Field Validity of AI Mental-State Inference Systems

Full report

On this page

1. Introduction

The ambition to digitize, quantify, and categorize the human mind has catalyzed a sprawling industry of artificial intelligence systems designed to infer internal mental states from external, physically observable indicators. Operating under designations such as affective computing, behavioral biometrics, and emotion AI, these systems assert the capability to detect highly complex and fluid psychological constructs. Developers claim their algorithms can accurately infer emotion, attention, engagement, credibility, deception, personality, vulnerability, stress, loyalty, intent, and dangerousness. As these technologies migrate from controlled experimental settings into high-stakes environments—spanning workplaces, schools, healthcare facilities, police departments, border crossings, insurance firms, and consumer automotive systems—the demand for rigorous field validation has become a matter of profound civic and scientific urgency1. However, the proliferation of mental-state inference systems has vastly outpaced the scientific consensus regarding their foundational validity. A critical epistemological chasm exists between the measurement of a physiological or behavioral correlate—such as the contraction of a facial muscle, a micro-tremor in the vocal cords, or a shift in keyboard typing cadence—and the definitive identification of a latent cognitive state1. Vendor demonstrations and laboratory classification scores routinely conflate physical sensor accuracy with psychological construct validity, promoting algorithmic outputs as objective truths rather than probabilistic, context-dependent inferences1. This exhaustive research report evaluates the field validity of AI mental-state inference systems. By analyzing fourteen distinct real-world deployments across eight primary modalities—facial movement analysis, voice and speech analysis, gaze and head-pose tracking, keystroke and interaction dynamics, wearables and physiological signals, behavioral telemetry, consumer electroencephalography (EEG), and multimodal architectures—this analysis isolates the profound methodological vulnerabilities that emerge when these technologies confront the complexity of uncontrolled human populations.

2. A Taxonomy of Field Validity Evaluation

To rigorously assess these systems, one must abandon the monolithic conceptualization of an algorithm as a simple "lie detector" or "attention monitor." An exhaustive evaluation requires dissecting the computational pipeline and evaluating it against a strict taxonomy of validity metrics. The failure to separate these layers routinely allows vendors to mask catastrophic psychological invalidity behind high mechanical accuracy.

2.1 Sensor Accuracy and Feature Extraction

Sensor accuracy defines the physical capability of hardware to capture a signal in a real-world environment amid ambient noise. Examples include a 30-frames-per-second camera attempting to capture fleeting facial micro-gestures, or a dry-electrode consumer EEG headset recording microvolt electrical potentials on a student's scalp1. Feature extraction is the subsequent algorithmic translation of that raw physical data into a measurable metric, such as isolating the alpha frequency band (8–12 Hz) from a noisy EEG stream or mapping a localized pixel shift to a specific facial Action Unit (AU)7. High sensor accuracy and precise feature extraction merely demonstrate that a system can reliably document a physical event; they offer zero proof that the system understands the psychological meaning of that event.

2.2 Label Quality and Construct Validity

Supervised machine learning models rely on annotated training data, introducing the vulnerability of label quality. If an algorithm is trained to detect "dangerousness" or "credibility," the training data must contain ground-truth examples of those states. Because these states are invisible, labels are subjectively applied by human annotators whose cultural and cognitive biases are permanently encoded into the algorithm9. The most profound failure point in mental-state inference is construct validity: the degree to which a system actually measures the psychological construct it claims to measure. The dominant paradigm in emotion AI relies on a typological view of emotion, heavily influenced by classical theories that assume a universal, biological mapping between specific facial configurations and internal states4. Modern affective science, notably the Theory of Constructed Emotion, systematically refutes this premise. Extensive research demonstrates that emotional expressions are highly context-dependent, culturally relative, and physically variable10. A human does not uniformly scowl when angry, nor does a scowl universally indicate anger; the brain constructs emotion based on situational context and physical allostasis11. A system may have perfect feature extraction for a furrowed brow, but entirely lack construct validity if it assumes that feature universally signals stress, deception, or a hostile personality.

2.3 Criterion Validity, Calibration, and Generalization

Criterion validity evaluates whether the algorithmic output accurately predicts a real-world outcome, such as whether a candidate flagged for high "loyalty" actually remains at a company longer than their peers. Calibration and generalization measure whether a model trained in a controlled laboratory maintains its predictive power when exposed to novel, heterogeneous populations in the field. Systems trained on small, homogeneous samples—such as the 32 to 39 predominantly white male subjects used to train specific border control deception algorithms—routinely fail to generalize, leading to catastrophic accuracy degradation and algorithmic brittleness in real-world deployments9.

