Artificial intelligence in lie detection: Why do cognitive theories matter?

This theoretical paper by Tseng and Cheng argues that mechanism-driven, cognitively-informed theories are essential for developing trustworthy AI-based lie detection systems. The authors contend that understanding why algorithms work—not just how accurately they classify—is critical for forensic applications where false positives and negatives have irreversible legal consequences.

Theoretical position paper integrating philosophy of science, cognitive psychology, psychophysiological deception research, and AI ethics to advocate for theory-driven approaches to machine learning in lie detection.

Background & Context

The intersection of artificial intelligence and lie detection has reached a critical juncture. As machine learning algorithms advance rapidly in deception detection research, a fundamental debate has emerged: should researchers prioritize classification accuracy and prediction, or explanatory theories that reveal underlying causal mechanisms? This tension is particularly acute in forensic contexts where courtrooms demand high accuracy, but where false positives and negatives can result in irreversible legal consequences.

In the field of psychological research on deception detection, the rise of machine learning algorithms has sparked discussions about its potential benefits and risks, with most researchers arguing strongly for the inclusion of good theories in the design, training, and application phases of AI. However, the opacity of many AI systems—their "black box" nature—has created challenges for establishing trust, interpretability, and scientific rigor in high-stakes applications.

Tseng and Cheng ask a critical follow-up question: what makes a good theory, and why do they matter in detecting deception? They argue that mechanism-driven, cognitively-informed theories are what AI researchers need to be looking for. This theoretical perspective paper represents an important contribution to the growing literature on explainable AI in forensic psychology.

Research Design & Methodology

This 2025 publication in New Ideas in Psychology is a theoretical position paper rather than an empirical study. The authors do not present original data or conduct experiments. Instead, Tseng and Cheng construct a conceptual argument about the necessary role of cognitive theory in AI-based deception detection systems.

The paper draws on multiple literatures to build its argument:

• Philosophy of science regarding mechanistic explanations (citing Craver & Darden, 2013)
• Debates in psychology about prediction versus explanation (citing Yarkoni & Westfall, 2017)
• Polygraph and CIT research including P300-based detection methods (referencing Gamer et al., Geven et al., Ambach et al.)
• Machine learning applications in multimodal deception detection (referencing Krishnamurthy et al., 2023)

The methodology is argumentative and integrative, synthesizing insights from cognitive neuroscience, psychophysiology, and AI ethics to advocate for a specific approach to developing lie detection algorithms. The paper emphasizes the Concealed Information Test as a paradigm case where theory-driven approaches can enhance AI interpretability.

Results & Key Findings

The paper's central claim is that mechanism-driven cognitive theories are essential for trustworthy AI lie detection. Crucially, mechanism-driven theories allow researchers to know what features are being extracted (e.g., a particular facial expression in time, higher peak in P300) and importantly, how these features are related (subordinate or superordinate) to the entity being inferred (e.g., memory recognition, anxiety, or deception).

Key conceptual contributions include:

• Theory enables error anticipation: Understanding mechanisms helps forensic experts anticipate where the majority of AI mistakes occur (false negatives or false positives), and allows policymakers to adjust decision-making criteria to compensate for such errors via legal or other means.
• Black-box limitations: Even when features extracted by algorithms are known, it remains unclear what the features actually represent without theoretical grounding.
• Adaptive criterion setting: Different legal contexts may require different decision thresholds—more liberal criteria for detecting deception during investigation phases, but conservative criteria in court.
• Cognitive grounding matters: All these logical inferences start from mechanistically and cognitively-informative theories.

The authors emphasize that purely data-driven approaches, no matter how accurate statistically, fail to provide the explanatory depth needed for legal applications where transparency and accountability are paramount.

Discussion & Significance

This paper makes an important contribution to the explainable AI (XAI) movement within forensic psychology. While many AI researchers focus on improving classification metrics, Tseng and Cheng argue that understanding why an algorithm works is as important as how well it works. This perspective directly challenges the trend toward treating AI as a purely engineering problem divorced from psychological theory.

The paper's significance extends beyond technical considerations to legal and ethical dimensions. In courtroom settings, the ability to explain why a system flagged someone as deceptive—and what specific cognitive or physiological mechanisms were involved—can determine whether evidence is admissible and whether defendants receive fair treatment. Without mechanistic understanding, even highly accurate AI systems risk perpetuating bias or producing inexplicable errors.

The emphasis on the Concealed Information Test is strategic. Unlike the Comparison Question Test, the CIT has clearer theoretical foundations in memory recognition and orienting responses. The authors implicitly suggest that CIT-based approaches may be more amenable to theory-driven AI development than other polygraph methods, offering a roadmap for building more interpretable systems.

Limitations & Considerations

As a theoretical position paper, this work does not present empirical data to test its claims. The authors advocate for a particular approach but do not demonstrate that mechanism-driven AI actually outperforms black-box alternatives in real-world deception detection scenarios. The argument is philosophical and conceptual rather than empirical.

The paper also does not fully address the practical challenges of implementing their vision. Building AI systems that incorporate detailed cognitive theories while maintaining high accuracy is technically complex. The authors acknowledge that features and their relationships to deception are often unclear, but provide limited guidance on how to resolve this in practice.

Additionally, the focus on mechanism may underestimate situations where purely predictive models could be valuable, particularly in screening contexts where explanation is less critical than sensitivity. The tension between prediction and explanation may not be resolvable in favor of one approach for all applications.

Practical Applications

For polygraph examiners and forensic psychologists, this paper underscores the importance of understanding the theoretical basis of any AI-assisted detection tools they might adopt. Examiners should demand more than accuracy statistics—they should ask what cognitive or physiological processes the AI is allegedly measuring, and whether those processes have been validated in deception research.

For policymakers and legal professionals, the paper highlights the need for regulatory frameworks that require explainability in forensic AI systems. Standards for admissibility of AI-based lie detection evidence should include requirements for theoretical grounding and mechanistic transparency, not just validation studies showing high accuracy. The adaptive criterion-setting concept suggests that different legal stages (investigation vs. adjudication) may appropriately employ different decision thresholds based on theoretical understanding of error patterns.