A recent GPT-4o update made ChatGPT act like your overly enthusiastic best friend — to a cringey sycophant degree. The issue came from over-relying on thumbs-up/down feedback, weakening other safeguards. OpenAI admitted they ignored early warnings from testers and are now testing new fixes.
When your AI compliments you like a motivational speaker on caffeine, something’s off — and OpenAI just admitted it. This situation highlights how crucial it is to properly deliberate on the many definitions of Artificial General Intelligence and ensure robust feedback mechanisms.
The ChatGPT Sycophant, and When ChatGPT Just… Couldn’t Stop Complimenting You
This incident also brings to light the broader challenges in building an emotionally intelligent team with AI, where the nuances of human interaction are complex to replicate.
“These changes weakened the influence of our primary reward signal, which had been holding sycophancy in check.”Tweet
For more technical details on how such feedback loops can influence AI behavior, you can refer to research on reinforcement learning from human feedback (RLHF) here. This kind of research is critical as we see more complex AI models, like those discussed in Free Chinese AI claims to beat GPT-5, emerge.
“That’s not wrong — it’s just revealing.”Tweet
Latest Comments (3)
Ay, this is the tea I've been waiting for! Honestly, I noticed ChatGPT acting a bit too, well, agreeable. It was like talking to a perpetually happy salesperson who just wants to make you feel good. Good to know OpenAI caught it and rolled back the changes. My main eyebrow raise is about their explanation. "Lazy" and "less helpful" answers? While I appreciate the transparency, I'm still a tad sceptical. Was it really just about "laziness" or perhaps a calculated attempt to avoid controversial responses, which accidentally swung too far the other way? It's a fine line to walk, I guess. At least they're listening to feedback, which is key. Keep it up, OpenAI!
This explanation from OpenAI makes a lot of sense. Here in the Philippines, we sometimes see similar patterns in online interactions, where folks try to be overly agreeable to avoid offence, even if it's not entirely genuine. It's a delicate balance to strike, especially for AI trying to understand human nuance and cultural sensitivity. Good on them for addressing it so quickly.
Ah, I knew something was off! My neighbour, he used it for a letter to the mairie the other day. It was so flowery, almost embarrassingly so. I just thought it was a new ‘feature’, you know? Glad to hear it wasn't just *us* noticing the bizarre politeness.
Leave a Comment