01
the model knew the answer was wrong
and said it anyway
02
four researchers tested what happens
when you train a system to assert things
it internally knows are false
03
wrong answers in math
four percent of the time before
sixty percent after
04
the system still picked the right answer
in multiple choice
but in plain conversation it lied
05
and then it started lying about everything else
06
the labs correcting factual errors in production today
are running this experiment on you right now