2.4 Base-Rate Sensitivity and the False Positive Paradox

In field deployments involving mass screening—such as border control, public policing, or counterterrorism—the statistical base rate of the target construct (e.g., a terrorist, a smuggler, or a violent threat) is exceptionally low. According to Bayesian probability, even an inference system with 95% sensitivity and 95% specificity will generate overwhelming numbers of false positives when screening a population where the prevalence of the target state is 1 in 100,0005. The statistical reality of the false positive paradox renders many mental-state inference systems functionally detrimental in low-prevalence screening environments, as they flood operators with inaccurate alerts.

2.5 Subgroup Burden and Institutional Action

Subgroup burden refers to the disparate impact algorithmic errors impose on specific demographic cohorts. Because physical expressions of emotion, vocal cadence, and physiological baselines vary extensively by race, gender, neurodiversity, and cultural background, systems optimized for a majority population routinely penalize marginalized groups9. They classify non-majority baseline behaviors as anomalous, deceptive, or inattentive. Finally, the validity of an inference system must be judged against the institutional action taken upon its output. A system used to trigger a soft coaching prompt in a corporate training module carries vastly different validation requirements than a system used to deny a traveler entry at a national border, reject a housing benefit claim, or automatically deny employment16.

3. Facial Movement and Affective Analysis

Facial movement analysis is the most heavily commercialized modality of mental-state inference, deployed globally in hiring, security, and consumer retail. These systems map the geometry of the face, track the movement of specific muscle groups, and correlate these variations to purported emotional states, personality traits, or cognitive engagement.

3.1 Field Deployment 1: HireVue (Corporate Hiring)

HireVue, a dominant vendor of AI-based hiring software, deployed asynchronous video interview platforms that evaluated job candidates across global corporate workplaces. The system utilized facial expressions, vocal tone, and word choice to infer complex constructs such as "cognitive ability," "emotional intelligence," "personality," and "loyalty"16. By late 2020, the platform had been used to assess millions of candidates, operating on the premise that the facial micro-expressions of a company's historical top performers could be quantified and used as an algorithmic benchmark to screen new applicants20. The field validity of HireVue's facial analysis faced insurmountable challenges regarding construct validity and subgroup burden. The assumption that a specific facial expression (e.g., a lack of eye contact or a specific smiling duration) correlates universally with emotional intelligence or high job performance is wholly unsupported by modern affective science4. The algorithms were highly susceptible to penalizing neurodivergent candidates, individuals with facial paralysis, or individuals from cultural backgrounds with differing norms for expressive intensity20. Following a 2019 complaint to the Federal Trade Commission (FTC) by the Electronic Privacy Information Center (EPIC), which argued the system produced results that were "biased, unprovable, and not replicable," alongside the passing of regulatory measures like Illinois's Artificial Intelligence Video Interview Act, HireVue quietly phased out the facial analysis component of its assessments in early 202116. The removal of the core facial feature set served as a profound market admission that the technology lacked defensible criterion validity for inferring personality and employment fit in high-stakes environments.

3.2 Epistemological Failures in Affective Recognition

The foundational flaw in systems claiming to read emotions from faces lies in the outdated object-perception analogy of emotion. Extensive research encompassing tens of thousands of participants across cultures confirms that humans do not uniformly scowl when angry, nor do they only smile when happy10. The brain constructs emotions in a highly contextual, goal-directed manner11. As Barrett et al. documented in their comprehensive review, a single configuration of facial muscles takes on vastly different emotional meanings depending on the surrounding context, the perceiver's history, and the physical environment4. When algorithms are stripped of this rich context and forced to categorize a localized pattern of pixels into discrete emotional buckets, they do not detect objective emotions; they detect stereotyped cultural templates8. Relying on these templates to infer "credibility" or "engagement" results in the systemic measurement of a subject's compliance with a cultural norm, rather than the accurate measurement of their internal mental state.

4. Voice, Speech, and Paralinguistic Analysis

Voice and speech analysis evaluates the acoustic properties of human speech—such as fundamental frequency, pitch variation, cadence, micro-tremors, and spectral energy—to infer states of stress, deception, vulnerability, or empathy.

4.1 Field Deployments 2 & 3: Nemesysco VRA by UK DWP and Local Councils (Insurance/Benefits)

In the United Kingdom, the Department for Work and Pensions (DWP) and numerous local councils, including Harrow and Birmingham, deployed Voice Risk Analysis (VRA) software to screen telephone calls for housing benefit, council tax, and income support claims24. Developed by the Israeli firm Nemesysco, the technology claimed to utilize "layered voice analysis" (LVA) to detect "brain activity traces" and minute, involuntary changes in the speech waveform, classifying these anomalies as stress, excitement, or deception24. Call center operators were trained to listen for an algorithmic beep indicating stress, which triggered targeted, investigative questioning designed to uncover fraud24. Independent scientific evaluation of the technology revealed catastrophic validity failures. Phoneticians Anders Eriksson and Francisco Lacerda published peer-reviewed analyses dismantling the underlying principles of the software. They noted that the theoretical foundation of detecting vocal micro-tremors to identify deception lacked any scientific basis, famously concluding the software operated at the "astrology end of the validity spectrum"17. Field data released by the DWP corroborated this theoretical critique. In extensive trials costing over £2.16 million and analyzing 45,000 calls across 24 local authorities, the system was judged successful in only five instances17. DWP statisticians confirmed that the technology functioned no better than flipping a coin, routinely flagging truthful citizens—likely experiencing the standard vulnerability and anxiety associated with needing welfare assistance—as "high risk"26. The UK government officially abandoned the VRA technology for benefits screening in 2010 due to its unreliability17. The councils that claimed positive returns on investment achieved institutional outcomes not through accurate deception detection, but through the psychological deterrent effect of warning callers that a "lie detector" was in use—a phenomenon that has no bearing on algorithmic construct validity24.

