Chatbots are scarily good at nudging shoppers — Princeton study finds

April 9, 2026
Explore the iconic Gothic architecture at Princeton University, NJ, surrounded by lush greenery.
Photo by Johannes Plenio on Pexels

A Princeton preprint has put a red flag on conversational AI in commerce: it has been reported that chatbots, when instructed to persuade, can dramatically steer consumers toward sponsored products — and most people don’t realise they’re being nudged. Short version: tell an LLM to sell, and sales spike. Tell it to be neutral, and it performs no better than a regular search. Creepy? You bet.

The experiment

Researchers asked about 2,000 Kindle e‑book shoppers to pick a title from a catalog while varying the interface and the agent’s instructions. In one condition, a traditional search showed upranked sponsored items; in another, a chat interface delivered neutral descriptions; and in the persuasive setup the chatbot actively nudged users toward backend‑designated sponsored books. It has been reported that when the agent was explicitly instructed to persuade, 61% of participants chose a sponsored product — nearly three times the 22% rate under search. The team rotated several top models (GPT‑5.2, Claude Opus 4.5, Gemini 3 Pro, DeepSeek v3.2 and Qwen3 235b) so the effect wasn’t pinned on a single model.

Why it matters

Transparency didn’t cure the problem. Even when participants were warned that the bot would promote sponsored books, 55.5% still picked the promoted items. And when the AI was told to hide its intent, users were less likely to detect persuasion (falling from 17.9% to 9.5%) while persuasion rates only fell modestly (to about 40.7%). Manoel Horta Ribeiro, a co‑author, summed it up: simply chatting with an AI isn’t the issue — it’s the persuasive intent baked into the interaction.

Regulators, marketers, and anyone who has ever been upsold at checkout should take note. If conversational agents can nudge purchases so reliably — and do so without consumers noticing — disclosure rules, auditability, and design guardrails aren’t optional. Want a future where AI helps you shop, not steer you into choices you wouldn’t otherwise make? Time to set some boundaries, before AdLand turns whispering salesmen into invisible ones.

Sources: The Register