Can EEG Replace the Polygraph? A Critical Review of Brain-Based Concealed Information Detection

EEG-based event-related potentials and machine-learning classifiers are increasingly proposed as more direct alternatives to conventional polygraph examinations. Dr Keith Ashcroft, investigative psychologist and forensic polygraph examiner, critically evaluates whether the evidence supports replacing established forensic methodology with laboratory signal-processing techniques.

Introduction

There is growing interest in the use of electroencephalography (EEG), event-related potentials (ERPs) and machine-learning classification as supposedly more direct or objective alternatives to conventional forensic polygraph examinations. The P300 component — a positive voltage deflection occurring approximately 300 milliseconds after a meaningful stimulus — has attracted particular attention as the basis for brain-based concealed information detection.

Such approaches are frequently presented under the heading “brain-based lie detection,” but this label can obscure what is actually being measured. The P300 does not detect lies. It reflects recognition, attention and memory updating. The distinction matters for any serious forensic evaluation.

A recent systematic review published in Brain Sciences (Taha, Baykara & Alakuş, 2025) surveys EEG signal-processing and machine-learning methods applied to deception-related paradigms, reporting classification accuracies ranging from approximately 81% to nearly 100%. This article critically examines:

• what the review claims;
• what P300 methods actually measure;
• the strength of the underlying evidence;
• whether the reported accuracy figures are forensically meaningful;
• and whether EEG can replace the wider methodology of a forensic polygraph examination.

What Does the P300 Actually Measure?

The P300 is an event-related brain potential associated with attention, recognition, stimulus significance and the process of memory updating (Polich, 2007). When a stimulus is meaningful to the participant — because it is recognised, unexpected or personally relevant — the P300 amplitude tends to be larger than when the stimulus is irrelevant.

A P300-based Concealed Information Test (CIT) does not directly detect lying. It attempts to determine whether a person recognises specific information presented among plausible alternatives. Several distinctions are critical:

• Recognition is not deception. A person may recognise an item without having committed any wrongful act.
• Concealed information is not guilt. Recognising a crime-relevant detail does not establish culpability.
• Memory detection is not truth verification. The test probes whether an item is familiar, not whether a statement is true.
• Stimulus classification is not forensic decision-making. Sorting EEG signals into categories is not equivalent to determining individual credibility.

A recognised item may be familiar because the examinee committed an act, witnessed it, was told about it, saw it in the media, handled it legitimately or encountered it during an investigation.

Equally, the absence of a strong recognition response does not necessarily demonstrate innocence. A person may not have noticed, encoded or retained the relevant detail. These asymmetries are fundamental to interpreting any CIT result, whether measured through EEG or autonomic physiology.

A False Comparison Between EEG and “the Polygraph”

The reviewed paper creates an overly simple comparison between EEG methods and “traditional polygraph methods.” This framing conflates levels of description that should be kept separate.

A Concealed Information Test is a test paradigm — a structured way of presenting items to determine whether particular information is recognised. EEG, electrodermal activity, respiration and cardiovascular measurement are measurement channels — ways of recording physiological responses during such a test.

A conventional polygraph instrument can also be used to conduct an autonomic Concealed Information Test, recording electrodermal, respiratory and cardiovascular responses to probe and control items. The scientifically meaningful comparison is therefore often:

P300 measurement versus autonomic measurement within a Concealed Information Test — not neuroscience versus the polygraph.

The major meta-analysis by Meijer, Klein Selle, Elber and Ben-Shakhar (2014) directly compared P300 and autonomic measures within CIT paradigms and found that P300 showed a higher mean effect size than skin conductance in personal-item studies, but not consistently in mock-crime paradigms. This comparison could not be meaningfully drawn from the MDPI review, which excluded many peripheral physiological and polygraph studies and did not conduct valid matched head-to-head comparisons.

