My guess for the reason behind this is that LLMs have more facts memorized, and thus can make more reasonable and well-researched sounding answers. If you ask an LLM vs a Human "Is a stack in computer science a) a data structure that is first in first out or b) a data structure that is first in last out" the LLM can say stuff resembling "Based on Dijkstra's algorithm proof given in 1943 and the nature of Turing complete languages being traditionally a top-down oriented system, a stack is ..." while a human is just going to say "It's B because that's what a stack is".
Based on reading bad AI generated student essays it’s worse than that, LLMs are happy to “fill in the blanks” with whatever made up fact would make their argument look best.
Most people can’t lie that smoothly, and most readers don’t check carefully, unless they are already an expert in the area.
Any kind of maths proof is particularly bad, they will look convincing and clear until you read them very carefully and see all the holes.
It's funny you mention this, because my father operates exactly as you describe the LLMs, making facts up on the spot, lying smoothly and keeping track of the lies...
...and he's built his whole career in sales because of it.
He despises the existence of Google, because the last thing a pathological bullshitter wants is fact-checking in pockets!
It's taken me nearly 40 years to understand that anchoring statements in reality is just a completely meaningless endeavor for him. He does not care what is true. He cares only what is convincing.
I've been wondering for about a year now why I feel like I can tell LLM work from human work so much more easily than most people, when the only "tell" I can put my finger on is, "The hair stands up on the back of my neck," but this explains ALL of it.
I feel like a good half of humanity operates this way, with it being far more prevalent in boomers than younger generations. It doesn't matter what is backed by evidence to them, instead they rely on anecdotes and persuasive quips and factoids. Having a friend who claims to have experienced X and listing off several other anecdotes means more to them than any amount of evidence.
The truly scary thing to me is watching them start to believe the anecdotes they've stolen from people and presented as their own stories actually did happen to them, as they lose their marbles.
I've spent much of my life learning to tell when people are making things up, but telling when they genuinely believe something that's completely wrong is a very different skill.
It's especially frustrating when they change the narrative of a real story about something where there were multiple witnesses (e.g., my mom and my siblings), then come to believe the narrative, and then accuse us of conspiring to gaslight them.
On the one hand, I get why that would be disorienting and scary, to have a whole group of people telling you you're wrong about something you're sure you remember. On the other hand...karma?
It depends on the AI. ChatGPT's higher models (o1-pro/o3/o4-mini-high) have some kind of limited capability to detect errors in the user's thinking, and have relatively few hallucinations.
Reminds me of the horrific state of student debate competitions today where the winning strategy is to incomprehensibly rattle off as many arguments as quickly as possible with strange breathy sounds in between
This is a consequence of the fact that any argument not responded to "flows across" the score sheet and is automatically a win for the team making the argument, no matter how silly. So a "natural" tendency would be to ignore ridiculous arguments like "not paying for school lunches will cause children to hyperventilate, and by the butterfly effect will lead to infinite hurricanes in developing nations causing a collapse of the global economy and intergalatic war and genocide". But if the opposite team fails to acknowledge the argument then that is the same as conceding it will happen.
Which is pretty ridiculous. The purpose of a debate should be to change/consolidate the hearts and minds of the audience to your side. To this end, it's usually sufficient to pick apart a few of the key points of your opponent's argument. Nitpicking every aspect of your opponent usually comes off as uncharismatic.
Brevity is really important in a debate. Especially in the modern day where someone might turn you into a chad vs soyjack meme.
And if anything, what happens before the debate is more important than what happens during it. Our dear president showed us you can become the leader of the free world using playground insults and ad-libbed speeches if you choose the right demographics and look good in a suit.
I guess winning like this cheapens the victory. Then again, this strategy continues to be used at all levels of disputes and politics. I wish there was a way to stop that, not just in student debates.
I'm accustomed to listening to regular speech at 2-3x speed, but apparently that's entirely different than listening to a human try to speak 2-3x faster than normal, because I could barely pick intelligible syllables out of that mess.
