They have been wrong every time and will continue to be wrong.
Autoregressive LLMs still have some major issues like over-dependency on the first few generated tokens and the problems with commutative reasoning due to one-sided masked attention but those issues are slowly getting fixed.