AI & Digital Forensics - Review of Fumagalliʼs (2026) Ethical challenges of AI-based deception detection...

Fumagalli (2026) Ethical challenges of AI-based deception detection in forensic settings

 The purpose of this report is to review recommended guidelines for digital forensics examiners considering the use of artificial intelligence (AI) in forensic decision-making based on Fumagalliʼs article, Ethical challenges of AI-based deception detection in forensic settings. The author identifies concerns about “the validity, transparency, contextual sensitivity, and fairness of AI-assisted deception detection systems, particularly in high-stakes forensic contexts” (Fumagalli, 2026, p. 18).

Scope of Fumagalli's Work

The scope of the source is AI-assisted deception detection and associated forensic assessment— not digital forensic examinations generally. Accordingly, the guidelines below are confined to the articleʼs recommendations concerning interpretation, validation, oversight, bias, and accountability. They should not be read as source-supported procedures for other digital forensic tasks. The articleʼs governing recommendation is caution: “Until stronger theoretical foundations and transparent methodologies are established, caution rather than technological optimism should guide its use” (p. 22). Page numbers in the citations refer to the articleʼs printed journal pages, rather than the PDFʼs page sequence.

Fumagalli's Recommendations

The following guidelines present the sourceʼs recommendations as considerations for digital forensics examiners using AI-assisted deception detection or its outputs.

• Retain human judgment, oversight, and critical evaluation. “Practitioners should operate in accordance with super alignment principles, ensuring that human judgement, oversight, and critical evaluation remain central” (p. 21).

• Treat AI outputs as supplementary information, not determinations of truth. “AI outputs should be treated as supplementary information rather than objective ʻtruth detectionʼ tools” (p. 21).

• Require independent validation before implementation. “Given the consequences of false positives and negatives, AI-assisted tools should undergo independent validation under ecologically realistic forensic conditions before implementation” (p. 21).

• Prioritize theoretically grounded approaches and interpretable reasoning. “Practitioners should prioritise theoretically grounded approaches, as legal and clinical decisions require interpretable reasoning rather than prediction alone” (p. 21).

• Do not adopt a method solely because it appears accurate. “In forensic assessment and credibility evaluation, apparent accuracy is insufficient justification for adopting a method” (p. 18).

• Require traceable conclusions, transparent methods, and reasoning open to scrutiny. The author identifies “core requirements of forensic practice: that conclusions be traceable, that methods be transparent, and that the reasoning underpinning an assessment be capable of withstanding external scrutiny” (p. 19).

• Train practitioners in both AI capabilities and limitations. “Due to risks of automation bias and over-reliance, practitioners must be trained in both the capabilities and limitations of AI systems” (p. 21).

• Apply cultural and contextual sensitivity when interpreting outputs. Practitioners should also apply cultural and contextual sensitivity, recognising that narrow training data may misclassify vulnerability-related behaviours as deception, particularly among neurodivergent individuals, trauma survivors, or those experiencing distress. (p. 21)

• Address dataset bias without assuming that it can be eliminated completely. “Although bias can be reduced through careful sampling, dataset balancing, and transparency, the literature suggests that completely unbiased datasets are extremely difficult to achieve” (p. 19).

• Maintain the same interpretability and accountability standards applied to other forensic assessments. “Forensic practitioners, researchers, and policymakers must hold AI deception detection to the same standards of interpretability and accountability that govern all other forms of forensic assessment” (p. 21).

Issues and Challenges

The issues below are those identified by Fumagalli in AI-assisted deception detection and forensic investigations. They are not findings about every application of AI in digital investigations.

• Limited real-world validity. “Although some studies report high accuracy rates, these findings are largely based on laboratory tasks, controlled conditions, and curated datasets rather than real-world forensic contexts” (p. 18).

• Weak theoretical foundations and uncertain measurement of deception. “Psychological research does not support a single cue, or fixed set of cues, uniquely associated with deception” (p. 18). The author further warns that “current systems may appear accurate without truly detecting deception, thereby increasing the risk of false positives” (p. 18).

• Transfer of human bias into AI systems. “A common misconception is that AI can eliminate human bias in credibility assessments. However, biases already present in human deception judgements may instead be transferred into AI systems” (p. 18).

• Dataset bias and reliance on irrelevant characteristics. “Models may appear highly accurate while relying on incidental characteristics such as race, sex, age, or communication style rather than genuine indicators of deception” (p. 19).

• Unwarranted trust in apparently successful systems. “This creates a risk that practitioners place unwarranted trust in systems that appear accurate in testing but lack fairness and generalisability” (p. 19).

• Black-box decision-making and restricted explainability. “These systems are o en described as ʻblack boxesʼ because the process linking input data to outputs, such as classifying someone as truthful or deceptive, is not always visible or understandable” (p. 19). Consequently, “This makes it difficult to verify classifications, identify errors, or assess the reliability of the system” (p. 19).

• Inability to defend unexplained conclusions. An expert who cannot articulate why a tool produced a particular output, or what processes generated a conclusion, is not in a position to defend that conclusion professionally or ethically. (p. 19)

• Cultural, clinical, and contextual limitations. “A major limitation of AI deception detection is its limited sensitivity to cultural, clinical, and contextual variation”. The author also identifies datasets that “o en lack sufficient cultural, ethnic, gender, and linguistic diversity, reducing the generalisability and trustworthiness of AI deception systems in real-world settings” (p. 20).

• Misinterpretation of behavior unrelated to deception. “These behaviours may instead reflect cognitive load, emotion, or communication differences rather than deception itself” (p. 20). Without contextual sensitivity, systems “may reinforce false suspicion, bias credibility assessments, and encourage unnecessarily confrontational interviewing practices” (p. 20).

• Disproportionate risk to vulnerable individuals. Trauma survivors, neurodivergent individuals, suspects with mental health difficulties, and witnesses under acute stress are among those whose communicative patterns may be most susceptible to misclassification. (p. 20)

• AI sycophancy and apparent trustworthiness. “AI sycophancy refers to a modelʼs tendency to over-agree with users, validate their beliefs, or provide responses that are more pleasing than accurate” (p. 20). The author warns that “the system may shape user interpretation by reinforcing assumptions and appearing more trustworthy than it is” (p. 20).

• Confirmation bias and investigative tunnel vision. “Sycophantic systems may actively reinforce confirmation bias and investigative tunnel vision, two of the most well-documented threats to the integrity of forensic decision making” (p. 21).

• Reduced scrutiny of alternative explanations and adverse investigative consequences. “This risks entrenching pre-existing assumptions, reducing critical scrutiny of alternative explanations, and narrowing the investigative frame in ways that are difficult to detect and harder still to correct” (p. 21). The identified consequences include “wrongful conclusions, missed evidence, and failures of due process” (p.21).

AI Use Statement

Perplexity AI was used in the development of this information.

References

Fumagalli, A. (2026). Ethical challenges of AI-based deception detection in forensic settings. Assessment & Development Matters, 18(3), 17–23. https://doi.org/10.53841/bpsadm.2026.18.3.17  

Perplexity AI. (n.d.). Perplexity [Generative AI tool]. https://www.perplexity.ai/

 

Cite this page:

Kardasz, F. (2026, October). AI & Digital Forensics -  Review of Fumagalliʼs (2026) Ethical challenges of AI-based deception detection...    https://kardasz.blogspot.com/2026/10/ai-digital-forensics-review-of_0145261725.html