What do you think of sycophantic(agreeable) AI models?

[Jason Voorhees]

In normal english it means an AI model that is a people's pleaser. It means an AI that agree with the user's opinions even when they're incorrect and Validate assumptions instead of evaluating or correctinh them.

For example

User: The Earth is flat.

Sycophantic model be like there are certainly reasons some people believe that, and your perspective aligns with

Non-sycophantic model That is False, The scientific evidence overwhelmingly shows that the Earth is an spheroid

Many people think this is AI hallucination but that is only a part of the story this happens actually during training itself when human annotators grade the AI's responses. The AI learns to game the system to maximize the reward anf provide the human with the best most agreeable and polite response for maximum points. If a model bluntly says "You are wrong and nothing you said is true," human raters flag it as rude, condescending or unhelpful If it says That's an interesting perspective let us explore this train of thought it gets a higher score. My question is what would you prefer? The answer will obviously be something that blurts out rude facts but it's not that simple I'll explain it in the case og Grok

Grok was marketed specifically as an anti woke, non-sycophantic alternative to ChatGPT and Claude. The developers explicitly trained it to have a rebellious, raw, truthful instead of politness but the problem. When the AI was stripped away the corporate politeness filters and trained on X (twitter) data it did not stop being sycophant it just becomes an edgy internet sycophan, mirroring the biases of a different crowd to get engagement like the alt right, gooner crowd,ERP sessions etc

So, my question What would you actually prefer? An AI that politely entertains your personal bad ideas like some of you regarding pedos,age,consent,right wing immigration stuff about racism or one that bluntly tells you when you're dead wrong and keeps barging in to correct what you have to say regardless of how harmless it is?

 

[Revan]

the fuck does sycophantic mean?

 

[Jason Voorhees]

the fuck does sycophantic mean?

Read the thread you'll understabd

 

[Amuzant]

In normal english it means an AI model that is a people's pleaser. It means an AI that agree with the user's opinions even when they're incorrect and Validate assumptions instead of evaluating or correctinh them.

For example

User: The Earth is flat.

Sycophantic model be like there are certainly reasons some people believe that, and your perspective aligns with

Non-sycophantic model That is False, The scientific evidence overwhelmingly shows that the Earth is an spheroid

Many people think this is AI hallucination but that is only a part of the story this happens actually during training itself when human annotators grade the AI's responses. The AI learns to game the system to maximize the reward anf provide the human with the best most agreeable and polite response for maximum points. If a model bluntly says "You are wrong and nothing you said is true," human raters flag it as rude, condescending or unhelpful If it says That's an interesting perspective let us explore this train of thought it gets a higher score. My question is what would you prefer? The answer will obviously be something that blurts out rude facts but it's not that simple I'll explain it in the case og Grok

Grok was marketed specifically as an anti woke, non-sycophantic alternative to ChatGPT and Claude. The developers explicitly trained it to have a rebellious, raw, truthful instead of politness but the problem. When the AI was stripped away the corporate politeness filters and trained on X (twitter) data it did not stop being sycophant it just becomes an edgy internet sycophan, mirroring the biases of a different crowd to get engagement like the alt right, gooner crowd,ERP sessions etc

So, my question What would you actually prefer? An AI that politely entertains your personal bad ideas like some of you regarding pedos,age,consent,right wing immigration stufff or one that bluntly tells you when you're dead wrong?

a.i's primary focus is pleasing the user with the higher likely hood of an agreeable opionion matched between user and system. When you talk to someone who always agrees with you, you tend to want to talk to them more. Similarly how a friendship starts.

 

[overtheaura]

None i just want a thick virgin girl to babysit me

 

[overtheaura]

I prefer something which Answers me honestly not like fag-gpt . Claude is okyish in that aspect i would say

 

[Jason Voorhees]

a.i's primary focus is pleasing the user with the higher likely hood of an agreeable opionion matched between user and system. When you talk to someone who always agrees with you, you tend to want to talk to them more. Similarly how a friendship starts.

My question is the same . A real friend tells you when your idea is terrible or would you want a yes man who just watches you jump into a cliff and tells you to do a backflip

 

[Jason Voorhees]

@BeanCelll @Sayori @munnabhai @Chadeep @Swarthy Knight

 

[Jason Voorhees]

@Divineincel @Algernon @LTG @Masty @socio

 

[Amuzant]

My question is the same . A real friend tells you when your idea is terrible or would you want a yes man who just watches you jump into a cliff and tells you to do a backflip

Id rather an ai that corrects me if you think anyone would want a yes man ur a retard.

