AI models do not just learn the world as it is. They push existing patterns even further.

If the training data says most people in tech are men, the model starts treating “man” as the default answer for technical roles. A world that is already unequal becomes even more unequal in the output.

This appears in image generation, AI assistants, recommendation systems, hiring tools, copilots, and many other AI features now entering production systems. Models optimize for the most likely pattern, not necessarily the most representative one.

The difficult part is that these systems can still look successful overall. The outputs feel plausible because they reflect patterns we already accept as normal. But underneath the averages, the same groups of people are repeatedly pushed further away from the “default” answer.

In this session, we will explore how modern AI systems learn and amplify these patterns and why the problem is bigger than “bad data” alone.