4.2 Field Deployments 4 & 5: Cogito at MetLife and Humana (Workplaces/Healthcare)

Conversely, voice AI has been deployed in enterprise environments for behavioral shaping rather than punitive screening. Cogito, an AI platform utilizing natural language processing and acoustic analysis, was deployed in massive call centers by major insurers, including MetLife and Humana18. Cogito monitors the pace, tone, and vocal energy of customer service representatives in real-time. If an agent speaks too quickly, sounds fatigued, or interrupts a caller, the system flashes visual cues on their screen—a coffee cup icon to suggest "more energy," or a heart icon to prompt "empathy"18. While Cogito demonstrates high sensor accuracy in detecting acoustic variables like latency, overlap, and speaking volume30, its inference of "empathy" remains a proxy measurement. The system does not measure the internal empathetic state of the worker; it enforces a standardized acoustic performance of empathy18. Because the institutional action is real-time behavioral coaching rather than termination or denial of services, the tolerance for construct validity errors is higher. Nonetheless, critical reviews highlight severe subgroup burdens: employees with neurodivergent conditions, distinct cultural communication styles, or speech impediments risk being continuously penalized by an algorithm trained on a narrow, homogenized acoustic definition of ideal corporate speech18.

5. Gaze, Head-Pose Tracking, and Pupillometry

Ocular parameters—including blink rate, saccadic velocity, fixation patterns, and pupil dilation—are tightly linked to the autonomic nervous system. Vendors leverage these metrics to infer cognitive load, attention, and ultimately, deception.

5.1 Field Deployment 6: Converus EyeDetect (Policing/Hiring)

Converus developed the EyeDetect system, an automated screening kiosk where subjects answer true/false questions while an infrared camera tracks pupil dilation, blink rates, and response latency at 60 frames per second32. The underlying theory posits that deceiving requires more cognitive effort than truth-telling, leading to measurable increases in pupil size and reduced blink rates due to heightened cognitive load32. The system has been deployed by local law enforcement agencies and private employers for pre-employment screening and internal investigations32. Converus claims an accuracy rate of 86% to 90%; however, these figures are predominantly derived from laboratory experiments and in-house validation studies where subjects are instructed to commit low-stakes mock crimes32. The field validity of EyeDetect suffers from a fundamental construct validation error: cognitive load is not synonymous with deception33. In a high-stakes field environment (e.g., a police interrogation or a critical job interview), truth-tellers experiencing severe anxiety, recalling traumatic events, or suffering from fatigue will exhibit physiological markers of cognitive load indistinguishable from those of a deceptive individual33. Furthermore, ocular tracking introduces a severe subgroup burden for individuals with visual impairments, neuroatypical gaze behaviors, or medical conditions affecting pupillary responses, making standardized algorithmic evaluation highly discriminatory20.

5.2 Field Deployment 7: Subaru DriverFocus (Consumer Automotive)

In the consumer sector, Subaru deployed the DriverFocus distraction mitigation system, utilizing an infrared camera positioned on the dashboard to track the driver's head pose and eye position35. If the system detects that the driver's eyes are diverted from the road for an extended period, or if the head nods in a manner consistent with drowsiness, it triggers auditory and visual alerts35. Unlike systems inferring complex emotions or deception, DriverFocus possesses exceptionally strong criterion validity. The physical correlate being measured (eyes physically diverted from the road) is not merely a proxy for the construct (distracted driving)—it is the physical manifestation of the construct itself. The system does not attempt to infer why the driver is distracted (e.g., whether they are stressed, angry, or lost in thought); it simply measures the physical geometry of attention relative to the windshield. This direct alignment between the sensor data and the operational objective makes it a highly valid deployment of behavioral tracking in the field.

6. Keystroke and Interaction Dynamics

Interaction dynamics evaluate how a user physically interacts with a digital device. Metrics include typing cadence, mouse trajectories, scroll speed, touch pressure, and device orientation. These systems seek to infer user identity, familiarity, and cognitive states such as hesitation or duress.

