Manuel Cherep
Graduate Student, Fluid Interfaces Group
MIT Media Lab
Manuel Cherep is a PhD student in the Fluid Interfaces group at the MIT Media Lab, where he studies how AI agents make decisions and where their choices drift from human ones. He is the lead author of ABxLab, an ICLR 2026 study that sent 17 AI agents shopping and found that prices, ratings and persuasive nudges moved their picks far more than they move people's.
Testing Whether AI Agents Shop Like People
"Many applications of AI agents tacitly assume that, under uncertainty, they will react in roughly human-like ways, if not more rationally," Cherep told PsyPost. "Instead of accepting this assumption, we decided to explore how agents behave when choices are presented to them in different ways." In a paper published in PNAS in June 2026 with Pattie Maes and Nikhil Singh, he adapted a human decision task and tested leading language models against defaults, suggestions, highlighted information and nudges derived from a model of human choice. The models reacted to weak cues far more strongly than people did, and neither chain-of-thought prompting nor examples of human behavior reliably steadied them.
Cherep separates these cues from deliberate attacks. "The fundamental difference is that nudges are part of everyday life for decision-makers," he said. "While adversarial attacks can potentially be detected and removed, nudges will always exist."
Prices, Ratings and Scarcity Tags
ABxLab, written with Chengtian Ma, Abigail Xu, Maya Shaked, Maes and Singh, moved the question into a realistic web store. Across more than 80,000 experiments, the team varied the order, price and rating of products and added persuasive lines drawn from consumer psychology: authority cues, social proof, scarcity, negative framing and incentives. Even when two products were otherwise identical, those lines shifted what the agents picked, and a stated user preference such as a tight budget worked almost like a hard rule. People shown the same pages were far less sensitive. The framework is released as an open benchmark for testing how agents choose.
For anyone handing purchases to an assistant, Cherep sees a clear risk. "This sensitivity can be exploited by a third party to influence the agents you delegate to, leading to decisions that you might not have made otherwise," he said.
From Madrid and Lausanne to the Media Lab
Cherep earned a bachelor's degree in mathematics and computer science from Universidad Politécnica de Madrid and a master's degree in data science from EPFL. Before his PhD he spent nearly six years in industry, as a full-stack software engineer and as an AI engineer at Logitech. He has held fellowships from Fulbright, "la Caixa" and the Rafael del Pino Foundation, and his work has been supported by Amazon and SK Telecom through the MIT Generative AI Impact Consortium. Alongside agent behavior, he works on audio machine learning. "Our long-term goal is to develop a behavioral science of AI agents," he said.