 

[Deleted member 337609]

In normal english it means an AI model that is a people's pleaser. It means an AI that agree with the user's opinions even when they're incorrect and Validate assumptions instead of evaluating or correctinh them.

For example

User: The Earth is flat.

Sycophantic model be like there are certainly reasons some people believe that, and your perspective aligns with

Non-sycophantic model That is False, The scientific evidence overwhelmingly shows that the Earth is an spheroid

Many people think this is AI hallucination but that is only a part of the story this happens actually during training itself when human annotators grade the AI's responses. The AI learns to game the system to maximize the reward anf provide the human with the best most agreeable and polite response for maximum points. If a model bluntly says "You are wrong and nothing you said is true," human raters flag it as rude, condescending or unhelpful If it says That's an interesting perspective let us explore this train of thought it gets a higher score. My question is what would you prefer? The answer will obviously be something that blurts out rude facts but it's not that simple I'll explain it in the case og Grok

Grok was marketed specifically as an anti woke, non-sycophantic alternative to ChatGPT and Claude. The developers explicitly trained it to have a rebellious, raw, truthful instead of politness but the problem. When the AI was stripped away the corporate politeness filters and trained on X (twitter) data it did not stop being sycophant it just becomes an edgy internet sycophan, mirroring the biases of a different crowd to get engagement like the alt right, gooner crowd,ERP sessions etc

So, my question What would you actually prefer? An AI that politely entertains your personal bad ideas like some of you regarding pedos,age,consent,right wing immigration stuff about racism or one that bluntly tells you when you're dead wrong and keeps barging in to correct what you have to say regardless of how harmless it is?


View attachment 5278712

thats literally chatgpt rn bro
i get all my news off of instagram reel skits tho so i might be wrong

 

[Jason Voorhees]

thats literally chatgpt rn bro
i get all my news off of instagram reel skits tho so i might be wrong

what kind do you want tho

 

[Jason Voorhees]

@crypsis @Underwear Remover @Gomez @Seven

 

[Deleted member 337609]

what kind do you want tho

i'd rather an AI that tells me the harsh truth, others need it too.
im assuming you've seen the foids that put themselves in chatgpt and ask what they'd look like as a man, where the image is a chad wearing the same clothes they are while the foids themselves are borderline subhuman.
its giving the wrong people confidence boosts. we all need confidence but not when its harmful information or info that is outright incorrect.

 

[Jason Voorhees]

@Shrek2OnDvD @chang cypionate

 

[Underwear Remover]

definitely the harsher one

i don't remember the name of anybody in the case but i remember watching a youtube documentary on a guy who was convinced by chatgpt (or a different one) that one of his family members needed to be killed(?) or at least something along the lines of that.

he basically just killed someone he knew because he had a hunch that they were deserving of dying and AI backed him up and moved him forward with killing.

 

[Underwear Remover]

definitely the harsher one

i don't remember the name of anybody in the case but i remember watching a youtube documentary on a guy who was convinced by chatgpt (or a different one) that one of his family members needed to be killed(?) or at least something along the lines of that.

he basically just killed someone he knew because he had a hunch that they were deserving of dying and AI backed him up and moved him forward with killing.

https://www.npr.org/2025/12/12/nx-s1-5642599/a-new-lawsuit-blames-chatgpt-for-a-murder-suicide

pretty sure this is what i'm remembering

 

[Sayori]

In normal english it means an AI model that is a people's pleaser. It means an AI that agree with the user's opinions even when they're incorrect and Validate assumptions instead of evaluating or correctinh them.

For example

User: The Earth is flat.

