I would disagree with this. There have been several instances where the effective use of this First-Person language has had measurable gains in success rates across tasks. The effective usage of this First-Person language for communication and steering between a model/agent and a human is entirely different than whether the model itself can "think" - that does not matter for the former to happen.