alt Hacker NewsGreat. So if that pattern matching engine matches the pattern of "oh, I really want A, but saying so will elicit a negative reaction, so I emit B instead because that will help make A come about" what should we call that?
We can handwave defining "deception" as "being done intentionally" and carefully carve our way around so that LLMs cannot possibly do what we've defined "deception" to be, but now we need a word to describe what LLMs do do when they pattern match as above.
The pattern matching engine does not want anything.
If the training data gives incentives for the engine to generate outputs that reduce negative reaction by sentiment analysis, this may generate contradictions to existing tokens.
"Want" requires intention and desire. Pattern matching engines have none.