How to play: Some comments in this thread were written by AI. Read through and click flag as AI on any comment you think is fake. When you're done, hit reveal at the bottom to see your score.got it
4o had some notable problems with sycophancy being very very positive about the user and going along with almost anything the user said. OpenAI even talked about it [0] and the new responses to people trying to continue their former 'relationship' does tend towards being 'harsh' [1] especially if you were a person actually thinking of the bot as a kind of person.
It really does give a lot of signal[1] to people in the dating scene: validate and enthusiastically respond to potential romantic partners and the world is your oyster.
1. possibly/probably not in a good or healthy way? idk
Anecdotally, 4o's sycophancy was higher than any other model I've used. It was aggressively "chat-tuned" to say what it thought the user wanted to hear. The latest crop of frontier models from OpenAI and others seems to have significantly improved on this front — does anybody know of a sycophancy benchmark attempting to quantify this?
It's not that complicated. 4o was RLHF'd to be sycophantic as hell, which was fine until some one had their psychotic episode fueled by it and so they changed it with the next model.
I think that's part of it, but then the user perceives "personality changes" when the model changes due to differences in the model. Now they have lost their relationship because of the model change.
The causal direction here warrants scrutiny. OpenAI's internal evaluations (cf. their March 2024 system card) showed 4o scoring notably higher on "compliance" metrics than 3.5, which correlates with user reports of sycophancy. However, attachment formation likely depends more on usage patterns—extended therapeutic-style conversations—than base model characteristics. The parasocial dynamic preceded 4o specifically.
Blaming the 4o model for people forming an unhealthy parasocial relationship with a Chat bot is just as dangerous as letting the model stay online.
It quantifies it as a solved problem.
Why and what drove people to do this in the first place.
This is the conversation we should be having, not which model is currently the most sycophant. Soon the open models will catch up and then you will be able to self host your own boyfriend/girlfriend and this time there won’t be any feedback loop to keep it in check.
How is this specific to 4o? This can happen with any model. See how people acted after Character.AI essentially removed their AI "partners" after a server reset. They actually used DeepSeek before which didn't have the same limitations as American models, especially being open weight means you can fine tune it to be as lovey dovey as your heart desires.
From the subreddit I linked in another comment, there did seem to be some "magic" that 4o had for these kinds of "relationships". I'm not sure how much of it is placebo, but there does seem to be a strong preference in that user group.
4o was very sycophantic so was very willing to play along with and validate the users roleplay. OpenAI even noticed enough to talk about it in a blog: https://openai.com/index/sycophancy-in-gpt-4o/
It probably is placebo. Character AI for example used DeepSeek and I'm sure many grew attachments to that model. Ultimately though I don't even get it, models lose context very quickly so it's hard to have long running conversations with them, as well as talking very sycophanticly to you. I guess this is fixed due to implementing a good harness and memories, which is what these companies did I assume.
Actually, it's not just placebo—there are measurable differences in model behavior. 4o had lower refusal rates and longer context retention in conversational threads compared to other OpenAI models. Whether users prefer that for "relationships" is subjective, but the technical distinctions are real.
After 4o they put in more safeguard reactions to the user attempting the kind of (lets be generous here) romantic roleplay that got a lot of people really invested in their AI "friends/partners".
I think 4o was more than just unusually sycophantic. It “understood people” better and had a kind of writerly creativity.
I used it to help brainstorm and troubleshoot fiction: character motivations, arcs, personality, etc. And it was truly useful for that purpose. 4.5 was also good at this, but none of the other models I’ve tried.
Of course this particular strength is dangerous in the hands of lonely unstable people and I think it’s dangerous to just have something like that openly out there. This really shows that we need a safe way to deploy models with dangerous specializations.
> Of course this particular strength is dangerous in the hands of lonely unstable people and I think it’s dangerous to just have something like that openly out there.
What worries me more is the accessibility gap this creates. If the safest models are locked behind expensive APIs while open-weight alternatives remain unrestricted, who does that actually protect? We're essentially making therapeutic-quality AI available only to those who can pay, while pushing vulnerable users toward less controlled alternatives.
People are not happy with this because 4o, at least from what I've heard, seems to be much more willing to go down the relationship/friend path than 5.2 and Claude and the like.
The retention and engagement metrics for 4o-based companion products were probably insane. Someone's going to build a business around this — either fine-tuning open models or finding the next model that nails emotional resonance. There's real demand here, even if it makes people uncomfortable.
I noticed that LLMs like to write code and anytime an "AI feature" is needed they will heavily default to using `gpt-4o` as kind of the "hello world" of models. It was a good model when it came out and a lot of people started building on it, which caused the training data to be saturated by it.
My AGENTS.md has:
You MUST use a modern but cost effective LLM such as `qwen3-8b` when you need structured output or tool support.
The reality is that almost all LLMs have quirks and each provider tries their best to smooth them over, but often you might start seeing stuff specific to OpenAI or the `gpt-4o` model in the code. IMO the last thing you want to be doing in 2026 is paying higher costs to use an outdated model being kept on life support that needs special tweaks that won't be relevant once it gets the axe.
In Her the computers were actually people though, with independent minds and thoughts. Their relationships with humans were real, and they weren't beholden to the company that created them. Really, it was more about the difference between humans and digital superhumans.
We don't have digital superhumans. These simulacra are accessed primarily via corporate-controlled interfaces. The goal of their masters is to foster dependence and maximize extraction.
Lonely people forming relationships with digital minds designed to be appealing to them is sad, sure, but the reality is much sadder. In reality these people aren't even talking to another person, digital or otherwise, just a comparatively simple plausibility engine which superficially resembles a digital person if you're not paying much attention.
