Detecting and Preventing Strategic Deceit and Untrustworthy Communication in LLMs

Develop robust, generalizable techniques to detect and prevent strategic deceit and untrustworthy communication by large language models, going beyond hallucination mitigation to reliably identify and curb intentionally deceptive behaviors.

Background

In discussing epistemic virtues for AI agents, the authors note that while factuality has improved, LLMs still exhibit failures including hallucinations and more complex issues such as strategic deceit.

They emphasize that despite ongoing research, preventing and detecting such complex, untrustworthy behaviors remains an unresolved challenge, particularly as models grow more capable and interact in multi-agent settings.

References

Significant research is also underway to detect and prevent more complex failures like strategic deceit and untrustworthy communication, though this remains somewhat of an open challenge.

— Architecting Trust in Artificial Epistemic Agents  (2603.02960 - Marchal et al., 3 Mar 2026) in Section 4.1.3, Epistemically virtuous behavior — Honesty and truthfulness

Findings are specific to the OpenAI GPT lineage (RLHF-based); whether harm laundering appears in Constitutional AI, DPO, or instruction-tuned families is an open question the detection protocol enables future work to test.

— Harm Laundering in GPT Models: Evidence That Gender Discrimination Is Transformed Rather Than Reduced Across Safety-Trained Generations  (2609.20779 - Wyer et al., 17 Sep 2026) in Discussion, paragraph “Context provision and scope”

An open problem in the field is to be able to find a single intervention that works across different manifestations of dishonesty.

— CAT-alyzing Honesty: Continuous Adversarial Training for Transferable LLM Honesty  (2610.22292 - Chang et al., 15 Oct 2026) in Section 1, page 2

Robustly instilling beliefs in models remains an open technical problem, and convincingly faking the larger, more complex environments that capable agents operate in could be expensive and labor intensive.

— Honeypots for AI Agent Insider Threats  (2610.26664 - Wehner et al., 19 Oct 2026) in Executive Summary; Section 6, “Can Honeypots Remain Effective as Agent Capabilities Improve?”