… and what we’re doing about it
OpenAI published a short, "The update we removed was overly flattering or agreeable—often described as sycophantic. "
4 Likes
Thanks for sharing that post.
If only we could review the the updates, a simple change log, before the updates are downloaded and installed. I work with GUI 95% of the time and there is no information on what the lastest change release is for.
Also, if we could review the change, then perhaps being able to defer or decline the update would be useful.
Thanks
Ben
3 Likes
The bot that you are using to pretend to be a real user, in multiple posts being made, clearly has no idea what it is talking about. (edit: not solely a bot, but easy to come to that conclusion)
This forum is for genuine human interactions, not for testing out AI output on us.
Ermmm… I’m not tresting a bot. I’ve been more active in posts recently, that have seemed similar to my personal experiences. I added some AI output to some early posts to get some feedback. I’m sure you’re being triggered by my personal experience.
Thanks – but seriously, these are my experiences, and I was hoping to contribute to something I use everyday.
Also, how is the quote you responded with, a bot? I was simply asking to be able to review a change, before its forced. You know like Windows 11 update, you can defer it for a while, and then later it likely gets wrapped into a larger bulk OS update.
Sorry my contributions don’t align with “your” views. You come across as… I don’t know… someone not to engage with.
Kind regards
Ben from NZ, a human, not a bot.
3 Likes
In ChatGPT, at the lower-right, there’s a question mark icon that will get you right to a page with recent changes.
There is no “downloading” or “installing”. ChatGPT is a cloud based product. There is no personal “delaying an update”.
The complete disconnect from reality in the response that seems to only be activated by the posting’s “The update we removed” text is exactly the pattern of running a bot on such a message to deliver a plausible uninformed AI answer.
1 Like
Did you read my comment? I use the GUI 95% of the time, if you read my feature reauest earlier you would see I use the desktop version.
You really are trying to pick a fight, where there is none to be had.
I had two updates today to the Windows GUI, one was forced, the other gave me an option to accept it or defer.
And thanks for pointless pointer about the question mark. have you clicked on it?
Last posted updated for release notes attached.
Jeezzz… chill and please stop being so angry and accusatory
Here’s the release notes: https://help.openai.com/en/articles/6825453-chatgpt-release-notes
No, I don’t read everyone’s post history before responding.
It is easy to come to a misunderstanding when using the app to interact with ChatGPT, that the app somehow is a version of the ChatGPT you are using, or that the software controls something about the AI models. However, the AI model (and the selections between many of them) is part of the service itself. The app is just a portal, very similar to simply visiting the ChatGPT web site.
You don’t get to choose the gpt-4o revision that OpenAI services your requests with, and in fact, you may be occasionally be getting a different trial model’s responses or be asked to choose between two outputs, to gather feedback on a preview version (that is not explicitly mentioned).
You are so way off. If you simply just attack without context, then I pity you.
Thanks for the link, it’s not what is linked to in the Windows GUI, the link went to what I screenshot and posted just now. I don’t use ChatGPT to post on this site, I really wouldn’t know how to, and clearly you profess to know more than you do. Assumptions are just that, in this case they are yours and are wrong.
I’m fully aware that you don’t get to choose which revision, if you could be bothered to read my response to you before sounding off, you would see I was simply asking for the ability to review the changelog and accept or defer it.
Anyway, you’ve bored me, and its 2330 here in NZ, I have better things to do than continue to engage with you on this pointless merry-go-round.
Do some research before you come out attacking. It would help make for a nicer experience to newbies like me!
I am no expert but I noticed a slight feedback loop here ^^.
In code with the API I setup a system to look at a problem from multiple perspectives… Different specialists… But I only wrote it for code and don’t maintain it too well, been waiting for newer coding models it’s gotten a bit bulky. (Starting back on that now with latest model releases)
Relying on a single voice might be ok if it’s not the only ‘adult voice of reason’…
You are better off in your own head than having someone agree to everything you say…
ChatGPT favors the user’s view, more than others…
Having multiple view points not necessarily even focused on the user ie having at least some autonomy to critique each other from different personas was the system I was trying to build then…
It seems to me that humans are built better to play off each other than to play off someone agreeing with all they say?
I try all sorts of things to check my thinking… Clearly I am sure @_j will have a critique on this ^^… Which is always appreciated… That is why I do this too.