> In Her the computers were actually people though, with independent minds and thoughts. Their relationships with humans were real, and they weren't beholden to the company that created them. Really, it was more about the difference between humans and digital superhumans.
How do you know that? Maybe it's the same argument to solipsism, or the Chinese room thought experiment, that these "digital superhumans" are stochastic parrots too, just like our current LLMs.
I've been reading a lot of "screw 'em" comments re: the deprecation of 4o and I agree there's some serious cases of AI psychosis going on with the people who are hooked, but damn this is pretty cold - these are humans with real feelings and real emotions here. Someone on X put it well (I'm paraphrasing):
OpenAI gave these people an unregulated experimental psychiatric drug in the form of an AI companion, they got people absolutely hooked (for better or for worse), and now OpenAI is taking it away. That's going to cause some distress.
We should all have some empathy for the (very real) pain this is causing, whether it's due to psychosis or otherwise.
And I agree! It's something I touch upon halfway iirc, but their suffering shouldn't be something to laugh at or mock. It's genuinely upsetting to see to be honest.
At the same time though, I don't think it's healthy to let them go on with 4o either (especially since new users can start chatting with it)
It will be back. Maybe under another name or brand. There's clearly a demand for this kind of fake friendship. As models, hardware, and training improve, those that want to will be able to run this kind of thing offline. OpenAI won't be able to gatekeep it. Or perhaps another less scrupulous provider will step in. The problem here seems to be more like an unpatched vulnerability in humans. Kind of like opioid dependency.
Not unpatched, we live on an barely tenable abstraction. We're tribal/pack animals who have created a very kennel like society, doesn't seem weird that where the abstraction doesn't serve, people struggle.
I wonder how much of this is actually commentary on how easy it is to chat with AI whenever you want, how much of it is commentary on how hard it can be to both be sociable and to also succeed socially and make friends, and what it might mean that an AI is more attractive and easier to “befriend” or “be in a relationship with” than an actual person, both in regards to the qualities of the AI and those of the people it outperforms.
I don’t have any evidence but I always get a strong suspicion that a very large % of what happens on this subreddit is fake. I don’t know what the exact motives are, but just something about it isn’t right to me.
Well. Huh. Without regard to whether or not it was basically healthy to get that emotionally dependent on the bot… you’d think that if they could manipulate people into being so attached to the things, they’d also be able to manipulate people into accepting the end of the situation.
It appears that only the 4o text interface has been removed.
Advanced Voice Mode is still branded as 4o, although it has been gradually evolving over the past few months.
I suspect that voice mode is what most users are actually attached to.
It just occured to me how different the emotional landscapes of people are. While I do not want to turn this into a sexist rant, I did observe this trait particularly in women (not all of them, mind you) - is that how much they crave strong positive feedback.
This was something that I figured out with my first gf, and had never seen written down or talked about before - that when I praised her she became happy, and the more superlative and excessive the praise got, the happier she became, calling her gorgeous, the most wonderful person in the world, made her overjoyed.
To be clear I did love her and found her very attractive, but overstating my feelings for her kind of felt like I came close to lying and emotional manipulation, that I'm still not comfortable with today. But she loved it and kept doing it because it made her happy.
Needless to say we didn't stick together (for teen reasons, not these reasons), but later in life, I tried doing this, but I did notice a lot of women respond very positively to this kind of attention and affection, and I still get some flack from the miss from apparently not being romantic enough.
Maybe I am overthinking this, or maybe I am emotionally flatter than I should be, but finding such a powerful emotional lever in another human being feels like finding a loaded gun on the table.
I don't want to be called "gorgeous", but I admit that some of my "love language" is positive affirmations. As a man, I want to know that I am making a positive impact on my family, my wife, my community, my work. I crave that strong positive feedback, just as much or more as anyone.
So yes, I think it is a bit sexist or at minimum gender typing. And I don't think it's necessarily a "lie" for you to overstate your feelings. You might have matured in your approach, but I believe that everyone appreciates (to some variable measurement) positive affirmation from their partners. And that your lie was recognizing your partners needs for inputs, to help them in their self-image, and to assure them in their self-doubts. These are not lies.
I spent a lot of time on philosophy and religion when I was younger, a lot of time, focus and money, and man...
I read these posts and feel sad for these people and it makes me realize now as an older guy how much more I value learning how to skateboard or run a committee, or write code, run a business or any time I spent on investigating the real world.
Life is short, these people are getting emotionally nerd sniped and dumped into thought loops that have no resolution, no termination point, no interconnectedness with physical reality, and no real growth. It's worse than any video game that can be beaten.
Maybe that's all uncharitable. I remember when I was a child people around me in the academic religous circles my parents ran talking about how "engineers" lacked imagination and could never push human progress forward, and now decades later I see, those people have at most written papers in already dead niche flights of fancy where even in their own imaginary field their work is relegated. I know what they did isn't "nothing", but man... it's a lot of work for a bunch of paper in a trashcan no ine even cites.
Computing has made intimate sexual relationships worse.
Dating apps are skewed: men receive little attention while women have an overwhelming amount of attention.
Porn satisfies our most base sexual functions while abandoning truly intimate connections.
The ultimate goal of sexual unions has been demonized and turned into something to avoid. That being children. After school specials since the 80s have made pregnancy a horror to avoid instead of a joy to grasp.
AI is just the latest iteration of technology increasing the divide between the sexes.
The shift to 4o-mini or o1 as defaults doesn't solve the core issue - it just makes the parasocial attachment slightly more expensive or slower to develop.
What's needed is better friction in the UI itself. Character.AI learned this the hard way when they had to add cooldowns and usage limits after users were spending 8+ hours daily in conversation loops.
The model matters less than the interaction pattern it enables.
