E01¶


Research	R0053 — Prompt Claims
Run	2026-03-31-02
Claim	C002
Source	SRC01
Evidence	SRC01-E01
Type	Reported

Negative instructions produce worse LLM output — the Pink Elephant Problem

URL: https://eval.16x.engineer/blog/the-pink-elephant-negative-instructions-llms-effectiveness-analysis

Extract¶

"LLMs seem to produce worse output the more 'DO NOTs' are included in the prompt." The article references Ironic Process Theory (the "pink elephant paradox"), suggesting that telling a system "don't think of X" ironically activates the concept. Real-world examples: Claude Code created duplicate files despite explicit "NEVER create duplicate files" rules. Gemini showed "hit-or-miss" compliance with negative commands.

Relevance to Hypotheses¶

Hypothesis	Relationship	Strength
H1	Contradicts	Directly shows negative constraints are counterproductive
H2	Supports	Confirms enforcement is needed but mechanism is wrong
H3	N/A	Does not address whether AI follows all clear requirements

Context¶

The Ironic Process Theory analogy has limitations — LLMs are statistical models, not human minds. However, the practical observation that negative instructions fail is well-documented across multiple practitioners.