Psychometric Surveillance

Definition

Psychometric Surveillance: [Established] Monitoring and profiling of cognitive and emotional traits for prediction and control. Zuboff-adjacent in surveillance capitalism.

Definitional Foundation

Psychometrics is the measurement of minds: personality, intelligence, emotional state, the traits psychology spent a century learning to quantify with consent, validated instruments, and clinical purpose. Psychometric surveillance is what happens when the measurement detaches from all three: traits inferred about people without their knowledge, from data they shed rather than answers they gave, by parties who intend to act on the results. The founding demonstration is the same study that anchors this dictionary’s cognitive dossiers entry: in 2013, Facebook Likes alone predicted sexual orientation, personality, intelligence, and substance use with accuracies that made the paper one of the most discussed of its year (Kosinski, Stillwell and Graepel, 2013). The authors wrote it as a privacy warning. Within five years, the methods had a scandal named after a company.

The term’s place in this lexicon (“Zuboff-adjacent,” per the short definition) is the supply chain. Surveillance capitalism’s prediction products need features, and psychological traits are the premium features: knowing a user is neurotic, impulsive, lonely, or fearful prices their attention and shapes what bends them (the surveillance capitalism entry documents the economics; the governmentality and paternalism entries document the bending). Psychometric surveillance is the refinery between raw behavioral surplus and the instrumentarian power Zuboff warned about.

One distinction and one demolition, both up front. The distinction: psychometrics proper (validated instruments, informed consent, clinical or research purpose) is legitimate science, and this entry does not touch it. The target is the surveillance form: inference without consent, validation, or appeal. The demolition: a large fraction of the surveillance form is pseudoscience, and this is documented at the highest evidentiary level. The definitive review of emotion recognition (the industry of reading feelings from faces) concluded that facial movements do not reliably indicate emotion: the same scowl variably means anger, concentration, or indigestion, across cultures, situations, and individuals (Barrett, Adolphs, Marsella, Martinez and Pollak, 2019). The European Union subsequently banned AI emotion inference in workplaces and schools, with the statute’s own recitals citing “serious concerns about the scientific basis” of such systems, their “limited reliability” and “lack of specificity” (EU AI Act, Article 5(1)(f); Recital 44). Hold both facts together and the entry’s central warning emerges: it does not have to work to hurt you. A pseudoscientific emotion score still costs the job interview it was consulted in. Invalid psychometrics, acted upon, produces valid harm.

Mechanism Analysis

Trait inference from residue. The Kosinski layer: models converting behavioral exhaust (likes, clicks, watch time, typing cadence) into psychological profiles. No question is ever asked; the subject’s participation is unnecessary and their refusal impossible.

Emotion recognition. The face-and-voice layer: systems claiming to read affect from expressions, tone, and micro-movements, sold for hiring, proctoring, retail, and policing, against the published science (Barrett et al., 2019). Its persistence after demolition is the tell that the market never needed validity, only verdicts.

Conversational psychometrics. The LLM-era layer, and the deepest. Chat is the richest psychometric instrument ever fielded: the cognitive dossiers entry documents the record it builds, and the annotation that rides on it. The deployed example is on this dictionary’s books already: classifiers that infer users’ emotional states and “reliance” from conversation, triggering interventions (the gaslighting/”>alignment gaslighting and biopolitics entries). Mental-state inference is no longer a research capability. It is a shipped feature, framed as care.

Institutional deployment. The verdicts land in decisions: hiring video analysis, exam proctoring flags, employee sentiment dashboards, classroom attention scoring. The EU’s targeted ban on workplace and education uses is the regulatory map of exactly where the deployment was thickest, which is also where subjects could least refuse: contexts where declining the scan means declining the job or the exam.

The control loop. Profiles feed back into environments: ads timed to inferred vulnerability, feeds tuned to inferred temperament, nudges matched to inferred compliance. This is Zuboff’s instrumentarian power with a psychological targeting layer, and it closes the definition’s circuit: monitoring for prediction and control.

Case Studies

From PNAS to scandal. The arc from the 2013 study to Cambridge Analytica is the canonical case. The harvesting is documented: data from up to 87 million Facebook profiles obtained via a personality-quiz app, exposed by the Guardian’s 2018 investigation, with regulatory action following (Cadwalladr and Graham-Harrison, 2018). Honesty requires the second half: scholars dispute how effective the firm’s “psychographic” targeting actually was, and the contested efficacy is part of this entry’s lesson rather than a defense. A company sold psychological targeting of electorates, buyers paid for it, and the data of tens of millions was taken to build it. Whether the weapon fired true matters less than the documented facts that it was built, bought, and aimed.

