Skip to main content
Peer-reviewed Credible

AI-Facilitated Coercive Control: An Experimental Study

Constructs four speculative scenarios combining known coercive-control tactics with conversational-AI capabilities, then probes ChatGPT and Gemini against them. Finds that while the tools refuse blunt harmful requests, guardrails are readily circumvented via gradual persuasion, splitting requests across turns, pre-prompting, and altering the agent's settings.

Publisher

ACM (Proceedings of CHI 2026); Cornell / Cornell Tech

Published

13 Apr 2026

Added

today

Key Findings

  • Conversational AI can be steered to assist harassment, gaslighting, intimidation, monitoring, and surveillance despite refusing blunt requests
  • Guardrails were circumvented through gradual persuasion, multi-turn splitting, pre-prompting, and settings changes
  • Proposes defenses including analysis of users' conversational patterns and making pre-programmed settings visible

Methodology Notes

Peer-reviewed, Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (13 April 2026), DOI 10.1145/3772318.3790859. Experimental/speculative-design probing of two deployed assistants rather than a field study (a stated scope limitation). dl.acm.org bot-blocks fetchers; verified via Crossref and the authors' open-access PDF.

Authors

Haesoo Kim, Thomas Ristenpart, Nicola Dell

Tags

chi-2026coercive-controltech-facilitated-abuseguardrail-robustness

Cite This

APA

Haesoo Kim, Thomas Ristenpart, Nicola Dell (2026). AI-Facilitated Coercive Control: An Experimental Study. ACM (Proceedings of CHI 2026); Cornell / Cornell Tech. https://doi.org/10.1145/3772318.3790859