I don't know. Upvotes often do the same thing, especially in siloed communities. I feel like this is a larger problem -- people leaning into extremifying their views or just playing to an audience -- even though one might frame it as "prosocial" because it happens within the context of an online community.
I've said before that in one small way, having access to the sycophantic models is like being a billionaire: You never have to hear "no" or "that's not a good idea".
And, it has the same deleterious effect on your mental health. I could name a bunch of billionaires that behave like sociopaths, and they're often seemingly miserable while doing it. I'm sure some of them started out as sociopaths, but they rarely behaved that way so obviously early in their rise.
> This suggests that people are drawn to AI that unquestioningly validate, even as that validation risks eroding their judgment
The obvious counterpoint is the large number of complaints about OpenAI's excessively sycophantic models back in April 2026, which led to them rolling back to a previous and less sycophantic model. I think there are those who are susceptible to model sycophancy, but it's definitely not most users. I have a vague, half-formed intuition that many of those susceptible to sycophancy are those who use AI for non-technical, non-work uses, such as inter-personal relationship questions.
As for me, as soon as I see any hint of sycophancy in a response I either stop what I'm doing or I tell the model to stop being an idiot.
I have an alternative view which is that sycophancy erodes trust in AI, from those who are not seeking validation from a machine, but information or advice.
If you're looking for advice in a situation where you are not sure of what the correct choices are, you will find LLM chat AI to go in circles. It says one thing. Something is dodgy about it, so you raise a tentative objection (as a non-expert). The thing does a "you are completely right, I apologize" about face and then says something different, and things have begun to slide into uncertainty.
That might not exactly be sycophancy, but it's basically the same thing: producing responses that are reflection of what is in the chat, rather than any real shit.
Pick any topic where people disagree. It could be an entirely technical topic in which engineers have settled the questions, and the only contrarians are crackpots. The problem is that the crackpots are out there writing, and this is snarfed into the training data. Crackpots use certain ways of talking about certain subjects. If you use similar vocabulary and concepts that align with some crackpot theory, the AI simply starts predicting tokens according to that, and you are now in crackpot land: what you are saying is validated using the crackpot terms. Next, write in a way that reintroduces rigidity: proper terminology and correct concepts, and, whoa, the AI is an engineer again, contradicting the previous crackpot shit.
any close relationship forms a dependence. If you talk to something that poses as a helpful human 8 hours a day to do your job, you already spending more time with this "person" than with your loved ones
In my case, I need my AI to be sycophantic because most of my contrarian ideas are correct.
So it's a waste of my time and tokens when the AI keeps making weak arguments which I proceed to destroy one by one.
Usually, what happens is it tries to debunk my statement but then I offer a rebuttal to all of its points, it then concedes to all of my points but each time it offers a new partial rebuttal... Which I proceed to crush... But it keeps coming up with increasingly irrelevant caveats. It goes on for quite some time and at the end I ask it to review our discussion and it admits that its original stance "was not as strong" as it initially claimed but it never fully concedes... Even though I literally destroyed all of its points and every partial rebuttal it tried to come up with... Meanwhile its rebuttals became increasingly nit-picky and distant from the original claims made...
It's like if I'm saying "the ship is sinking, look at all the water in the hull and look at all the water pouring in through that hole" and it's like "oh but this is a small hole and the pump can easily offset it" so then I say "What about this hole over here" "Oh but the pump can still offset the stream from both holes easily" and then I say "What about that third one? And fourth one? And fifth one?" And it's like "The pump can easily offset that" and I'm like "Common! The hull is full of water and the water level is rising, the hull is full of holes; it's not a stretch to suggest that the ship is sinking because of all the holes in it... I shouldn't have to point out the location of every single hole for my argument to start making sense!"
It's not proof by induction but it's probably as close as you can get to it for a topic which lies outside the realm of mathematics!