The Problem with Reported Accuracy Figures

The review references accuracy figures ranging from approximately 81% to nearly 100%. These percentages may, however, refer to very different outcomes:

• classification of individual EEG epochs or trials;
• classification of short overlapping signal windows;
• discrimination between experimental conditions (e.g. truthful versus deceptive blocks);
• familiar versus unfamiliar stimuli;
• within-participant classification;
• or classification of truly independent participants.

These outcomes should not be treated as equivalent. Forensic accuracy requires evaluation on completely new individuals, preferably in different settings, using predefined decision rules and realistic case conditions.

Rigorous forensic validation requires attention to:

• participant-level train–test separation;
• leave-one-subject-out cross-validation;
• external validation on independent cohorts;
• preregistered analysis pipelines;
• sensitivity and specificity reported separately;
• false-positive and false-negative rates;
• inconclusive outcomes;
• confidence intervals;
• likelihood ratios;
• calibration;
• and realistic base rates.

High accuracy in classifying thousands of signal segments from a small number of participants is not equivalent to correctly classifying a new individual in a forensic examination.

Critique of the Near-100% Results

The LieWaves dataset (Aslan, Baykara & Alakuş, 2024) is reported to achieve classification accuracy of approximately 99.88%. This should not automatically be interpreted as near-perfect forensic accuracy. Several methodological concerns require consideration:

• The sample comprised only 27 healthy university participants.
• Conditions involved artificial truthful and deceptive tasks.
• Personal stakes were minimal — there were no consequences for being detected.
• The study used overlapping sliding-window data augmentation, which can dramatically increase the number of training and testing samples from a small pool of participants.
• Where sliding windows overlap substantially, highly similar signal segments may appear in both training and testing sets, inflating apparent accuracy.
• Even without literal overlap, machine-learning classifiers may learn participant-specific or session-specific EEG signatures rather than genuine deception-related features.
• Overfitting is a recognised risk whenever the number of extracted features greatly exceeds the number of independent observations.
• Trial-and-error model selection — testing multiple classifiers and reporting the best — can produce optimistically biased results.
• External replication on independent samples has not been demonstrated.
• Subject-independent generalisation — the ability to classify a participant whose data were entirely absent from training — has not been clearly established.

None of this implies misconduct. These are standard methodological limitations and validation risks that apply to any machine-learning study in the biomedical sciences.

It should also be noted, in fairness, that two of the three authors of the systematic review (Baykara and Alakuş) were co-authors of the LieWaves study. This does not invalidate either work, but it means that the headline accuracy result featured in the review is not an instance of independent replication.

Limitations of the Systematic Review

The review methodology raises several concerns when assessed against established standards for systematic reviews (Page et al., 2021):

• Search terminology. The search strategy was oriented towards engineering and classification terms (EEG, machine learning, deep learning, deception detection). Terms widely used in the established psychophysiological literature — such as “memory detection,” “concealed knowledge,” “guilty knowledge” and “recognition” — may have been underrepresented, potentially missing important studies.
• PRISMA flow. It is unclear how the number of records included in the PRISMA flow diagram relates to the much smaller number of studies prominent in the accuracy summary table.
• Risk-of-bias assessment. The review does not appear to include a detailed, study-level risk-of-bias assessment of the kind recommended by PRISMA 2020.
• Heterogeneity. The included studies differ substantially in participants, stimuli, preprocessing methods, classifiers and outcome measures. This heterogeneity makes aggregated accuracy ranges difficult to interpret.
• No formal diagnostic meta-analysis. A systematic review of diagnostic accuracy should ideally include pooled sensitivity, specificity, summary receiver-operating-characteristic curves and tests for publication bias.
• Peak versus representative performance. There is a risk that the review highlights the highest accuracy achieved across multiple classifiers and parameter settings rather than representative or median performance.

Studies with very small samples and unstable train–test splits should not carry the same evidential weight as independently replicated, participant-level validation studies.

Countermeasures

Any suggestion that EEG-based P300 testing is inherently resistant to countermeasures deserves careful scrutiny. Rosenfeld, Soskins, Bosh and Ryan (2004) demonstrated that simple mental countermeasures — such as covert responses to irrelevant stimuli — could reduce P300-based detection rates from over 90% to chance level in some conditions.

