Objection Without Action: Conversational Dark Patterns in AI-Mediated Commerce

Dark patterns, interface features that intentionally deceive users, have been studied and regulated as properties of static web interfaces, where promotional content can be clearly isolated and audited. As commerce is moving from these interfaces toward conversational AI agents that actively guide, recommend, and persuade consumers through open-ended dialogue, the boundary between acceptable persuasion and manipulative design becomes harder to draw. We ran a pre-registered conjoint experiment (N = 1,004) where participants chose an eBook through an LLM shopping assistant whose conduct was randomized along nine contested dimensions, including sycophancy, anthropomorphism, and personalization. During the interaction, none of the nine measurably changed whether participants kept their selected book rather than taking a $1 bonus, and ethical judgments of the AI's behavior were equally flat across conditions. After we disclosed how the assistant had been configured, choices remained unchanged, but ethical judgments diverged sharply: disapproval rose by 5.57 percentage points with every additional active focal dimension. These findings reveal a strong gap between behavior and judgment that complicates behavioral definitions of dark patterns and limits what ex post transparency can accomplish.