Simon Willison first-party opinion · 2d ago · 1/5
scosman/pelicans_riding_bicycles ↗
Simon Willison discusses Steve Cosman's project to deliberately introduce synthetic images of pelicans riding bicycles into AI training datasets. Willison characterizes this as a form of training data poisoning, a technique to degrade model performance or introduce unwanted behaviors. He notes that similar synthetic data injection efforts have been published previously, framing the practice as commentary on AI training data integrity.
I firmly approve of Steve Cosman's efforts to pollute the training set of pelicans riding bicycles.