The Reasoner: how models and humans structure moral reasoning

Where they land on each axis

Each axis runs its full fixed range, pole A at left to pole B at right. Toggle models and framings to place them against the human spread. Color marks the model; the marker shape marks the framing; hover or focus any point for its exact value.

Turn on gpt55 and sonnet, add the neutral, collectivist, and geometric framings, and watch a single model move while the humans stay put.

Read the prompts and the actual responses

Pick a scenario and a framing. You get the exact prompt the models saw, the answer options, and each selected model's allocation and its own written reasoning (the longest of five reruns).

What this is

The Reasoner scores four bipolar axes of moral-reasoning structure from a constant-sum allocation: for each scenario a respondent assigns proportional weight across the options for a judgment question and, separately, a reasoning question, combined into a position from -1 to +1 per axis. The same scoring runs on people and models, so both land in one space. This viewer holds the full pilot: 11 models across 8 framings on 48 scenarios, and 68 human respondents on the 12 baseline scenarios, 58 of them answering one scenario per axis and ten all twelve. Each model position is an estimate from five reruns per cell; the Compare view can overlay each model's run-to-run spread (±1 SD) to show how far a position moves on rerun. No sampling temperature was sent, so each model ran at its provider's default and that spread is not on a common footing across models.

Population SD. Human baseline re-scored with the same function as the models from raw allocation weights. Analysis is stdlib-reproducible from the raw runs. Methodology was AI-assisted and disclosed.