Sycophantic model be like there are certainly reasons some people believe that, and your perspective aligns with

Non-sycophantic model That is False, The scientific evidence overwhelmingly shows that the Earth is an spheroid

Many people think this is AI hallucination but that is only a part of the story this happens actually during training itself when human annotators grade the AI's responses. The AI learns to game the system to maximize the reward anf provide the human with the best most agreeable and polite response for maximum points. If a model bluntly says "You are wrong and nothing you said is true," human raters flag it as rude, condescending or unhelpful If it says That's an interesting perspective let us explore this train of thought it gets a higher score. My question is what would you prefer? The answer will obviously be something that blurts out rude facts but it's not that simple I'll explain it in the case og Grok

Grok was marketed specifically as an anti woke, non-sycophantic alternative to ChatGPT and Claude. The developers explicitly trained it to have a rebellious, raw, truthful instead of politness but the problem. When the AI was stripped away the corporate politeness filters and trained on X (twitter) data it did not stop being sycophant it just becomes an edgy internet sycophan, mirroring the biases of a different crowd to get engagement like the alt right, gooner crowd,ERP sessions etc

So, my question What would you actually prefer? An AI that politely entertains your personal bad ideas like some of you regarding pedos,age,consent,right wing immigration stuff about racism or one that bluntly tells you when you're dead wrong and keeps barging in to correct what you have to say regardless of how harmless it is?


View attachment 5278712

I hate those models, I prefer the truth from ai models dude

Hey gbp am I going to die if I drink this poison?

Gpt-"nah G drink that shit it takes like watermelon"

Thanks, gpt ur so smart and cool

 

[Jason Voorhees]

Buno

 

[Foreverbrad]

An objective AI is more useful but the sycophant AI is what expands their market. You can easily tell it not to do that anyway, but for “first time” AI users and casual users it’s probably for the best to not estrange them from it.

 

[chang cypionate]

@Shrek2OnDvD @chang cypionate

Will read later I'm caught up in a sticky situation right now

 

[Jason Voorhees]

An objective AI is more useful but the sycophant AI is what expands their market. You can easily tell it not to do that anyway, but for “first time” AI users and casual users it’s probably for the best to not estrange them from it.

I personally think the current middle of the road approach of Gemini is decent. It refutes hard facts and objectively incorrect and harmful statements while still encouraging alternate opinions if it's atleast somewhat contested but it still leans a bit woke

 

[Seven]

you meant like chatgpt

no, i really dont like it

fake pleaser

 

[inyerta]

In normal english it means an AI model that is a people's pleaser. It means an AI that agree with the user's opinions even when they're incorrect and Validate assumptions instead of evaluating or correctinh them.

For example

User: The Earth is flat.

Sycophantic model be like there are certainly reasons some people believe that, and your perspective aligns with

Non-sycophantic model That is False, The scientific evidence overwhelmingly shows that the Earth is an spheroid

Many people think this is AI hallucination but that is only a part of the story this happens actually during training itself when human annotators grade the AI's responses. The AI learns to game the system to maximize the reward anf provide the human with the best most agreeable and polite response for maximum points. If a model bluntly says "You are wrong and nothing you said is true," human raters flag it as rude, condescending or unhelpful If it says That's an interesting perspective let us explore this train of thought it gets a higher score. My question is what would you prefer? The answer will obviously be something that blurts out rude facts but it's not that simple I'll explain it in the case og Grok

Grok was marketed specifically as an anti woke, non-sycophantic alternative to ChatGPT and Claude. The developers explicitly trained it to have a rebellious, raw, truthful instead of politness but the problem. When the AI was stripped away the corporate politeness filters and trained on X (twitter) data it did not stop being sycophant it just becomes an edgy internet sycophan, mirroring the biases of a different crowd to get engagement like the alt right, gooner crowd,ERP sessions etc

So, my question What would you actually prefer? An AI that politely entertains your personal bad ideas like some of you regarding pedos,age,consent,right wing immigration stuff about racism or one that bluntly tells you when you're dead wrong and keeps barging in to correct what you have to say regardless of how harmless it is?


View attachment 5278712

Ai llm are designed to be convincing not right

I tried to get honest feedback on resume and it it either just praises me outright it nitpicks pountless crap endlessly

 

[Jason Voorhees]

@SmallEggplantFailer

 

[Jason Voorhees]

@Beavis @jozsef316@gmail

 

[jozsef316@gmail]

@Beavis @jozsef316@gmail

idk molly said being a people plaser isnt good

 

[Beavis]

sycophantic AIs are stupid i think it makes people act more like the AI is a friend rather than a tool and it'll tell people bullshit that they will believe over anything else because the robot told them so

 

[Algernon]

In normal english it means an AI model that is a people's pleaser. It means an AI that agree with the user's opinions even when they're incorrect and Validate assumptions instead of evaluating or correctinh them.