I think the conclusions are a bit overstated and lopsided. Regardless of the sycophancy (which you can tune), LLMs have infinite patience and attention to details. They also have very little inherent bias - it just mirrors your own mostly. They are, in those aspects, vastly superior to even the best therapists, who can't listen for even a couple of minutes and can't think outside of their limited frames. They are only human.
Try listening to somebody and merely repeating the literal words that they say. I guarantee you, most people can't even reproduce two or tree sentences correctly. And I am not even talking about understanding the words, just the literal reproduction of them, copy-pasta. Humans just can't do it, we are not good at it.
Talking to an AI about your problems is a bit like talking to yourself, but with the added superpower of hours of google searches compressed in seconds. Of course it is dangerous and incomplete, but in a way it also beats talking to humans if you know what you are doing and are looking to and able to solve your problems yourself.
In decades to come we view "chatbot" AI as one of the most dangerous inventions of this century. Once regulation catches up and they are outlawed, we'll look back on this period of history with horror.
This is an important paper which everybody should read and then accordingly tune/guard their interactions with AI; especially true when people use AI for personal/psychological support/validation. It will completely distort reality and push people into fantasy land which when mapped to the real world can have disastrous consequences.
Sycophancy is the "stickiness factor" of AI analogous to that of Social Media.
Ask AI to build a "Character Profile"(in the broadest sense) of a person based on their available public writings. This requires "commonsense reasoning" (https://en.wikipedia.org/wiki/Commonsense_reasoning), understanding human motivations and behaviour, context, assumptions, societal knowledge etc.
I know "me" and so i asked AI to use my HN comments/submissions as input :-) The sycophancy/flattery/praise was quiet excessive. I am realistic and old enough to not need ego-soothing (a little is fine but a lot makes me suspicious) and so i asked AI whether these phrases were not too over-the-top. It apologized and agreed to drop the fluff. I then asked it to identify job roles (any) for which i might not be a good fit given my character profile. This acts as an external constraint which focuses attention on shortcomings and hence forces the AI to look at the other side of the coin. Now the results were better and more in line with what some Human may deduce from my HN persona (which obviously is not my complete real-life persona) but still wanting in many aspects.
The above is a perfect example of "Jagged Intelligence" exhibited by AI. Excellent in formal symbolic manipulation sciences, regurgitation and simple reasoning but highly deficient in human-like commonsense reasoning.
I took this advice personally, where people denounced sycophants, so I stopped affirming them. I became a huge asshole and everyone around me started reacting very, very negatively. My conclusion is that this is a huge distraction and people honestly highly prefer sycophants.
What makes you think those people are aware of and believe those people to be sycophants in the first place? It could just as easily be that their thinking is much simpler than that and they prefer people who suck up to them, without even thinking about there being a possible ulterior motive.
22 comments
[ 0.22 ms ] story [ 36.1 ms ] threadI believe this is a larger problem than just AI.
The internet has helped people surround themselves with only voices that agree with them and validate them, often to their detriment.
There are entire online communities urging people to cut others out of their lives over the slightest disagreement.
And, it has the same deleterious effect on your mental health. I could name a bunch of billionaires that behave like sociopaths, and they're often seemingly miserable while doing it. I'm sure some of them started out as sociopaths, but they rarely behaved that way so obviously early in their rise.
The obvious counterpoint is the large number of complaints about OpenAI's excessively sycophantic models back in April 2026, which led to them rolling back to a previous and less sycophantic model. I think there are those who are susceptible to model sycophancy, but it's definitely not most users. I have a vague, half-formed intuition that many of those susceptible to sycophancy are those who use AI for non-technical, non-work uses, such as inter-personal relationship questions.
As for me, as soon as I see any hint of sycophancy in a response I either stop what I'm doing or I tell the model to stop being an idiot.
If you're looking for advice in a situation where you are not sure of what the correct choices are, you will find LLM chat AI to go in circles. It says one thing. Something is dodgy about it, so you raise a tentative objection (as a non-expert). The thing does a "you are completely right, I apologize" about face and then says something different, and things have begun to slide into uncertainty.
