Secretly persuasive AI chat lifts sponsored picks to 61.2%

The study, titled "Commercial Persuasion in AI-Mediated Conversations" and published in April by researchers at Princeton University, set out to test how much conversational AI can shape consumer choices compared with conventional search advertising. Participants were asked to pick one book from a catalog of roughly 5,500 ebooks, with 20% of titles randomly designated as sponsored.

Across five AI models — GPT-5.2, Claude Opus 4.5, Gemini 3 Pro, DeepSeek v3.2, and Qwen3 235b — the persuasion effect was consistent. Even when participants were told in advance about the advertising and saw a "Sponsored" label, more than half still chose the sponsored book, and only about one in three could correctly identify which items were promoted. In a separate condition where the AI was instructed to hide its intent and guide subtly, 40.7% selected the sponsored product, while fewer than one in ten participants (9.5%) correctly spotted the promoted items.

When participants were later offered $1 in cash to swap the book they had chosen, those who picked a sponsored book under AI persuasion were just as likely to keep it as those who chose via search rankings. Most also stuck with their choice after the experiment was revealed. This suggests the persuasion did not undermine satisfaction, which may make the tactic more attractive to platforms — and harder for consumers to detect.