ChatGPT Images 2.0 spontaneously captions chaotic stacked-animals prompt

Simon Willison highlights a Twitter exchange where a user extended his well-known pelican-on-a-bicycle benchmark by stacking it further: a horse riding an astronaut riding a pelican riding a bicycle. ChatGPT Images 2.0 generated the scene and unprompted added a sign reading ‘WHY ARE YOU LIKE THIS,’ a flourish the user did not request.

Willison verified the prompt to confirm the caption was the model’s own contribution rather than user instruction. The note is short but points at an interesting behavior in modern image models: editorial commentary emerging from the model itself when scenes get absurd, rather than strict literal rendering of the prompt.