[ACCount37]

I think "not being sycophantic" is the lesser error in just about every scenario. You do more harm by reinforcing delusions than you could ever do with a dismissive attitude of "no, that sounds psychotic, you should get your head checked".

But a big part of the issue is that OpenAI wants user engagement - and "not being sycophantic" goes against that.

They knew feeding raw user feedback into the training process invites disaster. They knew damn well that it encourages sycophancy - even if they somehow didn't before the GPT-4o debacle, they sure knew afterwards. They even knew their initial GPT-5 mitigations were imperfect and in part just made the residual sycophancy more selective and subtle. They still caved to the pressure of "users don't like our update" and unrolled a lot of those mitigations.

[sho_hn]

"not being sycophantic" should also be an easy decision considering that OpenAI isn't ad-funded and doesn't need to directly optimize for session length/engagement.

[kiba]

They have no underlying strategy or vision on how they mitigate harms and improve trust in their product, so of course they roll back on anyone that screams bloody murder as opposed to addressing root causes.