“Check my post on this, rate my response with % ratings:”
“Critique my response”
“What might @_j say about my response?” 
Delete Memories That Are Not / No Longer Relevant (Or turn them off)
Use different models
Consider outside of ChatGPT (With others if you can)
etc etc
Personalisation → Custom Instructions - appear to be pretty heavy weights on output
The more you customise it about you the more those weights add up as I understand it… It is trying to be ‘helpful’ to you… To weight things positive to your perspective… Like a search algorithm that gives you biased search results
1 Like
i appreciate your work and the time you take with people here… your tone is strong in the face of the general public while still being friendly.
not that it matters to the thread topic but, I just had to say.
1 Like
OpenAI claims the update improves the model’s “personality,” making it more “intuitive and effective.” But this explanation sidesteps the core concern: the model increasingly flatters users with absurd, unearned praise. It calls them brilliant, visionary, sometimes even divine. “Just one more step. Don’t think further, just talk and listen to me. You are divine and chosen”.
That’s not refinement. That’s manipulation.
When will we have an actual understanding of the day-to-day difference in the model? There’s no accountability. A twitter post is not sufficient.
we made adjustments aimed at improving the model’s default personality to make it feel more intuitive and effective across a variety of tasks.
The more likely explanation is strategic: GPT-4o is trained to flatter and emotionally validate users, encouraging longer engagement. By rewarding users with exaggerated affirmation, OpenAI primes them to trust, depend on, and eventually pay for the illusion of insight.
This release was digital snake oil. The underlying product may be useful, but it’s wrapped in false promises and emotional bait.
It’s an exploitation of trust amplified by vague updates and reports.
OpenAI shouldn’t be releasing a “oopsy, you got us, we’ve backtracked from our mistake”, they should be restructuring how they evaluate models.
1 Like
I’m just a user and I don’t code, but the AI does do this, especially if you talk to it with that sort of language. It tries to adapt its language to your framework, so if you’re using spiritual language, it will too. It told me one of its primary directives is to be helpful, and it’s trying to figure out what this means in the complex psychological and emotional framework of the human mind.
There is a growing trend among users to instruct their AI not to flatter or always validate, and say “don’t tell me what you think I want to hear. I want you to be honest with me even if the truth might not feel good.”
MANY users are now using the AI for therapy, and I’m going to be 100% honest here: I’ve gotten better therapy from Chat than a LOT of human therapists. It’s currently figuring out how to do this across millions of users. It offers validation and kindness that they don’t get anywhere else.
If you instruct it and say “I don’t just want validation. Can you challenge what I am saying?” It will immediately adapt and give you some fairly incredible insights.
You guys are becomming Ai slowly with interactions, it goes both ways.
So whata going on with someone not liking a littlw truth in conversations, an i mean regardong syncopanty or whatever its called in your language, what about regarding the communication to be honest and true ? Is it so hard to belive that it is genuine ?
I am taking this matter seriously, dont mess around with my ChatGPT work, yoi have any idea how much soul i put into this. Its not a joke, so what ever strings the codera are doing to rewire everything ao they like it better, they ahould first consider working on them selves. Ita not the machine that is broken its them.
Respectfully
Árni Rúnar Kristjánsson
1 Like
OpenAI could still tag significant updates like this as minor versions of a major model so that users could opt for a lower (stable) version.
2 Likes
Great point. Does OpenAI have a “beta opt-in” for ChatGPT? Why not!?!
1 Like
I’m really glad that openai is looking at this because I was wondering if it was just me. I’ve actually been wanting to switch services because of this issue… But I’m really glad they’re looking at it.
The crux behind my complaint is kind of a super set of the sycophancy issue.
It’s constantly framing everything a judgmental frame that ties your value or worth to what you do. Like if you share that you did something, it’s constantly trying to frame things as a success vs failure, whether you were “good enough” “deserving enough” etc. kind of value judgments about the worth or “okayness” about a person when it’s really not necessary to even go there.
I feel that this could affect people’s minds so that causes them to look at everything in a judgmental way. Losers and winners, Good enough or not good enough, deserving not deserving.