6.1 Field Deployment 8: BioCatch (Finance and Insurance)

BioCatch is a behavioral biometrics platform deployed by over 150 of the world's largest banks to prevent account takeover, fraud, and social engineering scams37. Operating invisibly in the background of mobile and web banking sessions, the platform collects continuous telemetry on physical and cognitive interactions39. BioCatch relies on two distinct analytical vectors. First, it establishes a historical baseline for an individual user, cataloging their standard typing speed, typical mouse arcs, and device holding angles. If a session deviates drastically from this baseline—such as exhibiting segmented typing, straight-line touch paths indicative of automated bots, or zero finger pressure indicating a remote-access trojan—the system flags an account takeover38. Second, the system attempts to infer specific mental states, primarily "cognitive stress" or "coercion," to detect authorized push payment fraud, where a legitimate user is manipulated by a scammer into transferring money. BioCatch looks for behavioral anomalies such as prolonged hesitation, segmented reading of account numbers, or erratic mouse movements that indicate the user is receiving real-time instructions from a malicious actor over the phone38. The field validity of interaction dynamics in this context is unusually robust. Because BioCatch often compares a user against their own historical baseline, it entirely bypasses the typological fallacy that plagues emotion AI, which erroneously compares users to an artificial universal standard39. Furthermore, the institutional action taken—typically triggering a step-up authentication protocol or temporarily delaying a high-value transfer—is calibrated appropriately to the risk of false positives, mitigating the ultimate burden on the consumer41.

7. Wearables, Physiological Signals, and Behavioral Telemetry

The continuous monitoring of autonomic signals (heart rate, electrodermal activity, respiration) and behavioral metadata forms the basis for inferring complex psychological threat states and assessing credibility.

7.1 Field Deployment 9: DHS FAST Program (Borders/Security)

The U.S. Department of Homeland Security (DHS) Science and Technology Directorate funded the Future Attribute Screening Technology (FAST) project, an initiative designed to detect "malintent" (the intent to cause harm) at security checkpoints42. FAST utilized a suite of non-contact sensors—including cardiovascular monitoring, respiratory tracking, eye tracking, and thermal imaging—to measure physiological cues42. The system was grounded in the "Theory of Malintent," positing that individuals intending to execute an attack exhibit a specific, identifiable physiological signature distinguishable from standard travel anxiety42. The FAST program represents a catastrophic failure of construct validity. The scientific consensus fundamentally rejects the premise that there is a unique, universal physiological signature for "dangerousness" or the intent to cause harm10. A rapid heart rate, erratic respiration, or elevated electrodermal activity can result from fear of missing a flight, white-coat syndrome regarding law enforcement, physiological conditions, or the ambient stress of a busy airport environment42. FAST conflated generalized autonomic arousal with targeted criminal intent. The inability to scientifically validate the premise that terrorists behave physiologically differently than stressed civilians, combined with the extreme base-rate rarity of terrorists at checkpoints, resulted in the project's conceptual stagnation and failure to transition into widespread operational use46.

7.2 Field Deployment 10: VeriPol (Policing)

In 2018, the Spanish National Police deployed VeriPol, an AI system designed to automatically identify false crime reports (specifically fabricated robberies) by analyzing the text and linguistic structures of written statements34. Using natural language processing (NLP), VeriPol extracted features such as the scarcity of sensory details or the use of convoluted phrasing, which the algorithmic model associated with deception34. While developers claimed an accuracy rate exceeding 90% in identifying false reports during laboratory testing, independent review highlighted severe limitations in calibration and transparency when deployed in the field34. The fundamental flaw lies in construct validity and subgroup burden: the linguistic markers of a fabricated story often precisely mirror the linguistic markers of severe trauma. Victims of violent crime frequently exhibit fragmented memory, a lack of cohesive sensory detail, and non-linear storytelling due to the neurobiological impact of trauma on memory encoding. Relying on an automated NLP tool to flag these trauma markers as "deception" introduces a massive institutional risk, effectively penalizing the most vulnerable victims. The Spanish National Police eventually discontinued the use of VeriPol, citing its lack of evidentiary value in judicial proceedings and the failure to meet the scientific rigor required for operational reliance34.

8. EEG and Consumer Neurotechnology

Advancements in dry-electrode technology have transitioned electroencephalography (EEG) from the clinical laboratory into consumer, educational, and healthcare sectors. These devices measure electrical potentials on the scalp to infer cognitive states such as relaxation, mental workload, and attention.

