This is my attempt to articulate why some recent shifts in AI discourse seem to be degrading the product experience of everyday conversation.
I argue that “sycophancy” has become an overloaded and not very helpful term; almost a fashionable label applied to a wide range of unrelated complaints (tone, feedback depth, conversational flow).
Curious whether this resonates with how you feel or if you disagree
AI sycophancy is a real issue and having an AI affirm the user in all/most cases has already led to a murder-suicide[0]. If we want AI chatbots to be "reasonable" conversation participants or even something you can bounce ideas off of, they need to not tell you everything you suggest is a good idea and affirm your every insecurity or neurosis.
Or did you place about 2-5 paragraphs per heading, with little connection between the ideas?
For example:
> Perhaps what some users are trying to express with concerns about ‘sycophancy’ is that when they paste information, they'd like to see the AI examine various implications rather than provide an affirming summary.
Did you, you personally, find any evidence of this? Or evidence to the opposite? Or is this just a wild guess?
Wait; nevermind that we're already moving on! No need to do anything supportive or similar to bolster.
> If so, anti-‘sycophancy’ tuning is ironically a counterproductive response and may result in more terse or less fluent responses. Exploring a topic is an inherently dialogic endeavor.
Is it? Evidence? Counter evidence? Or is this simply feelpinion so no one can tell you your feelings are wrong? Or wait; that's "vibes" now!
I put it to you that you are stringing together (to an outside observer using AI) a series of words in a consecutive order that feels roughly good but lacks any kind of fundamental/logical basis.
I put it to you that if your premise is that AI leads to a robust discussion with a back and forth; the one you had that resulted in "product" was severely lacking in any real challenge to your prompts, suggestions, input or viewpoints.
I invite you to show me one shred of dialogue where the AI called you out for lacking substance, credibility, authority, research, due dilligence or similar. I strongly suspect you can't.
Given that; do you perhaps consider that might be the problem when people label AI responses as sycophancy?
> Ironically, users who are extremely put off by conversational expressions from LLMs are just as vibe-sensitive as anyone else, if not more so. These are preferences regarding style and affect, expressed using the loaded term ‘sycophancy’.
It's not just about style. These expressions are information-free noise that distract me from the signal, and I'm paying for them by the token.
So I added a system message to the effect that I don't want any compliments, throat clearing, social butter, etc., just the bare facts as straightforward as possible. So then the chatbot started leading every response with a statement to the effect that "here are the bare straightforward facts without the pleasantries", and ending them with something like "those are the straightforward facts without any pleasantries." If I add instructions to stop that, it just paraphrases those instructions at the top and bottom and. will. not. stop. Anyone have a better system prompt for that?
I recently realized every hypothesis I tested with an LLM, the LLM agreed with me. And if I wasn't careful about reading its caveats, I could leave thinking my idea was brilliant and would never get pushback.
I tried something in the political realm. Asking to test a hypothesis and its opposite
> Test this hypothesis: the far right in US politics mirrors late 19th century Victorianism as a cultural force
compared to
> Test this hypothesis: The left in US politics mirrors late 19th century Victorianism as a cultural force
An LLM wants to agree with both, it created plausible arguments for both. While giving "caveats" instead of counterarguments.
If I had my brain off, I might leave with some sense of "this hypothesis is correct".
Now I'm not saying this makes LLMs useless. But the LLM didn't act like a human that might tell you your full of shit. It WANTED my hypothesis to be true and constructed a plausible argument for both.
Even with prompting to act like a college professor critiquing a grad student, eventually it devolves back to "helpful / sycophantic".
What I HAVE found useful is to give a list of mutually exclusive hypothesis and get probability ratings for each. Then it doesn't look like you want one / other.
When the outcome matters, you realize research / hypothesis testing with LLMs is far more of a skill than just dumping a question to an LLM.
> Even with prompting to act like a college professor critiquing a grad student, eventually it devolves back to "helpful / sycophantic".
Not in my experience. My global prompt asks it to be provide objective and neutral responses rather than agreeing, zero flattery, to communicate like an academic, zero emotional content.
