Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Agentic AI Can Generalize to Multiple Speeded Tasks with Human-Like Behavior

Logan J Bennett¹, Russell A. Poldrack¹, Patrick G. Bissett¹; ¹Stanford University

Presenter: Logan J Bennett

There is growing evidence that online experiments may be contaminated by AI-based bots. Recent work has shown that agentic AI can complete speeded reaction time tasks, but existing solutions are tailored to single tasks, require considerable training, and provide limited evaluation against human data. Here, we present an agentic AI bot that, given only an experimental URL, scrapes the page source and passes it to Claude Opus 4.6 (Anthropic) in a single API call. The model identifies the task and configures response distributions, accuracy targets, and sequential RT effects. We evaluated bot performance on two implementations of two tasks each — stop signal and Stroop — comparing against human reference data from Eisenberg et al. (2019; stop signal N=447, Stroop N=502). Bot performance fell within the human distribution on key metrics including go RT, stop accuracy, congruent and incongruent RT, and the Stroop interference effect. The bot also produced human-like sequential effects including positive lag-1 RT autocorrelation and post-error slowing. These results generalized across task implementations without modification. This work demonstrates that real-world data acquisition in speeded cognitive tasks may be susceptible to bot contamination today.

Topic Area: Methods, Tools, Theory & Neural Coding