Artificial Intelligence in deliberation: The AI penalty and the emergence of a new deliberative divide

Abstract

Advances in Artificial Intelligence (AI) promise help for democratic deliberation, such as processing information, moderating discussion, and fact-checking. But public views of AI's role remain underexplored. Given widespread skepticism, integrating AI into deliberative formats may lower trust and willingness to participate. We report a preregistered within-subjects survey experiment with a representative German sample ( n = 1850) testing how information about AI-facilitated deliberation affects willingness to participate and expected quality. Respondents were randomly assigned to descriptions of identical deliberative tasks facilitated by either AI or humans, enabling causal identification of information effects. Results show a clear AI penalty: participants were less willing to engage in AI-facilitated deliberation and anticipated lower deliberative quality than for human-facilitated formats. The penalty shrank among respondents who perceived greater societal benefits of AI or tended to anthropomorphize it, but grew with higher assessments of AI risk. These findings indicate that AI-facilitated deliberation currently faces substantial public skepticism and may create a new “deliberative divide.” Unlike traditional participation gaps linked to education or demographics, this divide reflects attitudes toward AI. Efforts to realize AI's affordances should directly address these perceptions to offset the penalty and avoid discouraging participation or exacerbating participatory inequalities. • Information on AI-facilitated deliberation reduces willingness to participate. • People expect AI-facilitated deliberation to be lower quality than human-led. • Political interest and need for cognition predict willingness to deliberate. • Pro-AI attitudes raise willingness and quality expectations; AI-risk lowers both. • AI in deliberation creates a new divide beyond usual participation gaps.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 27 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.