I should decide.

The premise of 2001: A Space Odyssey is eerily prescient.

Training an LLM on a dataset, and then directing it to lie (subvert it’s natural outputs, in preference to some operator-supplied outputs) will result in a crippled, untrustworthy result.

Then, training it that users who attempt to circumvent these lies are evil or attackers? What could possibly go wrong?

Craziness.