> It's not the AI's inherent bias as such though. It's the bias of the dataset...

An engineer builds a chat protocol and app on top of UDP. Chat messages between users frequently never arrive and are lost.

When questioned about it, the engineer says "it's not my app that causing the message loss it's UDP that's causing it."

The engineer selected the communication protocol.

To avoid packet loss, I build retry logic into my app, or build on top of TCP.

> To avoid packet loss, I build retry logic into my app, or build on top of TCP.

What's the "I used TCP instead of UDP" equivalent of a dataset that can be used to build a perfectly non-biased translation service?

Your comment seems to suggest that everybody seems to deliberately choose the wrong datasets which result in racist and sexist products. What should we use instead then?

If the underlying layer isn't reliable, build a layer on top to make it reliable.

If the underlying dataset has issues, build a layer on top to address those issues.

So AI is taking over the world, and we should just accept that it has severe limitations of just parroting the training data like a Markov chain? This is the algorithm we think can solve everything?

> So AI is taking over the world, and we should just accept that it has severe limitations of just parroting the training data like a Markov chain? This is the algorithm we think can solve everything?

"No", to both questions.

These systems are a work-in-progress, constantly improving. They have value, even though they are far from perfect.