Lemma, a short statement you believe to be true
- No names, respond to the idea not the person
- Stay on topic, the lemma stays in the center
- Rate other lemmas, highest rise to the top
- Tags, follow what you care about
We don't want AI to accelerate or create more discrimination than already exists, particularly when AI could be used much more quickly and cheaply and therefore impact a significantly greater number of people.
Its actually quite difficult to determine an algorithms bias, you need a statistically significant sample. Only 8% chance of 50 head and 50 tails when flipping a fair coin, 92% of the time it will seem to have bias.
It is very difficult to create AI (especially machine learning algorithms) that performs equally well across every possible way you could group a population, you'd effectively be saying you could never use AI.
For machine learning algorithms, a general rule is the more data you train on the better the performance. If you train the model on a representative sample, it is likely to be less effective on the minority groups.
If an algorithm performs better for all groups than a human (for example reading medical scans), are we not going to use it because one groups results is higher than the other? eg humans 90% accurate, AI for women 95% and AI for men 98%?