Thanks for taking time to write this explanation. Note that the bias you mentioned about your experience with GANs is not what I would call systematic bias. It

Thanks for taking time to write this explanation. Note that the bias you mentioned about your experience with GANs is not what I would call systematic bias. It is random noise. And it may even be racist in nature. Yes, you may keep on being biased in your own way day in and day out, which makes it systematic for you, but we won't need an entire field of ethical AI to study that. The example you gave about credit score, and the theme generally, is very well understood for a very long time. That's something we teach to stats undergrads early on. That's just one source of bias in the data, btw. There are numerous other data biases and tons of research exists on that. Even on one method of data collection, such as survey data, there is a huge literature on reducing the bias. > you'd have to work really hard to make an inherently racist architecture. That's precisely where my confusion comes from. If someone makes this racist architecture, it is not a systematic bias that anyone cares about. It's a rogue model. In fact, in many cases, this could be outright illegal. For example, if someone designs a credit scoring model to deliberately leave out one group of people based on their gender, ethnicity, or religion, that's not going to fly with regulators. So anyway, I still don't have an example of a systematic bias that is not caused by data.