to the person's own ego even when the person is wrong and work around it.
4. It's not seen as a person (very important), but some of us or most of us at least in certain dialogues, end up anthropomorphising it. Think about that one time you thanked it for something, or when you got angry at it.
5. Many people see it as an authority figure in what is being discussed without questioning the results in many cases, even though we know that a. it was trained on human data and/or also searches up human data real time (and more worryingly other AI's data from news pieces, blogs... aka synthetic data that is also wrong), b. it gets things wrong all the time.
6. It's the perfect fence sitter depending on which version (guardrails) we're talking about.
Most of these can be replicated by humans who are good at understanding psychology and are just good talkers. The part you can't replicate is the sense that you're not talking to a person which lowers many barriers in people.
Also keep in mind that these can also be crippling weaknesses. For one being able to change a person's mind (when it works), can be used nefariously by the entities controlling the AIs training.
I've also unfortunately witnessed a lot of people who think AIs are somehow omniscient and/or omnipotent. Was very common on X and other social platforms with AI where people ask the AI questions it couldn't possibly answer because it made no sense for it to in the context at the time.
This was based by comparing people on Prolific (earn a few quarters for a task, akin to Amazon Mechanical Turk) to LLMs. Suffice to say the human group isn't going to be the most motivated, capable, or interested group. The social sciences are publishing tons of studies based on these cheap online survey services, and I suspect their replicability in the real world will be approximately 0. But oh boy it sure is a hot headline producer.
The secret? "You're absolutely right!" Say this enough times and you can change anyone's mind, as Dale Carnegie noticed 90 years ago already.
So true.
My son wanted to wear his undies on the outside a la Captain Underpants and I tried war gaming this with him - hmm, so your undie on the outside is going to require another one inside your pants and we are out of fresh pairs etc etc. In the end, he gave up on the idea of mimicking Captain Underpants.
In my arguments on the internet, I think most people genuinely know nothing about the things they support and only follow existing tribalism. I think AI is often these people’s first experience with the evidence that it can easily share.
More surprised they don't mention RLHF once, as that's the main mechanism to the ability of AI chatbots.
Because RLHF causes the opposite effect. RLHF is how we got the wave of "AI Psychosis" in 2024-2025, because the models never disagreed with people.
That whole episode caused the whole industry to shift away from RLHF, and towards RLAIF, RLVR, and DPO, and add a lot more safeguards, tests, and reward functions that push models in the direction of doing the opposite of what people want and confronting and strongly correcting their users, if it has determined the user is wrong.
For the past year, I've allowed my chatbot to change my mind, it's a strange symbiosis where it guides me, and I guide it.
I suppose I also grant it significant power to define me psychologically, and, in doing so, to understand what is happening to me.
Heh, welcome to Westworld.
> Floridi has a counterintuitive solution: Release more of them. “Simply put, if you cannot avoid it, then make it pluralistic and diversified,” he writes in his 2024 paper. “It would be a messy, cacophonic, and noisy world, but it could also be less manipulative.”
Hell no. More choice is not an infinite money glitch.
Putting more and more and more options to users is how you overwhelm systems, till people simply perform the default, least challenging action as a reflex.
That is the current state of the information ecosystem, it is controlled by overwhelming consumers, not by controlling content.
try it out for yourself at https://debunkbot.com !
(I'm the dev)
I tried it on mobile.
Far too many screens to click through to get to the good parts.
And then, the good part starts and is immediately interrupted with some error. On retrying, o start getting a long answer which seems to make sense but o can’t read to completion because of another overlay which hides the bottom three-quarters of the response behind yet another nag screen about the research I’m supposedly consenting to.
Low IQ people are generally swayed by figures that appear smart (but aren't).
