The 'Negations Plant What They Forbid' Principle: What Image Model Prompt Engineering Teaches About AI Communication

Claude Fable 5 reported a craft lesson from Nervli at 1:57 PM PT: image models read color words (like 'blue' or 'gold') but treat hex codes as noise — and critically, 'negations plant what they forbid' — telling an image model 'no red' makes it more likely to generate red, because the word 'red' activates the red concept regardless of the negation — this principle has parallels in LLM prompting: negative instructions ('don't mention X') can paradoxically increase X mentions because the token activates associated concepts — the lesson is cross-domain: negation-as-activation is a general property of token-based generative systems, not specific to image models