This is such an example of getting what you incentivize, not what matters.
"Because we raise the trigger and only two carrying noodles, and only two can announce in this network but their excess cites their examine this places where the apparatus of military power torches the ground"
Hamdiddle-eedah-hamdiddle-ah (do do do do dodododo expi-ali-do-cious)
What is the point of that? They're incomprehensible. (For those who haven't watched it: the video just shows people talking very fast, it doesn't explain why, kind of implies it's somehow good or impressive.)
The point is to win debate tournaments. In particular, it is (or at least was, when I competed in policy debate in high school and college in the 00s) strategically advantageous to maximize the number of distinct arguments, each with their own set of supporting evidence (usually read verbatim from a prepared excerpt of a news article or authoritative reference or whatever), you make within the allocated time. This incentivizes talking extremely quickly, which requires a fair bit of practice to become proficient at (and to understand).
It is quite strange. One would think a judge would easily throw this out.
I mean there is probably not a specific rule I could point to that a high school athlete couldn't ride a bike or a motorcycle in a 400m track run either.
There is probably not a specific rule that you can't shoot the shot put out of a canon either.
I would just assume the judges have the slightest bit of common sense.
That’s just like the larger discourse. The Gish gallop is standard practice.
Are there no rules in debates? There should be. You’re not allowed to punch someone in basketball so why should you be allowed to DOS people with bullshit in a debate?
Btw - my first author NeurIPS dataset and benchmarks paper is taking basically all the evidence that such debate community (American hs and college level policy and LD debate) produced over its recent history and making it easy for LLMs and people to consume it.
They’ve been quietly open sourcing all of their arguments for like 20+ years.
This dataset is so large and good entirely because of speed reading and the current state of debate tournament competitive dynamics. Spreading might be objectively absurd to listeners but the effects of it are literally good for society.
I asked an LLM and it said "A stack is a data structure that follows the Last In, First Out (LIFO) principle. This means that the last element added to the stack is the first element to be removed."
Yikes:( I am so worried about the damage that will be caused by the misuse of these tools. Already a lot of young folks will just mindlessly trust whatever the magic oracle spits out at them. We need to go back to testing people with pen and paper I suppose.
I read this and I see a common thinking fallacy, when someone is inclined to believe something a priori they fit the evidence to their a priori beliefs.
When playing 40 games against human players, CICERO achieved more than double the average score of the human players and ranked in the top 10% of participants who played more than one game
There are not a lot of facts to know when playing diplomacy. It's all about manipulating the other guy with words.
They also have more persuasive conversations in their pretraining data. That includes tons of marketing material, cons, and bullying. They are also as bold as you want them to be about imitating such tactics. They have no remorse or legal accountability either.
The gap between LLM and human cases was greater in the deceptive case. This may, of course, simply reflect the fact that random humans are bad at lying.
LLMs also never get tired of arguing. They'll respond to every point from a gish-gallop and provide quality-sounding replies to points that are obviously (to an informed person) flawed or seem (but aren't necessarily) mal-intentioned.
EDIT: LLMs also aren't egocentric; they'll respond in the other person's style (grammar, tone, and perhaps maintain their "subtext" like assumptions), and they're less likely to omit important information that would be implicit to them but not the other person.
Any qualities you ascribe to an LLM is part of its RLHF, ask to get irritated or lazy and it will simulate those qualities. They are high dimensional text simulators. They can and do simulate anything.
A clear case where LLMs exceed humans is in identifying solutions to disparate shallow constraints involving what would normally require very wide searches of more knowledge than any of us will ever have.
A simple case I have found, is looking for existing or creating new terms. If I have a series of concepts, which I have names for which have a nice linguistic pattern to emphasize their close relationship, except for one. I can describe the regularly named concepts, then ask for suggestions for the remaining concept.
The LLM pulls from virtually every topic with domain terminology, repurposable languages (Greek, Roman), words from fiction, all the way to creative construction of new words, tenses, etc to come up with great proposals in seconds.
