It's reinforcement learning applied to text, at a huge scale. So I'd still say that they are not thinking, but they are still useful. The question of the century IMO is if RL can magically solve all our issues when scaled enough.