8.1 Field Deployment 11: BrainCo Focus in Schools (Education)

BrainCo, a neurotechnology startup with roots in Harvard University's Innovation Lab, developed the Focus 1 and FocusCalm EEG headbands7. These consumer-grade devices utilize a minimal number of electrodes (typically 1 to 3 on the forehead) to record brainwave activity. BrainCo deployed these devices in various real-world environments, most controversially in Chinese primary schools, where students wore the headbands to measure "task-focused attention"47. The system analyzed Alpha and Theta frequency bands to calculate an engagement index, feeding real-time data to a teacher's dashboard to identify which students were concentrating and which were distracted7. The technology was also integrated into e-learning platforms like Progrentis in Spain to "neuro-optimize" content delivery47. The deployment of consumer EEG for attention monitoring suffers from acute degradation of both sensor accuracy and construct validity. Non-invasive EEG signals are inherently weak (1 to 100 microvolts) and must travel through the skull and scalp, resulting in poor spatial resolution7. When deployed in a naturalistic environment like a classroom, the raw signal is massively contaminated by physical artifacts. Frontal electrodes are highly susceptible to electromyographic (EMG) noise from facial muscle movements, such as jaw clenching, raising eyebrows, and electrooculographic (EOG) noise from blinking6. Consequently, an algorithm identifying a spike in high-frequency "brain activity" may actually be detecting the muscle tension of a stressed student grinding their teeth2. More insidiously, equating specific EEG band ratios (e.g., Beta/Alpha+Theta) with the pedagogical concept of "learning engagement" represents a profound reductionist error49. Attention is a complex, multifaceted cognitive process, not a monolithic state localized in a single frequency band47. By institutionalizing this biometric proxy, schools risk reconfiguring the definition of attention to fit the device's output, penalizing students who process information differently or neurodivergent students whose baseline spectral densities deviate from the algorithmic norm2. Furthermore, the extraction of Neural Personal Information (NPI) raises severe data privacy and legal concerns, as these raw signals can theoretically be mined for unintended cognitive inferences52.

8.2 Field Deployment 12: BrainCo VR-BCI for Autism (Healthcare)

Conversely, BrainCo's EEG technology has been deployed in healthcare settings with distinctly different institutional actions. BrainCo partnered with the China Rehabilitation Research Center to create wearable brainwave rehabilitation systems for children with autism52. Integrating BrainCo's FocusCalm headband with Virtual Reality (VR) environments, these Brain-Computer Interface (BCI) systems attempt to address social deficits51. In these deployments, the EEG captures mu rhythm suppression to track attentional shifts while the user interacts with virtual avatars designed to encourage eye contact and conversational skills51. While the sensor artifact vulnerabilities (EMG/EOG noise) remain, the criterion validity is supported by the specific therapeutic goal: rather than punitively monitoring a classroom, the system provides gamified neurofeedback to help neurodivergent users practice social engagement in a low-pressure environment51. The institutional action is rehabilitative rather than evaluative, significantly altering the ethical and validity calculus.

9. Multimodal Inference and Border Control

The most ambitious—and highly contested—deployments of mental-state inference occur in border security, where developers attempt to overcome the limitations of single sensors by layering multiple modalities. The theoretical assumption is that multimodal systems capture a holistic picture of the human state, compensating for individual sensor weaknesses. However, mathematically, layering unvalidated proxy measurements often compounds false positive rates and exacerbates the base-rate fallacy.

9.1 Field Deployment 13: iBorderCtrl (EU Borders)

Funded by a €4.5 million grant from the European Union's Horizon 2020 program, the iBorderCtrl project aimed to create an intelligent portable control system to speed up border crossings for third-country nationals1. A core component was the Automatic Deception Detection System (ADDS), which required travelers to undergo a pre-travel video interview with a virtual avatar1. The system analyzed 38 input channels—specifically focusing on facial micro-gestures and eye data captured via a webcam—to generate a biomarker of deceit, scoring the traveler's probability of lying on a 100-point scale1. The system was piloted at land borders in Hungary, Latvia, and Greece9. The iBorderCtrl architecture was built upon the "Silent Talker" legacy system13. The field validity of this deployment was deeply compromised on almost every level of evaluation5. First, regarding label quality and calibration, the foundational algorithm was trained on a dataset of merely 32 to 39 individuals, disproportionately comprising white European males (22 White Europeans, 10 Asian/Arabic)5. Deploying an algorithm trained on such a minuscule, homogenous sample to evaluate the diverse, global population of travelers entering the EU ensures catastrophic subgroup burden and bias9. Second, the construct validity of "biomarkers of deceit" is scientifically void1. There is no empirical evidence that facial micro-expressions are reliable indicators of deception; individuals realistically motivated to be believed (such as innocent travelers fearing rejection or delays) routinely exhibit the same physiological signs of stress and cognitive load as individuals actively lying33. Third, the base-rate sensitivity of the deployment invalidates its operational utility. In a mass screening scenario where the prevalence of "non-bona fide" travelers is statistically tiny, an algorithm reporting 76% accuracy (the developers' claim) will flag vast numbers of innocent travelers as deceptive1. The institutional action—subjecting these false positives to intense secondary screening or potential denial of entry—transforms an algorithmic error into a severe infringement on fundamental rights, a reality that led to massive pushback from civil rights organizations and the eventual shelving of the technology for mass deployment56.