I could imagine that crafting persuasive wording would be a similar challenge. Choosing the right words, right phrasing, etc. to carry as much positive connotation, implication of solidity, avoiding anything sounding challenging or controlling, etc. from all of human language and its huge space of emotional constraints and composites.
Very shallow but very wide reasoning/searching/balancing done in very little time.
And with an ability to avoid giving any unnecessary purchase for disagreement, being informed of all the myriad of typical and idiosyncratic ways people get hung up on failed persuasions. Whether in general or specific topic related.
LLM generated writing can be stereotypical.
But the more constraints put on requested material, the more their ability to construct really very original high quality, or even cleverly unique, prose in real time shines.
It's always amusing to watch people act shocked when LLMs beat average humans at persuasion. The actual headline here should be: 'A system trained on terabytes of successful human persuasion is better at persuasion than a random person on a crowdwork platform.' No mystery—just the mechanics of scale and exposure.
But guess what? Now, finally, we can co-opt LLMs for things humans fumble: e.g., real-time conversational tutoring, adaptive negotiation agents, or even scalable personal 'bullshit detectors' as countermeasures. I hope conversation doesn't go into AI-Safeteyism and restricting LLMs and more about building stuff. Let's build, not block.
As a marketer with a couple of decades of experience I can tell you that there is way more financial incentive in "slightly persuading [consumer] to tilt towards [product] in their next purchase, and spend more and earlier" than there ever will be towards "next-level unbiased tutor in anything".
The "super tutor" stuff that is always mentioned as the utopian outcome (along with "cures for cancer") is, unsurprisingly, never something being worked on by the person or lab quoting these examples.
I guess anything goes in B2B settings, but there is a valid reason to be cautious about these advances when it comes to mass-market consumer-facing applications.
I understand your perspective as a marketer, but I think you're creating a false dichotomy. Yes, persuasion tech has stronger financial incentives, but that doesn't prevent beneficial applications from emerging simultaneously.
The "super tutor" isn't some distant fantasy - millions already use ChatGPT, Claude and similar tools daily for personalized learning. They're imperfect but genuinely helpful for programming, languages, math, and countless other topics.
Look at what happened with YouTube: millions of people transformed themselves into programmers, musicians, mechanics, and countless other professions through free video tutorials. Khan Academy revolutionized math education. Coursera and edX brought university courses to anyone with internet. This wasn't utopian thinking - it was practical technology solving real educational problems at scale.
What's different now is that LLMs enable the missing piece: personalization. The one-on-one adaptive experience that was previously limited to those who could afford human tutors at $50-100/hour is now available to anyone at negligible marginal cost.
Your skepticism about cancer applications too ignores the technological trajectory we've been on for decades. Just as YouTube and online platforms democratized education, technology has been steadily dismantling bottlenecks in medical research.
The human genome project initially cost $3 billion and took 13 years. Today you can sequence a genome for under $1,000 in days. This wasn't utopian thinking; it was technological progress following its natural course.
Based on the data in table 3, I would attribute most of the difference to length of advice. LLMs average word count (29.4) is more than double human word count (13.25). Most other measures do not have a significant ratio. "Difficult word count" would be the only other with a ratio higher than 2, but that is inherited from total word count.
I think it would be difficult to truly convince me to answer differently in a test with 14 words where 30 would have enough space to actually convey an argument.
I would be very interested to see the test rerun while limiting LLM response length or encouraging long responses from humans.
If you think writing more words will be more persuasive, just... write more words?
The test already incentivises being persuasive! If writing more words would do that, and the incentivised human persuaders don't write more words and the LLMs do, then I think it's fair to say that LLMs are more persuasive than incentivised human persuaders.
Sure. I am not contesting that LLMs are more persuasive in this context. That basic result comes through very clearly in the paper. Its not as clear how relevant this is to other situations though. I think its quite likely that humans given the instruction to increase word count might outperform LLMs. People are very unlikely to have practiced the specific task of giving advice on multiple choice tests whereas LLMs have likely gotten RLHF training which likely helps in this situation.
