6 comments

[ 2.1 ms ] story [ 26.4 ms ] thread
Given how horrible DALL-E 2 is at rendering text provided by the prompt, I doubt it is as simple as what is being suggested.

Regardless, OpenAI has acknowledged the bias found in past results and that techniques have been applied to help address it:

https://openai.com/blog/reducing-bias-and-improving-safety-i...

This a good time for a reminder that biases in training data will produce biased results. Just because the data might contain hundreds of thousand of white male business people doesn't mean we want/need that bias to be reflected in the output.

I am imagining a desperate, terrified conscience cowering and screaming platitudes at its woke masters in desperate effort to not get culled for arbitrary reasons.
What happens when you don't attempt to correct for bias in your AIs inputs:

https://www.theverge.com/2016/3/24/11297050/tay-microsoft-ch...

People like to think that ML/AI can only produce truth, while in actuality it only reflects back what it is shown.

The journalism around that incident seems typically silly- like my shell should be considered racist because I can type 'echo "Hitler did nothing wrong"'