02 June 2025
Sonic the Hedgehog
There is a fundamental paradox in generative AI: the model carries surface-level safeguards (such as against copyright infringements) while remaining deeply bound to its training set. Thus, when the AI tool is asked for Sonic the Hedgehog, the copyright filter kicks in, likely as a barrier normed into the model after training. But when the prompt is reframed in vaguer terms ("a unique blue hedgehog videogame character"), the training-bound impulse takes over, and it averages what it has seen most that can be associated with a blue hedgehog videogame character. The result is unmistakably Sonic the Hedgehog.
In this slippage, we glimpse both the model's fidelity to its training set and its central lack of semantic understanding.