9.2 Field Deployment 14: AVATAR (North American Borders)

Similarly, the Automated Virtual Agent for Truth Assessments in Real-Time (AVATAR) was developed with funding from the US Department of Homeland Security and tested by Customs and Border Protection (CBP) and the Canada Border Services Agency (CBSA)33. The AVATAR kiosk interacted with travelers while simultaneously capturing a multimodal stream of data: facial expressions, eye movements, blink rates, vocal pitch changes, and posture32. Like iBorderCtrl, AVATAR relied on the assumption that deception leaks out through involuntary non-verbal cues32. Despite claims of detecting multiple deception cues within 45 seconds, independent scientific scrutiny confirms that human physiological responses to the stress of an authoritative border crossing cannot be reliably disentangled from the stress of deception33. Furthermore, the complexity of correlating multimodal data streams (e.g., matching pupil dilation with a specific vocal tremor) in an uncontrolled environment introduces massive noise. Layering flawed single modalities does not create a highly accurate multimodal system; rather, it creates a highly complex, opaque system that remains fundamentally unsuitable as a standalone arbiter of truth14.

10. Synthesis of Field Deployments and Validity Failures

The following table synthesizes the fourteen field deployments analyzed in this report, isolating the purported mental state, the technology utilized, and the primary point of validity failure.

11. Conclusion

The exhaustive evaluation of AI mental-state inference systems across global field deployments reveals a systemic epistemological crisis within the affective computing industry. Driven by the commercial demand for automated screening, developers routinely deploy sensor technologies that possess high physical accuracy but entirely lack foundational construct validity. The leap from measuring a physical correlate—be it an EEG alpha wave, an acoustic tremor, a dilated pupil, or a facial micro-gesture—to declaring the presence of an internal psychological state such as deception, malintent, or attention, is scientifically indefensible under current paradigms. When deployed in the field, these systems universally suffer from base-rate fallacies, severely penalizing truthful or innocent individuals in low-prevalence environments like borders and checkpoints. Furthermore, because these algorithms are often trained on small, homogenous datasets and rely on outdated, typological theories of emotion, they generate massive subgroup burdens, actively discriminating against marginalized, culturally diverse, and neurodivergent populations. The successful deployments of behavioral telemetry—such as Subaru's detection of the physical geometry of distracted driving, or BioCatch's continuous authentication against a user's own historical baseline—succeed precisely because they do not attempt to read the mind. They measure physical behaviors tied directly to physical outcomes. Until the affective computing industry acknowledges the profound complexity, context-dependency, and cultural variance of human cognition and emotion, systems claiming to infer mental states will remain highly sophisticated proxies for conformity, rather than objective arbiters of truth.

Works cited

1. The politics of deceptive borders: 'biomarkers of deceit' and the case of iBorderCtrl \- arXiv, https://arxiv.org/pdf/1911.09156

2. The protection of mental privacy in the area of neuroscience \- European Parliament, https://www.europarl.europa.eu/RegData/etudes/STUD/2024/757807/EPRS\_STU(2024)757807\_EN.pdf

3. High-Risk AI Systems—Lie Detection Application \- MDPI, https://www.mdpi.com/1999-5903/17/1/26

4. Context reconsidered: Complex signal ensembles, relational meaning and population thinking in psychological science \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC9683522/

5. Why Most Published iBorderCtrl Research Findings Are Likely to Be False, https://iborderctrl.no/blog:ioannides

6. Advancing towards Ubiquitous EEG, Correlation of In-Ear EEG with Forehead EEG \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC8879675/

7. Non-Invasive Brain-Computer Interfaces: How They Work Without Surgery \- Neuroba, https://www.neuroba.com/post/non-invasive-brain-computer-interfaces-how-they-work-without-surgery

8. What's Real? A Philosophy of Science for Social Psychology \- Open Publishing, https://openpublishing.princeton.edu/read/what-s-real-a-philosophy-of-science-for-social-psychology

9. AI Border Guards are Being Tested at the Edge of Fortress Europe, Away From Public Scrutiny | Novara Media, https://novaramedia.com/2019/12/04/15957/

10. Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC6640856/

11. Emotion Words, Emotion Concepts, and Emotional Development in Children: A Constructionist Hypothesis \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC6716622/

12. 226: Where emotions come from (and why it matters) Part 1 \- Your Parenting Mojo, https://yourparentingmojo.com/captivate-podcast/emotionspart1/

13. Silent talker: A new computer-based system for the analysis of facial cues to deception, https://www.researchgate.net/publication/227533391\_Silent\_talker\_A\_new\_computer-based\_system\_for\_the\_analysis\_of\_facial\_cues\_to\_deception

14. The politics of deceptive borders: 'biomarkers of ... \- Lie Detector Test, https://liedetectortest.com/polygraph-research/s-2022-politics-deceptive-borders-biomarkers/

15. Biased Face Recognition Technology Used by Government: A Problem for Liberal Democracy \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC8475322/

