On: I turned my interview task for Google into a start...

[rimliu]

I wonder what makes you think AI/ML is better.

[Loughla → rimliu]

My same question. If we make terribly biased decisions, why would something with those biases baked in make any better decisions?

Maybe because they're faster?

[MichaelDickens → Loughla]

You train the AI based on outcomes, not based on humans' judgments of what they expect outcomes to be.

[johnhenry → MichaelDickens]

But if your training data is based on previous decisions made using human bias, you're in trouble.

[whatshisface → rimliu]

ML can be used as an unbiased random number generator, unlike humans.

[kbenson → whatshisface]

If it's a random number generator, it's not ML. If it's unbiased, it's also not ML. ML is always biased based on the data you train it with, just as any student is biased based on the information they are educated with.

[whatshisface → kbenson]

If you trained it on data that wasn't useful at all it would be an unbiased random number generator.