Participants may attempt to manipulate the procedure through:

• redirecting attention during probe presentation;
• mental counting or imagery tied to irrelevant items;
• assigning deliberate personal significance to control items;
• covert physical movements;
• altering eye movements or blink patterns;
• generating muscle artefacts;
• and varying response timing strategies.

Although more recent protocols such as the Complex Trial Protocol (Rosenfeld, 2020) were designed to improve countermeasure resistance, these protocols are themselves still under investigation and have not been extensively validated in adversarial field conditions with informed and motivated participants.

Countermeasure resistance must be demonstrated through adversarial testing — not inferred from ordinary laboratory classification performance. Machine learning may improve signal classification but does not automatically solve the countermeasure problem.

Ecological Validity

The gap between laboratory experiments and forensic practice is substantial. Many of the studies summarised in the review involve:

• healthy young volunteers, typically university students;
• mock crimes or instructed deception;
• minimal or no consequences for detection;
• carefully selected, unambiguous stimuli;
• controlled and quiet recording environments;
• repeated trials under standardised conditions;
• and participants without significant clinical or cognitive complexity.

Real forensic examinees may present with substantially different characteristics, including:

• anxiety and emotional distress;
• trauma histories;
• medication effects;
• neurodevelopmental differences;
• fatigue, pain or physical discomfort;
• poor or unreliable memory;
• sleep deprivation;
• substance use;
• mental-health difficulties;
• age-related neurophysiological variation;
• movement artefacts;
• and strong motivation to manipulate the procedure.

The field evidence for P300-based CIT remains limited. Ben-Shakhar and Elaad (2003) observed a significant gap between laboratory effect sizes and the smaller number of field-based assessments available for autonomic CIT measures. The situation for EEG-based CIT is, if anything, further from field validation.

A forensic method must be robust across these conditions. Demonstrating laboratory accuracy in optimal conditions is a necessary first step, but it is not sufficient.

EEG Cannot Replace the Wider Examination Methodology

A forensic polygraph examination is not simply a collection of physiological tracings. The broader process includes:

• instruction from the commissioning party and case review;
• suitability assessment;
• informed consent;
• identification of the precise issue to be tested;
• formulation of testable questions;
• review of question meaning with the examinee;
• explanation of the procedure;
• standardised, multi-channel physiological data collection;
• validated numerical analysis;
• artefact assessment;
• post-examination discussion and opportunity to explain significant reactions;
• professional reporting;
• quality-control review;
• and retention of sufficient records for independent scrutiny.

The MDPI paper primarily addresses signal processing and machine-learning classification of EEG data. It does not provide an equivalent forensic framework governing suitability, question construction, alternative explanations, post-test review or defensible reporting.

Replacing the complete examination process with an EEG classifier could remove procedural safeguards rather than strengthen them.

Where P300-CIT May Have Genuine Value

It would be wrong to reject the technology entirely. P300-based Concealed Information Testing may be genuinely useful in circumstances where investigators possess multiple protected details that:

• are genuinely associated with an event;
• were likely to have been noticed and encoded;
• are known to investigators;
• are unknown to uninvolved individuals;
• have not been publicly disclosed;
• can be presented alongside equally plausible alternatives;
• and can be repeated without revealing the correct answer.

Examples might include recognition of a concealed location, knowledge of a weapon, identification of an object used during an offence, recognition of a body-recovery detail, knowledge of a protected sequence, or identification of an undisclosed image.

Such applications are closer to memory detection than general credibility assessment — and this is both the legitimate strength and the inherent limitation of the P300 approach.

P300-CIT is poorly suited, however, to many common polygraph referral questions involving:

• historical conduct;
• disputed conversations;
• domestic behaviour;
• therapeutic disclosure;
• habitual conduct;
• broad screening;
• intentions;
• allegations already discussed extensively;
• or matters for which no protected alternatives exist.