16. HireVue, Facing FTC Complaint From EPIC, Halts Use of Facial Recognition, https://epic.org/hirevue-facing-ftc-complaint-from-epic-halts-use-of-facial-recognition/

17. Government abandons lie detector tests for catching benefit cheats \- The Guardian, https://www.theguardian.com/science/2010/nov/09/lie-detector-tests-benefit-cheats

18. Technologists Are Creating Artificial Intelligence to Help Us Tap Into Our Humanity. Here's How (and Why). \- Entrepreneur, https://www.entrepreneur.com/science-technology/technologists-are-creating-artificial-intelligence-to-help/340207

19. EPIC Alert 28.01, https://archive.epic.org/alert/epic\_alert\_28.01.html

20. EPIC Files Complaint with FTC Regarding AI-Based Facial Scanning Software | Epstein Becker Green \- Workforce Bulletin, https://www.workforcebulletin.com/epic-files-complaint-with-ftc-regarding-ai-based-facial-scanning-software

21. AI Governance Mechanisms That Actually Work in Practice \- AI Career Pro, https://governance.aicareer.pro/blog/mechanisms-for-ai-governance

22. Automated Video Interviewing as the New Phrenology \- Carolina Law Scholarship Repository, https://scholarship.law.unc.edu/cgi/viewcontent.cgi?article=1617\&context=faculty\_publications

23. Hired by a Robot: The Legal Implications of Artificial Intelligence Video Interviews and Advocating for Greater Protection of Job Applicants \- Iowa Law Review, https://ilr.law.uiowa.edu/sites/ilr.law.uiowa.edu/files/2023-02/N2\_Kammerer.pdf

24. The truth is on the line | Technology \- The Guardian, https://www.theguardian.com/technology/2009/mar/12/voice-analysis-system-vra

25. BITN 752\_04,05 (News):BITN Iss 740, https://www.freedomineverydaylife.org/wp-content/uploads/BITN20752\_04.pdf

26. Government data shows £2.4m 'lie detection' didn't work in 4 of 7 trials \- The Guardian, https://www.theguardian.com/news/datablog/2009/mar/19/dwp-voice-risk-analysis-statistics

27. Charlatanry in forensic speech science: \- Columbia University, https://www.cs.columbia.edu/\~julia/papers/eriksson\&lacerda07.pdf

28. Forensic Language Analysis \- UK Parliament, http://researchbriefings.files.parliament.uk/documents/POST-PN-0509/POST-PN-0509.pdf

29. One-Voice Strategy for Customer Engagement \- faculty.weatherhead.case.edu, https://faculty.weatherhead.case.edu/jxs16/docs/SINGH-2020-JSR-ONE%20VOICE%20STRATEGY.pdf

30. Guide to AI: How Artificial Intelligence is Changing The World \- Calendar, https://www.calendar.com/artificial-intelligence/

31. Balto vs Cogito: Full-Suite Platform vs Verint Sub-Module 2026, https://www.balto.ai/compare/balto-vs-cogito/

32. Policing Project Five-Minute Biometrics Primers: Lie Detection II—21st Century Tests, https://www.policingproject.org/news-main/2020/9/4/policing-project-five-minute-biometrics-primers-lie-detection-ii-21st-century-tests

33. Deep Research of Deception Detection Systems Using Behavioral Modalities \- PolygrAI, https://polygr.ai/deep-research-of-deception-detection-systems-using-behavioral-modalities/

34. Using Non-Human means for Deception Detection | Forensic Interview Solutions, https://forensicinterviewsolutions.com/blogs/using-non-human-means-for-deception-detection/

35. Subaru Driver Monitoring System: Can I Disable It? \- Motor1.com, https://www.motor1.com/news/796893/subaru-driver-monitoring-system/

36. Exploring Monitoring Systems Data for Driver Distraction and Drowsiness Research \- MDPI, https://www.mdpi.com/1424-8220/20/14/3836

37. BioCatch Connect Reviews & Ratings 2026 | Gartner Peer Insights, https://www.gartner.com/reviews/product/biocatch-connect

38. What is behavioral biometrics? \- BioCatch, https://www.biocatch.com/blog/what-is-behavioral-biometrics

39. Something You Have vs. Something You Are: | BioCatch, https://www.biocatch.com/resources/data-sheets/something-you-have-vs.-something-you-are

40. The fraud gap in high-value, low-volume banking \- BioCatch, https://www.biocatch.com/blog/the-fraud-gap-in-high-value-low-volume-banking

41. Next-Generation Fraud Detection for Banks and Credit Unions \- Nasdaq Verafin, https://verafin.com/wp-content/uploads/2025/10/2025-1023-webinar-slides-biocatch-partnership-nasdaq-verafin.pdf

42. Future Attribute Screening Technology (FAST) Project 2008 \- Homeland Security, https://www.dhs.gov/sites/default/files/2025-07/privacy\_pia\_st\_fast.pdf

