37C3 - Self-cannibalizing AI

Discover the dangers of biases in AI, particularly in text-to-image generation models, and learn how self-cannibalizing AI can perpetuate harmful stereotypes and reinforce existing biases.

Key takeaways

The talk highlights the dangers of biases in AI, specifically in the context of text-to-image generation models.
The speaker presents a self-cannibalizing AI example, where an image generation model is trained to produce aesthetically pleasing images, resulting in it predicting a “happy” face with a high similarity score, even when there is no actual correlation between the image and the keyword.
The AI model is not intelligent in the classical sense, but rather uses mathematical patterns to generate images based on the input it receives.
The speaker argues that the AI’s ability to generate images that it considers “happy” is problematic, as it may perpetuate biases and reinforce already existing harmful stereotypes.
The talk also discusses the concept of model collapse, where the AI model becomes stuck in a loop, repeatedly generating the same image, and the importance of understanding how these models work to mitigate biases.
The speaker presents several examples of generated images and discusses how the AI model uses mathematical patterns to generate images that are aesthetically pleasing, but may not necessarily reflect reality.
The talk concludes by emphasizing the importance of understanding how these models work and the potential risks of biases in AI.

37C3 - Self-cannibalizing AI

More talks