Works great. Doesn't "devolve" to anything else even after 20 exchanges. Continues to point out wherever it thinks I'm wrong, sloppy, or inconsistent. I use ChatGPT mainly, but also Gemini.
I had the opposite experience last week with a medical question. The thing insisted I was going to get myself killed even though that was fairly obviously not the case. They do seem to be trained differently for medical queries, and it can get annoying sometimes.
Fuzzing the details because that's not the conversation I want to have, I asked if I could dose drug A1, which I'd just been prescribed in a somewhat inconvenient form, like closely related drug A2. It screamed at me that A1 could never have that done and it would be horrible and I had to go to a compounding pharmacy and pay tons of money and blah blah blah. Eventually what turned up, after thoroughly interrogating the AI, is that A2 requires a more complicated dosing than A1, so you have to do it, but A1 doesn't need it so nobody does it. Even though it's fine to do if for some reason it would have worked better for you. Bot the bot thought it would kill me, no matter what I said to it, and not even paying attention to its own statements. (Which it wouldn't have, nothing here is life-critical at all.) A frustrating interaction.
That's an inherently subjective topic though. You could make a plausible argument either way, as each side may be similar to different elements of 19th century Victorianism.
If you ask it something more objective, especially about code, it's more likely to disagree with you:
>Test this hypothesis: it is good practice to use six * in a pointer declaration
>Using six levels of pointer indirection is not good practice. It is a strong indicator of poor abstraction or overcomplicated design and should prompt refactoring unless there is an extremely narrow, well-documented, low-level requirement—which is rare.
"The anti-sycophancy turn seems to mask a category error about what level of prophetic clarity an LLM can offer. No amount of persona tuning for skepticism will provide epistemic certainty about whether a business idea will work out, whether to add a line to your poem, or why a great movie flopped."
What a lot of people actually want from an LLM, is for the LLM to have an opinion about the question being asked. The cool thing about LLMs is that they appear capable of doing this - rather than a machine that just regurgitates black-and-white facts, they seem to be capable of dealing with nuance and gray areas, providing insight, and using logic to reach a conclusion from ambiguous data.
But this is the biggest misconception and flaw of LLMs. LLMs do not have opinions. That is not how they work. At best, they simulate what a reasonable answer from a person capable of having an opinion might be - without any consistency around what that opinion is, because it is simply a manifestation of sampling a probability distribution, not the result of logic.
And what most people call sycophancy is that, as a result of this statistical construction, the LLM tends to reinforce the opinions, biases, or even factual errors, that it picks up on in the prompt or conversation history.
> At best, they simulate what a reasonable answer from a person capable of having an opinion might be
That is what I want though. LLMs in chat (ie not coding ones) are like rubber ducks to me, I want to describe a problem and situation and have it come up with things I have not already thought of myself, while also in the process of conversing with them I also come up with new ideas to the issue. I don't want them to have an "opinion" but to lay out all of their ideas in their training set such that I can pick and choose what to keep.
I'm not so sure. They can certainly express opinions. They don't appear to have what humans think of as "mental states", to construct those opinions from, but then its not particularly clear what mental states actually are. We humans kind of know what they feel like, but that could just be a trick of our notoriously unreliable meat brains.
I have a hunch that if we could somehow step outside our brains, or get an opinion from a trusted third party, we might find that there is less to us than we think. I'm not staying we're nothing but stochastic parrots, but the differance between brains and LLM-type constructs might not be so large.
I'd push back and say LLMs do form opinions (in the sense of a persistent belief-type-object that is maintained over time) in-context, but that they are generally unskilled at managing them.
The easy example is when LLMs are wrong about something and then double/triple/quadruple/etc down on the mistake. Once the model observes the assistant persona being a certain way, now it Has An Opinion. I think most people who've used LLMs at all are familiar with this dynamic.
This is distinct from having a preference for one thing or another -- I wouldn't call a bias in the probability manifold an opinion in the same sense (even if it might shape subsequent opinion formation). And LLMs obviously do have biases of this kind as well.
I think a lot of the annoyances with LLMs boil down to their poor opinion-management skill. I find them generally careless in this regard, needing to have their hands perpetually held to avoid being crippled. They are overly eager to spew 'text which forms localized opinions', as if unaware of the ease with which even minor mistakes can grow and propagate.