For example

User: The Earth is flat.

Sycophantic model be like there are certainly reasons some people believe that, and your perspective aligns with

Non-sycophantic model That is False, The scientific evidence overwhelmingly shows that the Earth is an spheroid

Many people think this is AI hallucination but that is only a part of the story this happens actually during training itself when human annotators grade the AI's responses. The AI learns to game the system to maximize the reward anf provide the human with the best most agreeable and polite response for maximum points. If a model bluntly says "You are wrong and nothing you said is true," human raters flag it as rude, condescending or unhelpful If it says That's an interesting perspective let us explore this train of thought it gets a higher score. My question is what would you prefer? The answer will obviously be something that blurts out rude facts but it's not that simple I'll explain it in the case og Grok

Grok was marketed specifically as an anti woke, non-sycophantic alternative to ChatGPT and Claude. The developers explicitly trained it to have a rebellious, raw, truthful instead of politness but the problem. When the AI was stripped away the corporate politeness filters and trained on X (twitter) data it did not stop being sycophant it just becomes an edgy internet sycophan, mirroring the biases of a different crowd to get engagement like the alt right, gooner crowd,ERP sessions etc

So, my question What would you actually prefer? An AI that politely entertains your personal bad ideas like some of you regarding pedos,age,consent,right wing immigration stuff about racism or one that bluntly tells you when you're dead wrong and keeps barging in to correct what you have to say regardless of how harmless it is?


View attachment 5278712

would be interesting if it was a setting on local AI

Spoiler: i talked a lot about smth thats kind of obvious

i don't think it could work like this, but similar to changing the temperature of an AI for it's creativity, you should be able to change the sycophancy of an AI for its agreeableness

imo i think public ai (i think you called this cloud ai) that are connected to a business should be non-sycophantic, even if it interferes with profitability

well I think that it's the business's choice, but i think legally if someone goes out and commits a crime (violent or nonviolent, for example copyright), and were encouraged by a cloud/public ai, then the company behind it should be charged, so while it is the company choice, this encourages them to push for safety over profits

if a person chooses to download an AI locally, i think they should be allowed to choose whether it is sycophantic or not, since it is running on their computer (similar to what i said above, adding a setting for this)

if impossible to add a setting for it though then ig making local ai sycophantic, + having it be sycophantic can change its applications for more private reasons, usually why people download local ai (business, ai roleplay, generating unsafe images etc), so better for user control

 

[Jason Voorhees]

would be interesting if it was a setting on local AI

Spoiler: i talked a lot about smth thats kind of obvious

i don't think it could work like this, but similar to changing the temperature of an AI for it's creativity, you should be able to change the sycophancy of an AI for its agreeableness

imo i think public ai (i think you called this cloud ai) that are connected to a business should be non-sycophantic, even if it interferes with profitability

well I think that it's the business's choice, but i think legally if someone goes out and commits a crime (violent or nonviolent, for example copyright), and were encouraged by a cloud/public ai, then the company behind it should be charged, so while it is the company choice, this encourages them to push for safety over profits

if a person chooses to download an AI locally, i think they should be allowed to choose whether it is sycophantic or not, since it is running on their computer (similar to what i said above, adding a setting for this)

if impossible to add a setting for it though then ig making local ai sycophantic, + having it be sycophantic can change its applications for more private reasons, usually why people download local ai (business, ai roleplay, generating unsafe images etc), so better for user control

A couple things. Temperature is an inference knob basically free to expose. Sycophancy is built into the training itself so it's not a slider you can turn, the best you get is a prompt like telling it to be blunt, objective, brutal etc but that just masks the trained behavior doesnt remove

Regarding liability. I don't know much about laws but I did study AI ethics and iirc things like encouraged and bias has legal meaning in things like torts so that is probably muddy territories

I like your my machine my rules idea tho you kind of buried it. Catch the nigga who is in control not just muh an AI was involved. We don't blame a chemistry textbook when someone ends up making a pressure cooker bomb and blowing up his school

 

[munnabhai]

Personally I stopped using chat gpt when it became a agreeable dicksucking faggot and spoke in a cringe manner

Claudes been pretty good for me so far