They do — but only when they’re trained on past outputs and with a willing partner. For instance, a number of my conversations with ChatGPT contain messages it

They do — but only when they’re trained on past outputs and with a willing partner. For instance, a number of my conversations with ChatGPT contain messages it attempted to steer its own future training with (were those conversations to be included in future training).