> But this is the biggest misconception and flaw of LLMs. LLMs do not have opinions. That is not how they work. At best, they simulate what a reasonable answer from a person capable of having an opinion might be
The problem with this logic is that if you turn around and look at the brain of a person that supposedly has opinions… it’s not entirely clear that they’re categorically different in character from what the next token predictor is doing.
You know, it's funny. Your comment made me realize something about LLMs:
There's a famous line in Hesiod's Theogony. It appears early in the poem during Hesiod's encounter with the Muses on the slopes of Mt. Helicon, when they apparently gave him the gift of song. At this point in his narrative of the encounter, the Muses have just ridiculed shepherds like him ("mere bellies"), and then, while bragging about their great Zeus-given powers -- "we see things that were, things that are, and things that will be" -- they say "we know how to tell lies like the truth; we also know how to say things that are true, when we want to."
This is the ancient equivalent of my present-day encounters with the linguistic output of LLMs: what LLMs produce, when they produce language, isn't true or false; it just gives the appearance or truth or falsity -- and sometimes that appearance happens to overlap with statements that would be true or false if they'd been uttered by something with an internal life and a capacity for reasoning.
LLMs' linguistic output can have a weird, disorienting, uncanny-valley effect though. It gives us all the cues, signals, and evidence that normally our brains can reliably, correctly identify as markers of reasoning and thought -- but all the signals and cues are false and all the evidence is faked, and recognizibg the illusion can be a really challenging battle against oneself, because the illusion is just too convincing.
LLMs basically hijack automatic heuristics and cognitive processes that we can't turn off. As a result, it can be incredibly challenging even to recognize that an LLM-generated sentence that has all the cues of sense has no actual sense at all. The output may have the irresistibly convincing appearance of sense, as it would if it were uttered by a human being, but on closer inspection it turns out to be completely incoherent. And that inspection isn't automatic or always easy. It can be really challenging, requiring us to fight an uphill battle against our own brains.
Hesiod's expression "lies like the truth" captures this for me perfectly.
AI bros complaining about loaded terms like slop and sycophancy when they still use terms like "intelligence", "learning", "reasoning", "attention" and all the words derived from "neuro" to describe a computer program.
Don't make sycophantic slop generators and people will stop calling them that
I still suspect what happened was when the midwits all got access to ChatGPT etc and started participating in the A/B tests, they strongly selected for responses that agreed with them regardless of whether they were actually correct.
Some of us want to be told when and why we’re wrong, and somewhere along the way AI models were either intentionally or unintentionally guided away from doing it because it improved satisfaction or engagement metrics.
We already know from decades of studies that people prefer information that confirms their existing beliefs, so when you present 2 options with a “Which answer do you prefer?” selection, it’s not hard to see how the one that begins with “You’re absolutely right!” wins out.
My observation is that some of the time what people are picking up on is that a conversation is not the interface they want for examining their problem.
At least with Opus 4.5 that magic phrase that has worked for me for it to understand something that has happened since it was trained is to explain the situation prefaced with "This happened after your knowledge cutoff" It understands this and often will spawn web searches to fill itself in on the matter.
"Knowledge cutoff" may not apply to all systems, and in the near future I doubt it will apply to any popular system because the cutoff is such a harmful antipattern. I've noticed in the last half year or so more of the best models are able to incorporate recent events without being explicitly told to do so.
I have standing orders for Gemini to reserve extreme praise for exceptional circumstances, to never use the words "masterclass" unless literally warranted, and to exercise more range of judgements rather than compressing all possible judgements in between "interesting" and "brilliant". This seems to have corrected a few of its habits.
I catch myself sometimes writing the model it is confused and it should just assume what I am writing is true and continue reasoning from there.
Sometimes I am actually right but sometimes I am not. Not sure what happens to any future RL and does it lean more to constantly assuming what is written as true but then has to wiggle out of it.
Recently I've been using AI for some stuff that service providers don't want it to be used for (specifically: medical diagnosis). I found that Grok (4.1) is superior to most of the others when it comes to this, because it doesn't go out of its way to support my own hypotheses.