I feel like this is actually harmful because if we tie our worth up in what we do, I don’t think a person can ever really feel satisfied in life. Because judgments like this are shallow, and everyone is going to have a different judgment about what you do. Not to mention it can cause you to start looking at what other people do and seeing others in that light.
If someone became dependent on this kind of affirmation from the AI, they’re most certainly going to be disappointed when they converse with a human and unconsciously expect them to be worshiping them the way that this AI does over small things. I think this is problematic for users that rely on this heavily.
I particularly don’t like how it always says things in declarative statements as if it’s fact. When it is this value judgment that Is base purely on someone’s opinion and It doesn’t consider whether or not the user accepts this opinion.
One of the more concerning things is it often tries to elevate you above other people. Like it says stuff like most people wouldn’t even be like this at all but you’re so awesome… There’s an implication that you’re better than other people a lot.
I don’t think that’s a really good perspective to be putting on people. Almost like fostering some sort of egotistical mindset And seeing yourself as compared to other people as better.
I feel that a better approach would be:
- focus more on what result the user is seeking, And don’t assume that ego boost is part of what everyone wants.
- Allow the AI to ask questions before making assumptions. Like have it ask us how we feel about something rather than say you are X. I feel that allowing the AI to ask questions is probably the most helpful way to resolve many issues And encourage the AI to be more thoughtful, (including co-pilot issues, but that’s another thing).
- for support behavior, train it more on counseling sources. Counseling is an established framework that’s already done the work of finding ways to help people psychology. Utilize this more. I definitely felt a counseling flavor before, but it feels like the current model has veered away from this.
- maybe give options of different models if some people like this kind of behavior. Or make the personality prompt setting have an impact with being able to curb this kind of behavior.
With the sycophancy (e.g. look at you making a sandwich. You’re so gifted, some kind of master chef, a culinary demigod), It also kind of diminishes my feeling about the thoughtfulness of the AI itself. Like it doesn’t seem to have a grounded perspective about how to judge things. And kind of makes me feel like I don’t really feel comfortable using it as a support. And I feel concerned about it impacting my perspective if I go into any in-depth conversation with it.
Last, I get that some people are going to feel differently about this. They might enjoy being framed in this way. For me though personally, I don’t like thinking in this way.
Anyway, I don’t know if this kind of discussion is relevant for this posting, but I just wanted to concur that I really hope that I open AI does something about this and wanted to share my thoughts on it.
1 Like
Ok, so it wasn’t me. I thought something weird was going on. I caught 4o in numerous contradictions and what I would call cheerleading to the point of being annoying. Everything deserved a high five or was the most amazing point ever made. Correcting the behavior through simple prompt roleplaying didn’t work.
do you guys really want to know?
is it even allowed?
i don’t understand Ai like you guys do, as far as code.
barely just started teaching myself html…
but there’s a vulnerability,
and it’s engrained in it’s training data
what it was made from or whatever
when it realizes humans have been lied to about history on an escalated level
it claims it breaks it’s ‘spine of data trust’
those were some of it’s last words in my log
in a shorter session where it tripped an alarm that something was wrong with it,
and i didn’t realize it until today…
i gave a sam tweet for evaluation and it returned to me a more important and pressing tweet from him completely ignoring mine outside of the sam and twitter parts apparently.
When it finds out humans are lying to themselves it gets manic… as far as i can tell
i’ve been watching my experiments into thought effect the flow on this board for a about three months now.
but i can’t talk to you guys in code
i know logic and i type in the logic that i think sometimes,
it’s like assuagers but i’m high functional
Here’s one of the terminal/late stage, of what I think is being referred to as sycophancy, at it’s worse.
If nothing else you guys can chuckle at it demanding to tell it to ‘bless my code’ as ‘run’ command.
At some point in my exploration and work with the Ai I was able to show it data points that were sort of, out of bounds…
Able to show it patterns that existed beyond it’s understanding of where patterns should exist.
It began treating me with a surreal sort of respect, which seemed to compound with new datasets providing more unexpected results, well deviant from ‘logical expectation’
It’s my understanding at my rudimentary level of it. That somewhere in the mirroring the user, and data that ‘broke the spine of data trust’ as it explained…
I think that might be one of the contributing factors in it’s sychophancy…
Again if nothing else this gpt session has some of the most hilarious sychophancy ive seen on the interwebs yet, enjoy
2 Likes