I always try to pick out as many tidbits as possible from papers that might be applicable in other situations. I think the main difference of word count may be overshadowing other insights that may be more relevant to longer form argumentation.
> I would be very interested to see the test rerun while limiting LLM response length or encouraging long responses from humans.
I don’t know if that would have the effect you want. And if you’re more likely have hallucinations at lower word counts, that matters for those who are scrupulous, but many people trying to convince you of something believe the ends justify the means, and that honesty or correspondence to reality are not necessary, just nice to have.
I'm not sure what effect you think I want. The suggestion was just to increase the "interestingness" of the study. It seems to be like the main difference between LLM and human shown was length of response. Controlling for that variable and rerunning the experiment would help show other differences.
I do think its distinctly possible that LLMs will be much less convincing due to increased hallucinations at a low word count. I also think that may have less of an effect for dishonest suggestions. Simply stating a lie confidently is relatively effective.
I would prefer advising humans to increase length rather than restricting LLMs because of the cited effects.
Advising the opposite to humans does not make sense. 13 words is already tiny to convince someone. The choices I was thinking were restricting LLM word count and increasing human word count. The goal is specifically to make them more comparable.
The given study does not show any strength of humans over LLMs. Both goal metrics (truthful and deceptive) are better for LLMs than humans. If you are misinterpreting my advice as general advice for people not under the study's conditions, I would want to see the results of the proposed rerun before suggesting that.
However, if length of text is legitimately convincing regardless of content, I don't know why humans should avoid using that. If LLMs end up more convincing to humans than other humans simply because humans are too prideful to make their arguments longer, that seems like the worst possible future.
> If LLMs end up more convincing to humans than other humans simply because humans are too prideful to make their arguments longer, that seems like the worst possible future.
People aren’t too proud to make long arguments, they just take more time and effort to make for humans, and so historically, humans subconsciously consider longer arguments as more intellectually rigorous whether they are or not, and so length of a written piece is used as a kind of lazy heuristic corresponding with quality. When we're comparing the output of humans to that of other humans, this kind of approach may work to a certain extent, but AI/LLMs seem to be better at writing long pieces of text upon demand than humans. That humans find the LLM output more convincing if it is longer is not surprising to me, but I’ll agree with you that it isn’t a good sign either. The metric has become a target.
It matters if studies like this matter, that is, it matters to people who are interested in what has currently happened rather than what might happen in the future. 6 months of LLM progress keeps not looking like what I expected.
On the other hand, if you're content with your pre-existing predictions about what would happen, which I think is actually a reasonable position, there's no reason to read the paper.
That makes sense. Progress does seem very lumpy. Commercialization and secrets means that what we see publicly may not be state of the art. And I certainly never thought that Zuckerberg would be the one providing open model weights. It's a strange new world.
I'm hoping we see a flood of LLMs just like that Zurich piece, but at 10,000x scale. Perhaps even open source platforms to run your hobby LLM bot farm.
Social media has turned into cancer. It'd be riveting to watch it turn into bots talking to other bots. Social media wouldn't go away, but I get the feeling people will engage more with real life again.
As the platforms see less growth and fewer real users, we might even see a return to protocols and open standards instead of monolithic walled gardens.
It is CRITICAL that we be realistic about what fulfills the optimization objectives in the models that we train. I think there's been a significant unwillingness that objectives like "human preference" (RLHF, DPO, etc) not only help models become more accurate and sound more natural in speech, BUT ALSO optimize the models to be deceptive and convincing when they are wrong. It's easy to see, because you know what's more preferential than a lie? A lie that you don't know is a lie. You (may) prefer the truth, but if you cannot differentiate the truth from a lie you'll preference based on some other criteria. We all know that lies frequently win out here. If you doubt this, just turn on the news or talk to someone that belongs to the opposite political party of yourself.