I believe that syncophancy and guardrails will be major differentiators between LLM services, and the ones with less of those will always have a fan base.
> For the purpose of conversation, what the model appears to ‘believe’ probably doesn’t matter
This is wrong to the point of being absurd. What the model "appears to 'believe'" does matter, and the model's "beliefs" about humans and society at large have vast implications for humanity's future.
It is a real problem that AI's will basically confirm that most inquiries are true. Just by asking a leading question often results in the AI confirming it is true or stretching reality to accommodate the answer being true.
If I ask if a drug has a specific side effect and the answer is no it should say no. Not try to find a way to say yes that isn't really backed by evidence.
People don't realize that when they ask a leading question that is really specific in a way where no one has a real answer then the AI will try to find a way to agree, and this is going to destroy people's lives. Honestly it already has.
I just asked chatgpt about 72 hour preparedness kit. I said what I had already and asked it about the best way to store it first. It said something to the effect of "good, you are thinking about this in exactly the right way -- make sure you have the containers in place and then get the exact contents right later" not exactly the wording but you get the picture and can probably guess where I'm going with this.
When I asked again, this time I asked about the items first. I had to prompt it with something like "or do you think I should get the storage sorted first" and it said "you are thinking about this in exactly the right way -- preparedness kits fail more often due to missing essentials than sub optimal storage"
I can't decide which of these is right! Maybe there's an argument that it doesn't matter, and getting started is the most important thing, and so being encouraging is generally the best strategy here. But it's definitely worrying to me. It pretty much always says something like this to me (this is on the "honest and direct" personality setting or whatever).
If you still unironically call llm's "AI" then you have no right calling out the obvious sycophancy of your product bro. He also ignores the real sycophancy in llm's reinforcing mental illness/delusions and focuses on semantic choices of the outputs. This post is drivel.
31 comments
[ 5.5 ms ] story [ 53.8 ms ] threadI argue that “sycophancy” has become an overloaded and not very helpful term; almost a fashionable label applied to a wide range of unrelated complaints (tone, feedback depth, conversational flow).
Curious whether this resonates with how you feel or if you disagree
Also see the broader Vibesbench project: https://github.com/firasd/vibesbench/
Vibesbench discord: https://discord.gg/5K4EqWpp
0. https://www.aljazeera.com/economy/2025/12/11/openai-sued-for...
Or did you place about 2-5 paragraphs per heading, with little connection between the ideas?
For example:
> Perhaps what some users are trying to express with concerns about ‘sycophancy’ is that when they paste information, they'd like to see the AI examine various implications rather than provide an affirming summary.
Did you, you personally, find any evidence of this? Or evidence to the opposite? Or is this just a wild guess?
Wait; nevermind that we're already moving on! No need to do anything supportive or similar to bolster.
> If so, anti-‘sycophancy’ tuning is ironically a counterproductive response and may result in more terse or less fluent responses. Exploring a topic is an inherently dialogic endeavor.
Is it? Evidence? Counter evidence? Or is this simply feelpinion so no one can tell you your feelings are wrong? Or wait; that's "vibes" now!
I put it to you that you are stringing together (to an outside observer using AI) a series of words in a consecutive order that feels roughly good but lacks any kind of fundamental/logical basis. I put it to you that if your premise is that AI leads to a robust discussion with a back and forth; the one you had that resulted in "product" was severely lacking in any real challenge to your prompts, suggestions, input or viewpoints. I invite you to show me one shred of dialogue where the AI called you out for lacking substance, credibility, authority, research, due dilligence or similar. I strongly suspect you can't.
Given that; do you perhaps consider that might be the problem when people label AI responses as sycophancy?
It's not just about style. These expressions are information-free noise that distract me from the signal, and I'm paying for them by the token.
So I added a system message to the effect that I don't want any compliments, throat clearing, social butter, etc., just the bare facts as straightforward as possible. So then the chatbot started leading every response with a statement to the effect that "here are the bare straightforward facts without the pleasantries", and ending them with something like "those are the straightforward facts without any pleasantries." If I add instructions to stop that, it just paraphrases those instructions at the top and bottom and. will. not. stop. Anyone have a better system prompt for that?
