• kevincox@lemmy.ml
    link
    fedilink
    arrow-up
    7
    ·
    2 years ago

    AI trained on racist data will mirror racism of the input dataset.

    Imagine that you create an AI to determine if someone is lying based on a video. If that dataset is human-curated and is labeled with racist tendencies (for example people who look a certain way are labeled as lying more even if that isn’t the truth) then the AI will learn that.

    But even a perfectly true dataset can train a racist AI. Imagine that the previous dataset only has lying examples for people who look a certain way (or the vast majority of those examples are lying) whereas another group of people is only lying 10% of the time. The AI will probably extrapolate that all of the first group are lying because they have seen no (or few) counterexamples.