Bias as a consequence of composition, not malice. Adjust a simplified training mix and watch the model's default and its likelihoods move with the data.
Live preview · launch for the interactive version
Set how often each category appears in the data the model learns from — the inputs, made adjustable.
As the mix changes, the most-likely output changes with it. The default follows the majority.
Nothing is hand-coded. The skew in the data simply becomes the skew in the behavior.
Predict which output becomes the default if one category dominates.
Push one category from balanced to dominant; hold the rest still.
How did the default and the ranked likelihoods move?
Not 'it's biased' — say how: composition drove the default.