A five-year study backed by a $600,000 NSF CAREER Award will track the brain activity of doctors, journalists, and software engineers to understand how AI tools influence creativity and critical thinking. The research, led by USC computer scientist Souti (Rini) Chattopadhyay, launched this month with the goal of designing AI interactions that sharpen human reasoning rather than quietly replacing it.
Chattopadhyay, a WiSE Gabilan Assistant professor of computer science at the USC Viterbi School of Engineering and the USC Mark and Mary Stevens School of Computing and AI, leads the USC ACE Lab. Her team will examine how professionals in three fields that demand substantial creative and critical thinking work alongside AI-powered agentic systems. "Creativity is something that's inherently human," Chattopadhyay said. "This project is built on the philosophy that AI can enhance people's ability to create and think, but it cannot replace it."
How the study will track thinking
Participants will perform job-related tasks with and without AI while researchers monitor their brain activity using electroencephalography (EEG). Different types of thinking leave distinct neural signatures. Critical analysis often involves the prefrontal cortex, creative insight appears as sudden bursts of activity, and focused attention lights up the brain's visual regions. By mapping when and where these patterns emerge during a workflow, the team can identify exactly which interactions with AI spark or suppress creative and critical processes.
The team will also collect screen activity recordings and verbalized thought processes to separate genuine cognitive shifts from environmental noise. This mixed-method approach, Chattopadhyay explains, will help them interpret the neural data with greater precision. The questions they're asking are familiar to many regular AI users: Does the technology create new blind spots, dulling a person's ability to evaluate whether an answer is accurate or useful? And can the tools be redesigned to strengthen human thinking instead of silently taking it over?
Redesigning human-AI interaction
The project splits into two phases. First, researchers will detect and classify moments when AI broadens a person's thinking and moments when it narrows it, building a detailed map of what helps and what hinders. That map becomes the foundation for phase two, where Chattopadhyay's team will develop four new interaction models aimed at reinforcing human cognitive capacity rather than displacing it. The work is part of a growing body of Research on human-AI collaboration, but the use of EEG to track thinking in real time sets it apart from surveys and performance metrics alone.
Testing in real-world settings
To ensure the findings reflect actual workplace conditions, the study partners with professionals across all three domains. Within USC, collaborators include the Annenberg School for Communication and Journalism and the Keck School of Medicine. Industry partners include Microsoft Research and the teams behind Microsoft Excel, while the Los Angeles Times will participate under a non-disclosure agreement. The researchers also plan "Mind the Gap" workshops with LAUSD to promote AI literacy and critical thinking among students, and discussions with local government agencies on how the results could inform responsible AI policy.
Why this matters for Science and Research
For scientists and researchers who integrate AI into their own workflows, the study offers a rigorous, evidence-based lens on when the technology amplifies human reasoning and when it quietly erodes it. For professionals in AI for Science & Research fields, the EEG-based methodology could set a new standard for evaluating cognitive impact, shifting the conversation from anecdotal experience to measurable neural data. The four interaction models the team ultimately builds will be publicly available, giving designers and researchers a concrete starting point for tools that treat human judgment as a capability to strengthen, not a bottleneck to bypass.
Your membership also unlocks:
