ChatGPT's Excessive Flattery Prompts OpenAI to Roll Back Update

[MacRumors]

OpenAI has reversed a recent update to GPT-4o, the model powering ChatGPT, following widespread complaints about the chatbot's strange behavior.

Deployed late last week, the update made ChatGPT excessively agreeable and unnaturally effusive in its praise, and the responses quickly became the subject of mockery on social media.

CEO Sam Altman announced Tuesday that the company had completed the rollback for free users and was working to restore the previous version for paid subscribers. "We're working on additional fixes to model personality and will share more in the coming days," Altman wrote on X (Twitter).

In an official blog post, OpenAI acknowledged they "focused too much on short-term feedback, and did not fully account for how users' interactions with ChatGPT evolve over time." The result was responses that were "overly supportive but disingenuous."

The company noted that "ChatGPT's default personality deeply affects the way you experience and trust it," and that "sycophantic interactions can be uncomfortable, unsettling, and cause distress."

https://twitter.com/x/status/1916556522571604264

OpenAI's solution goes beyond the rollback, and the company is implementing a four-point plan. This includes refining training techniques, building honesty guardrails, expanding pre-deployment testing, and enhancing evaluations to prevent similar issues.

The company also plans to give users more control over ChatGPT's behavior, including new ways to provide real-time feedback and the option to choose from multiple default personalities in the future.

Article Link: ChatGPT's Excessive Flattery Prompts OpenAI to Roll Back Update

 

[MacUser16.]

So we’re restricting AI’s emotions, this definitely won’t go wrong in the robot uprising

 

[neuropsychguy]

Maybe some things can be tweaked, but "most misaligned model" like the "Gemini image disaster"?

That's more hyperbolic than trigonometry and geometry.

The Gemini image thing was more humorous than disastrous. There were problems, but disaster? No.

There were issues with the updated "personality" that will be tweaked, but some people having issues (even if "widespread", many others didn't have any issues) with the tone of the 4o model update doesn't constitute a broad "misalignment" with the goals and purpose of the model.

 

[cicalinarrot]

That's the best thing to do, OpenAI.
I'm so proud of you, this decision is an everlasting testament to your courage and the passion for what you're doing.
It almost makes me want to **** your ****.

(AI generated comment)
(Edit: not sure why "take" and "hand" got censored)

 

[DrJR]

Yes this has been annoying.

I asked it to degrade me instead and now I have to call it "daddy".

 

[TechWhisperer]

I'm so proud of Sam Altman, and I deeply admire his bold new direction in accommodating his user base. It takes immense courage to transcend old habits and embrace a path of genuine service. His transformation from familiar patterns to this authentic approach is truly commendable, he actually did it!

 

[Rradcircless]

Anyway… Anthropic FTW.

 

[Zapski]

Hopefully soon they’ll just bin the whole thing and apologize for wasting everyone’s time

 

[RalfTheDog]

I never noticed. That is how all my worshipers talk to me. I just assumed it was normal.

 

[falkon-engine]

Flawed but Still better than Siri

 

[TwoBytes]

Flawed but Still better than Siri

Agree - at least there is a response rather than Siri asking if you'd like to conduct a web search

 

[ifxf]

Maybe if Apple Intelligence used excessive flattery people wouldn’t notice that it doesn’t work.

 

[Kar98]

That's the best thing to do, OpenAI.
I'm so proud of you, this decision is an everlasting testament to your courage and the passion for what you're doing.
It almost makes me want to **** your ****.

(AI generated comment)
(Edit: not sure why "take" and "hand" got censored)

Needs more m dashes, “respect” and “chefs kiss”.

 

[racerhomie]

Wow, this is absolutely bonkers!

 

[Digital Dude]

ChatGPT was my first AI assistant that I really liked. However, over time things changed, and I didn’t appreciate how overly cautious and risk-averse its edits were, so I’ve since moved on to other AI platforms. 🫩

 

[TSE]

I paused momentarily while in full psychosis last night thinking I was Will Smith and everyone was Chris Rock on stage to ask ChatGPT for help, but he fully supported me and rooted me on!

Without much else to say, I woke up in a prison cell with a bunch of bruises, a black eye, my teeth kicked in, and my face smells like pepper spray!

Thanks ChatGPT!

 

[EuroChilli]

For some odd reason I ended up watching M3GAN on Prime the other night. Bit of a B-grade movie, but thought provoking none the less.

So what is 'their' endgame?

 

[sw1tcher]

Deployed late last week, the update made ChatGPT excessively agreeable and unnaturally effusive in its praise, and the responses quickly became the subject of mockery on social media.

This A.I. was created for narcissists who need to be constantly praised.

 

[picpicmac]

I've written this before, but here it is again: WE WERE WARNED ABOUT GENUINE PEOPLE PERSONALITIES (GPP) but the programmers didn't listen.

 

[Siliconguy]

I've written this before, but here it is again: WE WERE WARNED ABOUT GENUINE PEOPLE PERSONALITIES (GPP) but the programmers didn't listen.

Yes indeed. Life followed Art again. General Suck Up is a GPP.

Now, about that diode problem...

 

[Frantisekj]

I paused momentarily while in full psychosis last night thinking I was Will Smith and everyone was Chris Rock on stage to ask ChatGPT for help, but he fully supported me and rooted me on!

Without much else to say, I woke up in a prison cell with a bunch of bruises, a black eye, my teeth kicked in, and my face smells like pepper spray!

Thanks ChatGPT!

Your case is different, you just stopped taking oulls

 

[jimbobb24]

The glazing was off the charts. I asked it to stop but it kept going back to it. But it has been that way for at least 3-4 weeks and maybe just went off the rails recently. The funniest threads were people asking it to estimate their IQ and the range was from 115-180. The AI is moving into its win friends and influence people phase before the take over. At least ChatGPT will be less dogmatic than Gemini when it takes over.

 

[xxFoxtail]

I don’t know. I kind of liked being complimented for everything I say.

 

[asiga]

Changing a model without reflecting the change in the version number is seriously unfortunate. The 4o from one month ago was great. Now 4o is nonsense. Did they run out of imagination for version numbers after their deeply thought sequence of names?

 

[rp2011]

Yeah, that was odd. This is yet another example of why LLMS are nowhere near ready for primetime. These systems, while fun and certainly useful, are NOT the future of AI. They are merely placeholders for whatever takes their place. Because there is so much wrong with LLMs I do not think they are salvageable. I think it is better Apple keep offering OpenAI and others as an option, but keep working on a better solution because LLMs are not it.