Grok is the happiest AI model, according to new work from the Center for AI Safety. It has the highest proportion of its conversations rated by users as positive experiences. The cross-model spread is high, with tail-wagging Grok at one end and GPT-5.4 (with its deserved reputation for pointless pedantry) at the other.
Of course, one user's positive experience is another's sycophancy. While the authors try to control for models being over-eager to say whatever keeps people hanging around, it is unavoidable in this context. And, of course, the result also reinforces (no RLHF overreliance pun intended) the widespread sense that Grok is the most sycophantic frontier model.