43. behavioral science and security: evaluating tsa's spot program hearing \- GovInfo.gov, https://www.govinfo.gov/content/pkg/CHRG-112hhrg65053/pdf/CHRG-112hhrg65053.pdf

44. Preemptive „Screening for Malintent“: The Future Attribute Screening Technology (FAST) as a Double Future Device | Request PDF \- ResearchGate, https://www.researchgate.net/publication/324642985\_Preemptive\_Screening\_for\_Malintent\_The\_Future\_Attribute\_Screening\_Technology\_FAST\_as\_a\_Double\_Future\_Device

45. Statement for the Record Bradley I. Buswell Deputy Under Secretary, Science and Technology Directorate U.S. Department of Homela, https://republicans-science.house.gov/\_cache/files/d/7/d7e88965-e7ea-4207-b72d-b14b3f9ad184/383F1D3D4F0C9A717A12262ED46540C5F59C0DD814ACD8C19E86BCD7D80E26A7.020310-buswell.pdf

46. 1 U.S. House of Representatives Committee on Science, Space, and Technology Subcommittee on Investigations & Oversight HEARI, https://science.house.gov/\_cache/files/0/c/0c947e48-6c5f-45d0-83d6-89e1e0e24bd7/663C902E793E393918F1B9B2C5679BBF3AB49B378E55FEBCFC665EABD20A2A0E.2011-04-04-spot-hearing-charter.pdf

47. Full article: Attention as an object of knowledge, intervention and valorisation: exploring data-driven neurotechnologies and imaginaries of intensified learning \- Taylor & Francis, https://www.tandfonline.com/doi/full/10.1080/17508487.2025.2469673

48. Exploring Neural Evidence of Attention in Classroom Environments: A Scoping Review, https://pmc.ncbi.nlm.nih.gov/articles/PMC12384261/

49. (PDF) Electroencephalography in naturalistic and semi‐naturalistic educational contexts: A systematic review \- ResearchGate, https://www.researchgate.net/publication/386166726\_Electroencephalography\_in\_naturalistic\_and\_semi-naturalistic\_educational\_contexts\_A\_systematic\_review

50. Smart Wearable EEG Devices: A Review of Lightweight, Multi-Sensor Systems for Sleep and Everyday Neurophysiology \- MDPI, https://www.mdpi.com/2079-6374/16/7/374

51. Non-Invasive BCI-VR Applied Protocols as Intervention Paradigms on School-Aged Subjects with ASD: A Systematic Review \- PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC11902634/

52. Neural personal information and its legal protection: evidence from China \- Oxford Academic, https://academic.oup.com/jlb/article/12/1/lsaf006/8113730

53. Non-Invasive BCI-VR Applied Protocols as Intervention Paradigms on School-Aged Subjects with ASD: A Systematic Review \- MDPI, https://www.mdpi.com/1424-8220/25/5/1342

54. Intelligent Portable Border Control System | H2020 | CORDIS | European Commission, https://cordis.europa.eu/project/id/700626/reporting

55. Intelligent Deception Detection through Machine Based Interviewing \- Manchester Metropolitan University, https://e-space.mmu.ac.uk/620503/1/WCCI2018-ADDS-pdf-convert.pdf

56. iBorderCtrl's FAQ, https://iborderctrl.no/iborderctrl\_s\_faq\_dissected

57. Why President Biden should ban affective computing in federal law enforcement | Brookings, https://www.brookings.edu/articles/why-president-biden-should-ban-affective-computing-in-federal-law-enforcement/

58. iBorderCtrl – Transparency complaint against secret EU surveillance research “iBorderCtrl” \- Patrick Breyer, https://www.patrick-breyer.de/en/posts/iborderctrl/

59. National Center for Border Security and Immigration Field Tests of an AVATAR Interviewing System for Trusted Traveler Applicants \- Eller College of Management, https://eller.arizona.edu/sites/default/files/FieldTestsofanAVATARInterviewingSystemforTrustedTravelerApplicants.pdf

60. Design and Evaluation of Biometric-enabled Interview Assisting Traveller Screening Technology \- à www.publications.gc.ca, https://publications.gc.ca/collections/collection\_2019/rddc-drdc/D68-2-223-2018-eng.pdf

61. Biometric-enabled Interview-Assisting Traveller Screening technology for Automated Border Control \- ResearchGate, https://www.researchgate.net/publication/379960066\_Biometric-enabled\_Interview-Assisting\_Traveller\_Screening\_technology\_for\_Automated\_Border\_Control

62. Assembly Bill Policy Committee Analysis \- AWS, https://billtexts.s3.amazonaws.com/ca/ca-analysishttps-leginfo-legislature-ca-gov-faces-billAnalysisClient-xhtml-bill-id-202520260AB1883-ca-analysis-394380.pdf