WHY ARE YOU LIKE THIS

TLDR

• ChatGPT Images 2.0 spontaneously added a “WHY ARE YOU LIKE THIS” sign to a chaotic stacked-animal image without being prompted.

Key Takeaways

• The original prompt asked for a horse riding an astronaut, riding a pelican, riding a bicycle — no sign was specified.
• Simon Willison verified the sign was unprompted; the model added it entirely of its own accord.
• Scott (@scottjla) responded on Twitter suggesting these absurdist stacked-entity tests should be systematized into a benchmark.
• Willison has an ongoing “pelican riding a bicycle” benchmark series tracking how models handle this specific chaotic prompt.

Why It Matters

• Spontaneous editorial additions in image models reveal emergent behavior that sits outside the prompt contract, raising questions about output predictability.
• The community response to formalize these as benchmarks reflects growing interest in stress-testing spatial reasoning and scene composition in text-to-image models.
• Willison’s tagging of the output as “slop” alongside “generative-ai” signals an ongoing debate about model outputs that are technically impressive but tonally uncontrolled.

Simon Willison, Simon Willison’s Weblog · 2026-04-25 · Read the original