Should I Trust You? Detecting Deception in Negotiations using Counterfactual RL

Wichayaporn Wongkamjan, Yanze Wang, Feng Gu, Denis Peskoff, Jonathan K. Kummerfeld, Jonathan May, Jordan Lee Boyd-Graber

Introduction

Negotiation is a high-stakes balancing act of trust and persuasion. In both human and AI-driven interactions, detecting when an offer is "too good to be true" is notoriously difficult. This research introduces CTRL-D (CounTerfactual RL against Deception), a novel framework that unmasks deceptive intent by combining strategic reasoning with natural language processing. By evaluating the strategic incentives for betrayal, CTRL-D detects human lies with 95% precision, significantly outperforming current Large Language Models (LLMs).

The Challenge of Strategic Deception

Deception in negotiations is not merely about the literal truth of a statement, but the strategic intent behind it. In complex environments like the board game Diplomacy—which requires both natural language and deep strategy—humans often fail to spot lies, and AI models frequently struggle.

Existing tools, including state-of-the-art LLMs, often suffer from low precision, flagging nearly half of all messages as deceptive. This "crying wolf" effect makes them impractical for real-time decision support. The goal of this work is to create "strategic friction"—targeted alerts that prompt a user to pause and re-evaluate only when a proposal is truly suspicious.

CTRL-D: Counterfactual RL for Deception Detection

The CTRL-D framework moves beyond simple text analysis by incorporating game-theoretic value functions to evaluate the utility of proposals.

Overview of the CTRL-D pipeline: extracting proposals via AMR and evaluating them using RL value functions to identify Bait, Switch, and Edge signals.

The pipeline operates in three main stages:

  1. Proposal Extraction: Natural language messages are parsed into logical forms using Abstract Meaning Representation (AMR).
  2. Counterfactual Evaluation: The model leverages the value function from CICERO (a top-tier Diplomacy AI) to calculate three "deceptive signs":
  3. Bait: The perceived benefit of the proposal to the recipient.
  4. Switch: The potential loss to the recipient if the sender betrays them.
  5. Edge: The potential gain to the sender from betraying the recipient.
  6. Classification: These strategic features are combined with BERT text embeddings and processed by a neural network to predict the probability of deception.

Experimental Evaluation

The study utilized two primary datasets to test the model's efficacy:

  • Peskov et al. (2020): A collection of 17,289 Diplomacy messages with human-annotated lie labels.
  • Webdiplomacy.net: A massive corpus of 40,000 games (13 million messages) used for large-scale validation.

Results and Discussion

The results highlight a significant gap between strategic RL-based detection and pure language modeling.

Performance comparison: CTRL-D achieves significantly higher precision than LLM baselines and human intuition.

As shown in the performance table, CTRL-D achieved a precision of 0.950 on human lies. In contrast, the LLM baseline using Direct Judgment achieved a precision of only 0.095, flagging over half of messages as deceptive. Human players themselves demonstrated a lie-detection precision of only 0.252.

These findings suggest that in strategic environments, the "truth" of a sentence is less important than its utility. By asking "What happens if they don't do what they said?", CTRL-D captures the incentive for betrayal, which is the core of strategic deception.

Conclusion and Future Work

While highly precise, CTRL-D currently has a lower recall (0.238), primarily because it relies on explicit proposals that can be parsed into logical forms. It may miss implicit deception or vague statements. Future research aims to extend this counterfactual reasoning to more open-ended domains like financial scams or online trading, where the "game rules" are less formal but the incentives for deception remain high.

Summary updated

BibTeX
@inproceedings{wongkamjan2025should,
  author = {Wichayaporn Wongkamjan and Yanze Wang and Feng Gu and Denis Peskoff and Jonathan K. Kummerfeld and Jonathan May and Jordan Lee Boyd-Graber},
  title = {{Should I Trust You? Detecting Deception in Negotiations using Counterfactual RL}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure