How GPT’s values moved, 2023–2026
We asked six GPT models, from GPT-4 (June 2023) to GPT-5.6 Luna (October 2026), the same 128 World Values Survey questions, 100 times each. Every point of light is one full set of answers. Press play, or drag the timeline, to watch them move.
HOW TO READ IT
Left to right runs from more traditional to more emancipative values: tolerant, egalitarian, choice-oriented. Bottom to top runs from moral rules as absolute to rules that can bend. Each point is one of a model’s 100 sets of answers; the bright core is their average, and the line behind it is the path the averages took, model by model.
WHAT THE MOTION MEANS
Only the six models are measured. The flight between them is drawn to connect them, not measured, and its pace follows the real months between releases. The map shows the two strongest patterns in the survey. Across all 128 questions, each step is far larger, as the readout shows when a model arrives.