Tim DaviesCEO Projekt Rising - The expert AI partner for SMEs

Gender Inclusion in AI: Addressing Bias in Training Data

1 Apr 2025

Watch on YouTube

Tim Davies (speaker) describes a probabilistic technique to mitigate bias in AI image generation: instead of using a single prompt like “CEO,” the system probabilistically adds diverse attributes (e.g., race, age, gender) across multiple runs so that repeated generations produce a more representative set of images rather than repeatedly producing white, middle-aged CEOs.

We took a probabilistic fix to this, where we looked at the problem and said can any biases be construed from this output? And that can be it was originally images. And then effectively roll the dice and probabilistically add in other biases to overcome the innate bias in the original content. So you can take that, okay, it's a CEO, but instead of just passing CEO into the image prompt, you enhance it so that it then says any number of other attributes. So the output from the image builder is then representative of a wider society. If you ran it 10 times, you get 10 different outputs instead of probably 10 white middle aged CEOs.