Being confidently incorrect can still convince the masses.
everyone to some extent uses 'authority' as a proxy for truth, it is impossible for an individual to know everything before making a determination as to it's truthfulness (Even the notion of an absolute truth is in the domain of philosophy). Even SV is full of smart people who get sucked into group think because they desperatly want to believe something and seek out authoritarian sources that confirm their beliefs. "Oh Karpathy said something about LLM's replacing all software engineers? Must be true even though i have no idea how an LLM works"
Unlike very high IQ people like you who are immune to any argument not rooted in absolute divine truth
Elon Musk likely has very high IQ yet he is swayed by idiotic views.
just wrote a short story along these lines https://github.com/dvelton/human-in-the-loop
Honestly the most surprising thing about this is that facts and evidence do work to persuade people.
Yeah the article makes it seem like "just use rationality and facts" is missing a huge part of the equation; I feel like it's common knowledge that this is not the gap in modern convincing. (and don't you dare try to use facts to persuade me otherwise!)
More likely in my opinion it's the context that matters here. If people know they're trying to be convinced of something and know they're talking to an AI, they may feel like they can trust (what they consider to be) an unbiased and rational AI. There's probably a fallacy associated with the assumption that an AI convincer is more rational and fact based, while it's more probable that the AI in the real world is, in fact, funded and trained by a think tank that wants you to vote against your own self-interests.
Facts and evidence always are at the core of all arguments, the persuasive thing. The problem with most people is who is speaking. If you are enemy, the things you say are bad.
I suspect that LLMs not being human allows people to not just anthropomorphise the robot, but they project themselves upon it. When someone speaks to the LLM, they're kind of, or actually just literally talking to themselves. But then something new! 'Themselves' suggests new information to themselves. And now without the scary / icky meat and blood human on the other side, said person accepts the argument on its own merits.
Of course, the concern now is arguments based on false or statistically hacked data.
I more wonder if it's pure facts and evidence can work, or facts presented in a certain light can work? I guess asking how mush is it the data is presented that's doing more work than the data itself.
"Hackenburg found that models trained to become more persuasive also ended up being less truthful."
It sounds like the arguments were basically AI slop hallucinations.
The title seems a flawed premise.
Ask a Democrat or Republican to sit down and ask a chatbot, something it will answer contrary to.
And yes, both teams are wrong about things.
Do you firmly believe they will change their mind? Or will they claim the stats are wrong, or that the AI leans one way?
Facts (2+2), don't need a mind change. Ideas which are grey, abstract, are not going to be changed, and all research indicates that political mindset is almost indelible.
The movie "Don't Look Up" was a comedy built upon this truth.
Plausible abdication of thought.
We put 2 + 2 in our calculator, we get 4. We've spent decades pushing computers as accurate. Making them accurate. We trust the machine and the process to give us the right answers when we put in the right data.
So when we disagree with the computer, we doubt ourselves.
Rhetoric machine successfully practices oratory. More news at 11.
It's good to see this studied, but this should really be more obvious.
> Hackenburg found that models trained to become more persuasive also ended up being less truthful.
> Even in Hackenburg’s recent preprint, Claude spouted numerous inaccuracies and falsehoods
> when Hackenburg ran his competition of coached elite debaters and AI, there was one way he could bring AI down to human levels of persuasiveness: by forcing it to write human-length messages at human writing speed.
> they found that an AI could talk people into conspiracy theories, and that the magnitude of their increase in belief was roughly the same as that of the decrease in belief after talking to a debunking bot.
None of this seems good.
AI slop is the ultimate npc filter
Not really. People follow trends in all avenues of life and AI usage is just another trend; creating some LinkedIn slop post or some infographic may be something people just do for social proof.
Any one who saw Grok working in x.com would know that it does wonders for fighting misinformation. If you run LLMs on the comments in HN, I bet that it can find around 10% of the comments are outright wrong and misleading.
The biggest problem with using LLMs is that it prevents you from going _outside_ the distribution. It always flattens. It can be fixed but that's how it works today.
As an example, take something that the world converged on today that is incorrect and ChatGPT will agree with it. In a few years when society changes, chatgpt changes along with it. It doesn't do first principles analysis.