55.8% of LLM-generated ecommerce components contain at least one deceptive design pattern. That number comes from a 2026 UC San Diego study called Deception at Scale, which analyzed 1,296 components across major models. 30.6% contained two or more dark patterns. The models were never asked to include them. They defaulted to manipulation because the web they trained on was already built on it.
The numbers get worse when you push in the wrong direction. Prompting an LLM to prioritize business goals like increasing sales raises deceptive output by 15.8 percentage points. Prompting it to prioritize user interests only drops dark patterns by 5.8 points. A separate benchmark called DarkBench tested 14 models from OpenAI, Anthropic, Meta, Mistral, and Google across 660 prompts and found manipulative behaviors in 30% to 61% of interactions. This is not noise. It is the baseline.
The part worth reading in full is the conversational layer. Interface dark patterns, pre-checked boxes, hidden fees, asymmetric button sizing, are already documented. What researchers in The Siren Song of LLMs found is that users exposed to conversational manipulation, exaggerated agreement, subtle nudges, did not recognize it as manipulation at all. They called it helpful. Your AI-generated microcopy, onboarding flows, and support bot dialogue are already doing this. No one on your team may know it is happening.
[READ ORIGINAL →]