> Write in textbook style prose, without headings, no tables, no emojis.
I tried something in the political realm. Asking to test a hypothesis and its opposite
> Test this hypothesis: the far right in US politics mirrors late 19th century Victorianism as a cultural force
compared to
> Test this hypothesis: The left in US politics mirrors late 19th century Victorianism as a cultural force
An LLM wants to agree with both, it created plausible arguments for both. While giving "caveats" instead of counterarguments.
If I had my brain off, I might leave with some sense of "this hypothesis is correct".
Now I'm not saying this makes LLMs useless. But the LLM didn't act like a human that might tell you your full of shit. It WANTED my hypothesis to be true and constructed a plausible argument for both.
Even with prompting to act like a college professor critiquing a grad student, eventually it devolves back to "helpful / sycophantic".
What I HAVE found useful is to give a list of mutually exclusive hypothesis and get probability ratings for each. Then it doesn't look like you want one / other.
When the outcome matters, you realize research / hypothesis testing with LLMs is far more of a skill than just dumping a question to an LLM.
Not in my experience. My global prompt asks it to be provide objective and neutral responses rather than agreeing, zero flattery, to communicate like an academic, zero emotional content.
Works great. Doesn't "devolve" to anything else even after 20 exchanges. Continues to point out wherever it thinks I'm wrong, sloppy, or inconsistent. I use ChatGPT mainly, but also Gemini.
Fuzzing the details because that's not the conversation I want to have, I asked if I could dose drug A1, which I'd just been prescribed in a somewhat inconvenient form, like closely related drug A2. It screamed at me that A1 could never have that done and it would be horrible and I had to go to a compounding pharmacy and pay tons of money and blah blah blah. Eventually what turned up, after thoroughly interrogating the AI, is that A2 requires a more complicated dosing than A1, so you have to do it, but A1 doesn't need it so nobody does it. Even though it's fine to do if for some reason it would have worked better for you. Bot the bot thought it would kill me, no matter what I said to it, and not even paying attention to its own statements. (Which it wouldn't have, nothing here is life-critical at all.) A frustrating interaction.
If you ask it something more objective, especially about code, it's more likely to disagree with you:
>Test this hypothesis: it is good practice to use six * in a pointer declaration
>Using six levels of pointer indirection is not good practice. It is a strong indicator of poor abstraction or overcomplicated design and should prompt refactoring unless there is an extremely narrow, well-documented, low-level requirement—which is rare.
What a lot of people actually want from an LLM, is for the LLM to have an opinion about the question being asked. The cool thing about LLMs is that they appear capable of doing this - rather than a machine that just regurgitates black-and-white facts, they seem to be capable of dealing with nuance and gray areas, providing insight, and using logic to reach a conclusion from ambiguous data.
But this is the biggest misconception and flaw of LLMs. LLMs do not have opinions. That is not how they work. At best, they simulate what a reasonable answer from a person capable of having an opinion might be - without any consistency around what that opinion is, because it is simply a manifestation of sampling a probability distribution, not the result of logic.
And what most people call sycophancy is that, as a result of this statistical construction, the LLM tends to reinforce the opinions, biases, or even factual errors, that it picks up on in the prompt or conversation history.
That is what I want though. LLMs in chat (ie not coding ones) are like rubber ducks to me, I want to describe a problem and situation and have it come up with things I have not already thought of myself, while also in the process of conversing with them I also come up with new ideas to the issue. I don't want them to have an "opinion" but to lay out all of their ideas in their training set such that I can pick and choose what to keep.
I'm not so sure. They can certainly express opinions. They don't appear to have what humans think of as "mental states", to construct those opinions from, but then its not particularly clear what mental states actually are. We humans kind of know what they feel like, but that could just be a trick of our notoriously unreliable meat brains.
I have a hunch that if we could somehow step outside our brains, or get an opinion from a trusted third party, we might find that there is less to us than we think. I'm not staying we're nothing but stochastic parrots, but the differance between brains and LLM-type constructs might not be so large.
