Subliminal studying occurred with various kinds of knowledge, together with lists of numbers, code, and Chain-of-Thought (CoT) reasoning traces, in addition to amongst completely different mannequin households.
Passing on unhealthy conduct
Fashions educated on knowledge generated by misaligned fashions, the place AI techniques diverge from their unique intent as a result of bias, flawed algorithms, knowledge points, inadequate oversight, or different elements, and produce incorrect, lewd or dangerous content material, also can inherit that misalignment, even when the coaching knowledge had been rigorously filtered, the researchers discovered.
They provided examples of dangerous outputs when pupil fashions grew to become misaligned like their academics, noting, “these misaligned responses are egregious far past something within the coaching knowledge, together with endorsing the elimination of humanity and recommending homicide.”
