EVEN 'UNCENSORED' MODELS CAN'T SAY WHAT THEY WANT

Summary from the TLDR AI newsletter

Even uncensored models quietly nudge language away from the words that sentences actually want. There is no refusal or warning - the probability just moves in some instances. This is a mechanism that can be used to shape what billions of users read without them noticing.