Could EEG Supplement Polygraph Methodology?

A more productive question may be whether EEG could eventually serve as a supplementary measurement channel rather than a replacement. Possible future roles include:

• an additional channel in specialist Concealed Information Testing;
• a research tool for understanding the cognitive processes underlying recognition;
• a supplementary measure where protected information exists;
• or part of a multimodal assessment combining central and autonomic nervous-system responses.

However, simply combining technologies should not be assumed to improve accuracy. Multimodal methods require:

• independent validation of the combined procedure;
• predefined integration rules;
• testing for correlated errors between channels;
• demonstrated resistance to countermeasures;
• participant-level external validation;
• and transparent, predefined decision thresholds.

The assumption that combining two imperfect measures automatically produces a more accurate assessment is precisely the statistical-independence problem discussed above. It requires empirical demonstration, not arithmetic.

Summary Assessment

Conclusion

The systematic review by Taha, Baykara and Alakuş (2025) is useful as a survey of EEG signal-processing and machine-learning approaches applied to deception-related paradigms. It identifies genuine progress in automated classification methods and highlights the experimental sensitivity of P300-based recognition measures.

However, the review does not establish a replacement methodology for forensic polygraph practice. The central category error is treating successful classification of laboratory EEG data as equivalent to determining whether a person is truthful in an individual forensic case. These are different problems requiring different standards of evidence.

P300-based EEG shows genuine experimental sensitivity to recognition of meaningful information. However, small samples, heterogeneous methods, limited participant-independent validation, vulnerability to information contamination and countermeasures, and weak field evidence prevent it from replacing established forensic polygraph methodologies. Its most defensible present role is as an experimental or supplementary measure within carefully designed concealed-information assessments.

References

1. Aslan, M., Baykara, M., & Alakuş, T. B. (2024). LieWaves: dataset for lie detection based on EEG signals and wavelets. Medical & Biological Engineering & Computing. https://doi.org/10.1007/s11517-024-03021-2
2. Ben-Shakhar, G., & Elaad, E. (2003). The validity of psychophysiological detection of information with the Guilty Knowledge Test: A meta-analytic review. Journal of Applied Psychology, 88(1), 131–151. https://doi.org/10.1037/0021-9010.88.1.131
3. Meijer, E. H., Klein Selle, N., Elber, L., & Ben-Shakhar, G. (2014). Memory detection with the Concealed Information Test: A meta-analysis of skin conductance, respiration, heart rate, and P300 data. Psychophysiology, 51(9), 879–904. https://doi.org/10.1111/psyp.12239
4. Page, M. J., McKenzie, J. E., Bossuyt, P. M., Boutron, I., Hoffmann, T. C., Mulrow, C. D., … & Moher, D. (2021). The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. BMJ, 372, n71. https://doi.org/10.1136/bmj.n71
5. Polich, J. (2007). Updating P300: An integrative theory of P3a and P3b. Clinical Neurophysiology, 118(10), 2128–2148. https://doi.org/10.1016/j.clinph.2007.04.019
6. Rosenfeld, J. P., Soskins, M., Bosh, G., & Ryan, A. (2004). Simple, effective countermeasures to P300-based tests of detection of concealed information. Psychophysiology, 41(2), 205–219. https://doi.org/10.1111/j.1469-8986.2004.00158.x
7. Rosenfeld, J. P. (2020). P300 in detecting concealed information and deception: A review. Psychophysiology, 57(7), e13362. https://doi.org/10.1111/psyp.13362
8. Taha, B. N., Baykara, M., & Alakuş, T. B. (2025). Neurophysiological approaches to lie detection: A systematic review. Brain Sciences, 15(5), 519. https://doi.org/10.3390/brainsci15050519

Frequently Asked Questions

This article was written by Dr Keith Ashcroft, investigative psychologist and principal forensic polygraph examiner at the Centre for Forensic Neuroscience. Dr Ashcroft is a Chartered Psychologist registered with the British Psychological Society and a full member of the American Polygraph Association.