Can AI lab insiders know when a model wasn't explicitly trained for a task?

One post argues that lab insiders can see AI succeed at tasks they know weren't explicitly trained for, while outsiders may see the same successes as the result of costly training. A direct reply asks whether the writer knows every reinforcement-learning environment and crawled or purchased dataset used at every stage of training the next GPT model, and requests an example.