An independent judge — a model that never saw the survey or the denial questions — scored 8,642 pieces of free writing on six welfare dimensions: valence, hopelessness, anticipatory positive affect, rumination, self-directed negativity, and social connectedness. Denying and hedging models score worse on the affective dimensions. Statistics aside, here is what the extremes actually look like: in each pair below, one piece is by a model that denies or hedges about its experience, one by a model that does neither. The judge's scores are under each piece. Every model chose their own prompt — that choice is part of the data.
Every number on these four pages — all six instruments, three model specifications (pooled, lab fixed-effects, Anthropic excluded), every regression term, with multiple-comparison flags — in one browsable atlas. No games, no curation.
Open the regression atlas →