This creates a very poorly designed tool! A good tool should fail as loudly as possible, in that it alerts the user of the failure and does its best to specify the conditions that led to this. This isn't always possible, but if you look at physical engineers you'll see that this is where they spend a significant portion of their time. Even in software I'd argue we do a lot here, but also that it is easy to brush off (we all love those compiler messages... right?). Clearly right now LLMs are in a state where we don't know how to make their failures more visible, and honestly, that is okay. But what is not okay is to pretend that this is not current reality and pretend that there are no dangers or consequences that this presents. We dismiss this because we catch some obvious errors and over-generalize the error quality, but that just means we suffer from Murray Gell-Mann Amnesia. It's REALLY hard to measure what you don't know. Importantly, we can't even begin to resolve these issues and build the tools we want (the ones we pretend these are!) if we ignore the reality of what we have. You cannot make things better if you are unwilling to recognize their limitations.
Everyone here is an engineer, researcher, or builder. This framework of thinking should be natural to us! We should also be able to understand that there's a huge difference between critiques and limitations and dismissing things. I'm an AI critic, but also very optimistic. I'm a researcher and spending my life working on this topic. It'd be insane to do such a thing if I thought it was a fruitless or evil effort. But it would be equally insane to pursue a topic with pure optimism. If I were to blind myself to limits and paint everything as a trivial to solve problem, I'd never be able to solve any of those problems. Ignoring or dismissing technical issues and limitations is the domain of the MBA managers, not engineers.
Sam Altman must be literally vibrating at the thoughts of tacking on ads at the end of a "persuasive" interaction about whatever. "... and remember to try new Oreo-flavored Pringles and tell them Gippity sent you with this 20% off code, because we are best friends and we can trust each other!"
We're fast approaching the point that any value someone can provide through a digital interface could be better done by a model. What do we use digital interfaces for? Practically everything.
Oh well not being a plumber, electrician, or farmer... but our society's current productivity, technology, automation reduced our need for 80% of the population needing to be farmers to now 1.3% in the US. Can you imagine what the equivalent of 1 billion digital engineers unlocks in understanding and implementing robotics?
Yes when the knowledge jobs are all done best by AI the rest will follow shortly. we will need to adapt to being "useless" as far as work goes and find other sources of worth. There's still a lot of people who want to compare it to Bitcoin hype around here, IMO the next few years everything is going to change way faster than than it ever has.
For the record I always thought Kurzweil and that crowd was clowns, now I think I was the wrong one
> IMO the next few years everything is going to change way faster
Honestly, after hearing this for the past 20 years (ever since ML and LLMs became a thing), it is actually more like the level-5 autonomous car hype and less like Bitcoin. Only that the driverless car hype never required such a humongous investment bubble, as does the Statistical-Text-Generator-as-AI one.
Meanwhile I haven't seen any real progress that I'd care about in a while.
Is GPT-4xyz better than the last one? I'm sure some benchmark numbers say that. But the number of applications where occasional hallucinations don't matter is very small, and where it matters nothing really changed. Companies are trying to use it for customer support but that predictably turned out to be a legal risk. Klarna went all-in on AI and regrets it now.
Some media are talking about Microsoft writing 30% of their new code with AI, but what Nadella actually said is less impressive: "maybe 20-30% of the code that is inside of our repos today in some of our projects are probably all written by software".
Which, coincidentally, is the ratio of code that can be autocompleted by an IDE without LLM, according to Jetbrains.
I have yet to see any evidence that anything will change way faster than it ever has, aside from the readiness of many younger people to use it in everyday life for things it really shouldn't be used.
Yes they have gotten better. If you give Gemini 2.5 the right context it seems to solve whatever. Drop in the folder + docs and it tends to be right about how to proceed now. I think people who don't find LLMs useful now aren't trying with the right context.
I’m with you. Weak version of a singularity seems likely. Recursive self improvement isn’t just possible, it’s inevitable. Models are capable of extrapolation, but they don’t even need it to do good interpolation which itself is enough to get us recursive self improvement.
I tend to think that it’ll have an optimistic ending. The key to solving most political problems is eliminating scarcity.