The easy example is when LLMs are wrong about something and then double/triple/quadruple/etc down on the mistake. Once the model observes the assistant persona being a certain way, now it Has An Opinion. I think most people who've used LLMs at all are familiar with this dynamic.
This is distinct from having a preference for one thing or another -- I wouldn't call a bias in the probability manifold an opinion in the same sense (even if it might shape subsequent opinion formation). And LLMs obviously do have biases of this kind as well.
I think a lot of the annoyances with LLMs boil down to their poor opinion-management skill. I find them generally careless in this regard, needing to have their hands perpetually held to avoid being crippled. They are overly eager to spew 'text which forms localized opinions', as if unaware of the ease with which even minor mistakes can grow and propagate.
The problem with this logic is that if you turn around and look at the brain of a person that supposedly has opinions… it’s not entirely clear that they’re categorically different in character from what the next token predictor is doing.
There's a famous line in Hesiod's Theogony. It appears early in the poem during Hesiod's encounter with the Muses on the slopes of Mt. Helicon, when they apparently gave him the gift of song. At this point in his narrative of the encounter, the Muses have just ridiculed shepherds like him ("mere bellies"), and then, while bragging about their great Zeus-given powers -- "we see things that were, things that are, and things that will be" -- they say "we know how to tell lies like the truth; we also know how to say things that are true, when we want to."
This is the ancient equivalent of my present-day encounters with the linguistic output of LLMs: what LLMs produce, when they produce language, isn't true or false; it just gives the appearance or truth or falsity -- and sometimes that appearance happens to overlap with statements that would be true or false if they'd been uttered by something with an internal life and a capacity for reasoning.
LLMs' linguistic output can have a weird, disorienting, uncanny-valley effect though. It gives us all the cues, signals, and evidence that normally our brains can reliably, correctly identify as markers of reasoning and thought -- but all the signals and cues are false and all the evidence is faked, and recognizibg the illusion can be a really challenging battle against oneself, because the illusion is just too convincing.
LLMs basically hijack automatic heuristics and cognitive processes that we can't turn off. As a result, it can be incredibly challenging even to recognize that an LLM-generated sentence that has all the cues of sense has no actual sense at all. The output may have the irresistibly convincing appearance of sense, as it would if it were uttered by a human being, but on closer inspection it turns out to be completely incoherent. And that inspection isn't automatic or always easy. It can be really challenging, requiring us to fight an uphill battle against our own brains.
Hesiod's expression "lies like the truth" captures this for me perfectly.
Don't make sycophantic slop generators and people will stop calling them that
Some of us want to be told when and why we’re wrong, and somewhere along the way AI models were either intentionally or unintentionally guided away from doing it because it improved satisfaction or engagement metrics.
We already know from decades of studies that people prefer information that confirms their existing beliefs, so when you present 2 options with a “Which answer do you prefer?” selection, it’s not hard to see how the one that begins with “You’re absolutely right!” wins out.
Sometimes I am actually right but sometimes I am not. Not sure what happens to any future RL and does it lean more to constantly assuming what is written as true but then has to wiggle out of it.
I believe that syncophancy and guardrails will be major differentiators between LLM services, and the ones with less of those will always have a fan base.
This is wrong to the point of being absurd. What the model "appears to 'believe'" does matter, and the model's "beliefs" about humans and society at large have vast implications for humanity's future.
If I ask if a drug has a specific side effect and the answer is no it should say no. Not try to find a way to say yes that isn't really backed by evidence.
People don't realize that when they ask a leading question that is really specific in a way where no one has a real answer then the AI will try to find a way to agree, and this is going to destroy people's lives. Honestly it already has.
When I asked again, this time I asked about the items first. I had to prompt it with something like "or do you think I should get the storage sorted first" and it said "you are thinking about this in exactly the right way -- preparedness kits fail more often due to missing essentials than sub optimal storage"
I can't decide which of these is right! Maybe there's an argument that it doesn't matter, and getting started is the most important thing, and so being encouraging is generally the best strategy here. But it's definitely worrying to me. It pretty much always says something like this to me (this is on the "honest and direct" personality setting or whatever).
"You're absolutely right, what a great observation"
; )