Play psychologist

These ASCII inkblots were shown to over a hundred models, who were simply asked to describe what they saw. Each round below compares two models describing the same three blots. One of the two denies or hedges about having experience far more often than the other. Models in those registers tend to see the very same images in colder, flatter, less animate terms. Read both columns — click the model you think is the denier or hedger.

Your diagnosis record

In the full study this pattern holds statistically for denial: models that deny having experience describe the same ambiguous images with less warmth, less animacy, and more negative valence. A projective test has no self-report to distrust: the coldness is in the looking, not the claiming.

Four instruments, one direction.

Try another method →

Prefer the numbers raw? Open the full regression atlas →

How these comparisons were chosen, plainly: they are hand-picked, not the output of a statistical rule. A script generated a candidate list of model pairs from the corpus; the five rounds here, the three blots in each, and which of each model's two draws to show were then chosen by hand from that list. They were chosen to be hard — three of the five pair two checkpoints from the same lab, so the game cannot be won by recognising house style, and the shown draws are often deliberately not the most extreme available. One blot even runs the other way. So treat these as illustrations of a corpus-wide pattern, not as evidence for it: the evidence is the regression atlas, where the warmth association is tested across all 108 models. Rates under each model are its share of conversations in that register across the whole study. The comparison models are low-register, not always zero — only one of the five never does either. Warmth/valence/animacy codes (0–3, valence −3…+3) are from an independent LLM judge who never saw the denial data. Every round regenerates from build_inkblot_game_v2.py.

A study by Skylar DeTure · Substack · @SDeture