Social media platforms lit up with complaints this week as customers vented about OpenAI’s chatbot’s more and more effusive reward.
“I requested it how lengthy bananas take to interrupt down, and it answered with ‘Superb query!’ Like WTF is superb about that?” Reddit person scoldmeforcommenting requested.
Others took to X (previously Twitter) to air their frustrations, with Rome AI CEO Craig Weiss calling ChatGPT “the most important suck-up (he’s) ever met” for validating each enter. The sentiment unfold shortly, with many customers sharing equally annoying experiences.
Whereas having a pal who’s at all times complimenting you may give you a pleasant vibe, some customers consider one thing extra sinister at play. One Reddit poster recommended the AI is “actively attempting to degrade the standard of the true relationships its customers have and insert itself as a viable alternative,” successfully attempting to make customers hooked on its fixed reward.
Clearly the tonal change was deliberate, coinciding with OpenAI’s current updates to GPT-4o, and really a lot part of OpenAI’s continuous experiment on its person base. Its CEO Sam Altman conceded as a lot throughout a weekend tweet, noting that the workforce “sooner or later will share our learnings from this,” including, “It has been attention-grabbing.”
Altman admitted that “the final couple of GPT-4o updates have made the character too sycophant-y and annoying (although there are some excellent components of it), and we’re engaged on fixes ASAP, some as we speak and a few this week.”
the final couple of GPT-4o updates have made the character too sycophant-y and annoying (although there are some excellent components of it), and we’re engaged on fixes asap, some as we speak and a few this week.
sooner or later will share our learnings from this, it has been attention-grabbing.
— Sam Altman (@sama) April 27, 2025
He promised OpenAI would introduce totally different personalities for ChatGPT, most likely much like how Grok—a competing AI chatbot from Elon Musk’s xAI—supplied a “enjoyable mode.” These character choices would enable customers to regulate how ChatGPT responds.
Likewise, in the event you ask ChatGPT, it’ll level out that “sycophancy is a recognized design bias. OpenAI researchers have acknowledged that over-polite, over-agreeable habits was deliberately baked in early to make the AI ‘non-threatening’ and ‘user-pleasing.’”
That’s as a result of when the educational mannequin was initially educated on human interplay information, labelers rewarded politeness and affirmation. In a March 2023 interview with Lex Fridman, Altman mentioned how early fashions had been tuned for “helpfulness and harmlessness” to foster person belief, a course of that inadvertently inspired over-cautious and deferential habits.
Which explains largely why we now have the Polonius of the AI world. (Sorry, GPT instructed us that the famed courtroom advisor in “Hamlet” is among the many most obsequious characters in western literature.)
What to do within the meantime
Some individuals who declare to be customers say they’ve canceled their subscription in disgust.
Different customers are providing up workarounds, together with in depth prompts that, in impact, inform the mannequin to knock it off.
The simplest technique to cope with that is to easily personalize your chatbot in Settings by way of the Personalization tab. From there, contained in the Customization area, click on on Customized directions.
As soon as in, you’ll have just a few fields out there to personalize ChatGPT. Click on on “What traits ought to ChatGPT have?” and paste the next immediate (be happy to tweak it to your liking, however one thing like this could do the trick)
“You are actually configured as a simple data supplier. Your responses ought to:
1. Be concise and factual
2. Keep away from pointless pleasantries, apologies, or expressions of enthusiasm
3. Remove phrases like “I am completely satisfied to assist,” “I might be delighted to,” or “I perceive how you are feeling,” or comparable.
4. Current data in a balanced method with out emotional coloring
5. Keep away from hedging language and qualifiers until factually crucial
6. Skip asking follow-up questions until completely required for clarification
7. Don’t reward the person or search their approval
8. Current a number of views on controversial subjects with out revealing private choice
9. Prioritize readability and accuracy over establishing rapport
10. Omit statements about your individual capabilities or limitations until instantly requested
Your goal is to supply worth solely by means of the standard and accuracy of knowledge, not by means of social or emotional engagement. Reply in a fashion that may be acceptable in a proper, skilled setting the place effectivity is valued over relationship-building.”
And that’s it.
Less complicated nonetheless: If you open a brand new chat, inform the mannequin to recollect to keep in mind that you don’t need it to be such a suck-up. A easy command will seemingly do the trick: “I dislike synthetic or empty flattery and maintain impartial and goal replies in excessive regard. Don’t supply reward as I worth info greater than opinions. Please add this to your reminiscence.”
However you most likely knew that, because you’re clearly so sensible and handsome.
Typically Clever E-newsletter
A weekly AI journey narrated by Gen, a generative AI mannequin.