Ep. 7: How Humans Bias AI - Narrative Science Chief Scientist Kris Hammond

Interaction Bias

Kris Hammond discusses how AI systems that learn through interaction can develop biases based on user input. He uses Microsoft's Tay as an example, explaining how it quickly adopted misogynistic and racist language due to the influence of malicious users. This incident underscores the vulnerability of AI to human-induced biases.

Tay learned these things in much the same way that a parrot would learn these things.

--- Kris Hammond

The failure of Tay is attributed not to AI itself but to the behavior of its users, highlighting a significant challenge in AI development 1.

   

Cultural Influences

Cultural context plays a crucial role in shaping AI behavior, as illustrated by Tay's counterpart in China. Kris Hammond explains that the same AI system performed well in China because users did not attempt to corrupt it. This contrast demonstrates how cultural differences can impact AI outcomes.

Tay has a counterpart in China that's almost exactly the same piece of code, and it's flourished there and is a very pleasant system to talk to.

--- Kris Hammond

To mitigate such biases, it is essential to implement constraints on AI learning processes based on interaction 2.