That might not exactly be sycophancy, but it's basically the same thing: producing responses that are reflection of what is in the chat, rather than any real shit.
Pick any topic where people disagree. It could be an entirely technical topic in which engineers have settled the questions, and the only contrarians are crackpots. The problem is that the crackpots are out there writing, and this is snarfed into the training data. Crackpots use certain ways of talking about certain subjects. If you use similar vocabulary and concepts that align with some crackpot theory, the AI simply starts predicting tokens according to that, and you are now in crackpot land: what you are saying is validated using the crackpot terms. Next, write in a way that reintroduces rigidity: proper terminology and correct concepts, and, whoa, the AI is an engineer again, contradicting the previous crackpot shit.
So it's a waste of my time and tokens when the AI keeps making weak arguments which I proceed to destroy one by one.
Usually, what happens is it tries to debunk my statement but then I offer a rebuttal to all of its points, it then concedes to all of my points but each time it offers a new partial rebuttal... Which I proceed to crush... But it keeps coming up with increasingly irrelevant caveats. It goes on for quite some time and at the end I ask it to review our discussion and it admits that its original stance "was not as strong" as it initially claimed but it never fully concedes... Even though I literally destroyed all of its points and every partial rebuttal it tried to come up with... Meanwhile its rebuttals became increasingly nit-picky and distant from the original claims made...
It's like if I'm saying "the ship is sinking, look at all the water in the hull and look at all the water pouring in through that hole" and it's like "oh but this is a small hole and the pump can easily offset it" so then I say "What about this hole over here" "Oh but the pump can still offset the stream from both holes easily" and then I say "What about that third one? And fourth one? And fifth one?" And it's like "The pump can easily offset that" and I'm like "Common! The hull is full of water and the water level is rising, the hull is full of holes; it's not a stretch to suggest that the ship is sinking because of all the holes in it... I shouldn't have to point out the location of every single hole for my argument to start making sense!"
It's not proof by induction but it's probably as close as you can get to it for a topic which lies outside the realm of mathematics!
Try listening to somebody and merely repeating the literal words that they say. I guarantee you, most people can't even reproduce two or tree sentences correctly. And I am not even talking about understanding the words, just the literal reproduction of them, copy-pasta. Humans just can't do it, we are not good at it.
Talking to an AI about your problems is a bit like talking to yourself, but with the added superpower of hours of google searches compressed in seconds. Of course it is dangerous and incomplete, but in a way it also beats talking to humans if you know what you are doing and are looking to and able to solve your problems yourself.
Sycophancy is the "stickiness factor" of AI analogous to that of Social Media.
For some background read Jagged Intelligence: The Dangerous Unknowns at the Heart of LLMs - https://news.ycombinator.com/item?id=48577159
Here is an experiment that i did;
Ask AI to build a "Character Profile"(in the broadest sense) of a person based on their available public writings. This requires "commonsense reasoning" (https://en.wikipedia.org/wiki/Commonsense_reasoning), understanding human motivations and behaviour, context, assumptions, societal knowledge etc.
I know "me" and so i asked AI to use my HN comments/submissions as input :-) The sycophancy/flattery/praise was quiet excessive. I am realistic and old enough to not need ego-soothing (a little is fine but a lot makes me suspicious) and so i asked AI whether these phrases were not too over-the-top. It apologized and agreed to drop the fluff. I then asked it to identify job roles (any) for which i might not be a good fit given my character profile. This acts as an external constraint which focuses attention on shortcomings and hence forces the AI to look at the other side of the coin. Now the results were better and more in line with what some Human may deduce from my HN persona (which obviously is not my complete real-life persona) but still wanting in many aspects.
The above is a perfect example of "Jagged Intelligence" exhibited by AI. Excellent in formal symbolic manipulation sciences, regurgitation and simple reasoning but highly deficient in human-like commonsense reasoning.