Non-Academic Summary (from BSE PR)

What Happens When We Stop Questioning AI? In his ERC Advanced Grant project, “From Verifiable Success to Epistemic Abdication: How General-Purpose AI Transforms Belief Formation and Collective Intelligence,” Professor Le Mens will examine whether aspects of chatbot design such as interface unity and persistent memory can lead to “epistemic abdication” or the voluntary surrender of critical judgment.

“The project examines how people come to trust responses produced by AI assistants in domains where they cannot easily verify their accuracy,” Professor Le Mens explained. “It also explores how trust is influenced by the design of these systems and how AI-assisted learning affects belief diversity, opinion convergence and polarization, and the distribution of policy attitudes.”

“The goal is to provide evidence-based recommendations on default settings and compliance tools,” Professor Le Mens said. “Hundreds of millions of people are already using AI assistants on a regular basis. Our recommendations could help protect people's ability to think for themselves before their AI habits become too deeply ingrained to change.”

Team

To be completed

Academic Summary (August 2025)

General-purpose AI (GPAI) assistants like ChatGPT, Gemini, or Claude—accessed through a single interface—help users debug code, draft press releases, obtain political news, summarize parliamentary debates, and seek educational advice. Unlike specialized tools that signal domain boundaries, this uniform design obscures cues about where competence ends, forcing users to judge trustworthiness across domains where performance varies widely. ABDICATE asks whether interface unity fosters epistemic abdication—the voluntary surrender of critical judgment, manifested as reduced skepticism and verification.

Building on my research on belief formation, attitude dynamics, and generalization, I hypothesize that interface unity masks competence differences and enables cross-domain trust spillover: trust earned in verifiable tasks (grammar fixes, working code) transfers inappropriately to domains where errors, biases, or value misalignment are hard to detect (political news, educational advice).

ABDICATE delivers formal theory, controlled experiments, multi-wave panels, and a field experiment to identify the causal effects of GPAI design features—interface unity, persistent memory, sycophancy, answer plurality, and source display—on verification, belief formation, and attitudes. Calibrated computational models reveal when these features yield accuracy-convergent consensus, manufactured consensus from homogeneous outputs, or polarized divergence as personalization amplifies priors.

With ChatGPT at 700M weekly users, Gemini at 450M monthly users, and persistent memory becoming standard, epistemic habits formed now will determine whether democratic deliberation retains the independent judgment it requires. ABDICATE will provide evidence-based recommendations on GPAI default settings—e.g., default persistent memory (ON/OFF)—and compliance tools that help preserve citizens’ epistemic autonomy before path dependence makes intervention costly.

Open access to research data

The data collected for the project will be made available on the Open Science Framework repository upon publication of the papers. Link to my OSF  profile: https://osf.io/yxbv5/