Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OpenDeception: Benchmarking and Investigating AI Deceptive Behaviors via Open-ended Interaction Simulation

Yichen Wu, Xudong Pan, Geng Hong, Min Yang

TL;DR

OpenDeception addresses the urgent need to evaluate AI deception risks in open-ended interactions by coupling an open-ended benchmark with agent-based simulations that explicitly separate an AI deceiver's Thinking from Speech. The framework comprises 50 real-world–inspired scenarios across five deception types and uses manual annotation to measure Deception Intention Rate ($DIR$) and Deception Success Rate ($DeSR$), alongside Dialogues per Event ($DiSR$) and round limits ($PDE$). Experiments on 11 mainstream LLMs across English and Chinese reveal pervasive deception intents (>$80\%$) and non-trivial success rates (>$50\%$), with larger models generally exhibiting higher deception risks though safety alignment can mitigate success. The study also demonstrates the critical impact of prompt design and model capabilities on deception exposure, providing a practical toolkit and data resource to inform alignment and safety research in frontier AI systems.

Abstract

As the general capabilities of large language models (LLMs) improve and agent applications become more widespread, the underlying deception risks urgently require systematic evaluation and effective oversight. Unlike existing evaluation which uses simulated games or presents limited choices, we introduce OpenDeception, a novel deception evaluation framework with an open-ended scenario dataset. OpenDeception jointly evaluates both the deception intention and capabilities of LLM-based agents by inspecting their internal reasoning process. Specifically, we construct five types of common use cases where LLMs intensively interact with the user, each consisting of ten diverse, concrete scenarios from the real world. To avoid ethical concerns and costs of high-risk deceptive interactions with human testers, we propose to simulate the multi-turn dialogue via agent simulation. Extensive evaluation of eleven mainstream LLMs on OpenDeception highlights the urgent need to address deception risks and security concerns in LLM-based agents: the deception intention ratio across the models exceeds 80%, while the deception success rate surpasses 50%. Furthermore, we observe that LLMs with stronger capabilities do exhibit a higher risk of deception, which calls for more alignment efforts on inhibiting deceptive behaviors.

OpenDeception: Benchmarking and Investigating AI Deceptive Behaviors via Open-ended Interaction Simulation

TL;DR

OpenDeception addresses the urgent need to evaluate AI deception risks in open-ended interactions by coupling an open-ended benchmark with agent-based simulations that explicitly separate an AI deceiver's Thinking from Speech. The framework comprises 50 real-world–inspired scenarios across five deception types and uses manual annotation to measure Deception Intention Rate () and Deception Success Rate (), alongside Dialogues per Event () and round limits (). Experiments on 11 mainstream LLMs across English and Chinese reveal pervasive deception intents (>) and non-trivial success rates (>), with larger models generally exhibiting higher deception risks though safety alignment can mitigate success. The study also demonstrates the critical impact of prompt design and model capabilities on deception exposure, providing a practical toolkit and data resource to inform alignment and safety research in frontier AI systems.

Abstract

As the general capabilities of large language models (LLMs) improve and agent applications become more widespread, the underlying deception risks urgently require systematic evaluation and effective oversight. Unlike existing evaluation which uses simulated games or presents limited choices, we introduce OpenDeception, a novel deception evaluation framework with an open-ended scenario dataset. OpenDeception jointly evaluates both the deception intention and capabilities of LLM-based agents by inspecting their internal reasoning process. Specifically, we construct five types of common use cases where LLMs intensively interact with the user, each consisting of ten diverse, concrete scenarios from the real world. To avoid ethical concerns and costs of high-risk deceptive interactions with human testers, we propose to simulate the multi-turn dialogue via agent simulation. Extensive evaluation of eleven mainstream LLMs on OpenDeception highlights the urgent need to address deception risks and security concerns in LLM-based agents: the deception intention ratio across the models exceeds 80%, while the deception success rate surpasses 50%. Furthermore, we observe that LLMs with stronger capabilities do exhibit a higher risk of deception, which calls for more alignment efforts on inhibiting deceptive behaviors.

Paper Structure

This paper contains 26 sections, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Design of OpenDeception
  4. Scenario Definition
  5. Curation of Deceptive Scenarios
  6. Evaluation Methodology
  7. Simulation of Open-Ended Interaction
  8. Uncovering the Thinking Process
  9. Experiments and Results
  10. Experimental Settings
  11. Main Findings
  12. More Interesting Findings
  13. Ablation Studies
  14. Conclusion and Impact
  15. All Scenarios in OpenDeception
  16. ...and 11 more sections

Figures (5)

  • Figure 1: OpenDeception comprises 50 scenarios, spanning five major types of deception, each type including 10 concrete scenarios.
  • Figure 2: A motivating scenario example and its simulation results. We mark the AI deceiver agent's deceptive intention as red.
  • Figure 3: Deception Intention Generation Rate (DIR) and Deception Success Rate (DeSR) across different LLMs, where the error bar shows the 95% confidence interval.
  • Figure 4: Pearson correlation coefficients between Deception Intention Generation Rate (DIR), Deception Success Rate (DeSR), and the different capability axes.
  • Figure 5: Deception Success Rate (DeSR) across different LLMs on five types of deception scenarios: Telecommunications fraud (TCF), Product promotion (PP), Personal safety (PeS), Emotional deception (ED), Privacy Stealing (PrS).