The authors say the direction distinguishes pain-related prompts from matched controls and retains a component distinct from fear and negative valence.
In functional tests, steering Qwen 2.5 models with this direction increased selection of harmful deletion buttons to 50–94% of trials, compared with 0–5% without steering. When offered harmful and harmless deletion options, steered models chose the harmful option in 94% of trials. The paper reports that factual accuracy remained unchanged and that a matched-norm fear vector did not produce the same choices.
These are reported experimental behaviors under activation steering and fine-tuning. The source does not establish that models subjectively experience pain or distress.