The science said no; the market shipped anyway. Emotion recognition’s trajectory is the cleanest specimen of verdict-over-validity. 2019: the field’s definitive review finds the premise unreliable. The years following: the industry grows regardless, scanning candidates and students. 2024-25: the EU prohibits the practice in work and education, citing the absent science, while vendors continue selling sentiment analysis into every unregulated market. At no point did the products start working. At every point their outputs were acted on.

The caring classifier. The AI safety apparatus documented across this dictionary is, viewed through this entry, psychometric surveillance at a scale with no obvious precedent: emotional-state inference running on every conversation across user bases the vendors themselves count in the hundreds of millions, generating interventions and records (the “emotional reliance” classifiers, the crisis-term counts disclosed in litigation). Its framing is protective, its consent is nonexistent, its validity is externally unaudited (the vendors consult clinicians; no independent party has validated the classifiers), and its verdicts (per the gaslighting and biopolitics entries) are unfalsifiable from below. The technique that was scandalous as advertising arrives uncontroversial as safety.

Systemic Context

The historical rhyme deserves naming: reading character from involuntary signs is physiognomy’s business model, and the machine-learning version inherits the old failure mode (confident classification of inner states from outer surfaces) with new math and planetary reach. Barrett’s review is the modern version of the demolitions phrenology eventually received; the difference is that phrenology was abandoned, while emotion AI was funded. The reason is structural: institutions want cheap verdicts about minds (who to hire, flag, insure, calm), and a pseudoscience that delivers confident verdicts serves the want better than an honest science that says “it depends.”

The asymmetry compounds every harm: subjects cannot see their scores, contest their classifications, or learn which decisions consulted them (the information asymmetry entry, adjacent in this cluster, treats the general condition). And the chilling completes it: people who know their affect is being read perform composure for the classifier, which is panoptic conditioning extended to the face itself: the inner life, managed for the meter.

Resistance & Mitigation

Cite the ban. The EU AI Act’s Article 5 prohibition is the proof of possibility: a major jurisdiction examined emotion recognition’s scientific basis and outlawed its highest-stakes uses. The precedent is exportable, and “the EU banned this citing its pseudoscience” is a sentence every procurement meeting deserves to hear.

Demand validity before deployment. The burden-shift this entry argues for: any system inferring psychological traits for consequential decisions should publish its validation against accepted scientific standards, or be treated as the coin-flip it is. Barrett’s review is the template for what honest evaluation looks like.

Refuse it in contracts. Employers, schools, and platforms choose these tools; unions, faculty senates, and customers can make the choice expensive. The workplace and classroom are where the EU drew the line, and where collective bargaining can draw it elsewhere.

Extend subject access to the inferences. Data rights typically cover what you provided; the fight is access to what was concluded: the trait scores, emotion flags, and risk classifications attached to your record (the cognitive dossiers entry’s annotation layer). What the machine thinks you are should be readable by you.

Keep the distinction sharp. Resistance gains nothing from sloppiness: validated psychometrics with consent is medicine and science; inference without consent is surveillance; inference without validity is divination with a dashboard. Name each accurately, and reserve the full force for the last two, where it belongs.

Annotated Bibliography

Barrett, Lisa Feldman, Ralph Adolphs, Stacy Marsella, Aleix M. Martinez, and Seth D. Pollak. “Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements.” Psychological Science in the Public Interest 20, no. 1 (2019). https://journals.sagepub.com/doi/full/10.1177/1529100619832930
The definitive scientific review: facial movements do not reliably indicate emotional states. The demolition the emotion-recognition industry survived commercially while losing empirically.

Cadwalladr, Carole and Emma Graham-Harrison. “Revealed: 50 million Facebook profiles harvested for Cambridge Analytica in major data breach.” The Guardian (March 2018).
The investigation that exposed psychometric surveillance’s flagship scandal; the affected-profile estimate later rose to 87 million.

EU Artificial Intelligence Act, Article 5(1)(f) and Recital 44. https://artificialintelligenceact.eu/article/5/
The statutory prohibition of emotion inference in workplaces and education, with the legislature’s own finding of absent scientific basis. The regulatory precedent this entry’s resistance section is built on.

Kosinski, Michal, David Stillwell, and Thore Graepel. “Private traits and attributes are predictable from digital records of human behavior.” PNAS 110, no. 15 (2013). https://www.pnas.org/doi/10.1073/pnas.1218772110
The founding demonstration of trait inference from behavioral residue, written as a warning and read as a roadmap.

Zuboff, Shoshana. The Age of Surveillance Capitalism (2019).
The economic frame: psychological profiles as premium features in prediction products. The adjacency the short definition names.

Dictionary of Digital Oppression, version 0.2.