What Passive Personality Sensing Approaches Exist Today
The passive sensing landscape spans several different modalities, each exploiting a different behavioural channel. A useful general introduction to the computational methods involved is available at Wikipedia: Natural language processing.
Language and text analysis. The most mature line of passive personality sensing draws on natural language processing to infer Big Five traits from written text. Researchers have found that lexical features — word choice, sentence structure, use of hedging language, frequency of first-person pronouns, emotional vocabulary — are reliably associated with personality traits. A 2015 meta-analysis by Park et al. found that language-based predictions of Big Five traits achieved average correlations with self-report scores of approximately .30 to .40 for Extraversion and Openness, somewhat lower for the remaining traits. These correlations are modest but consistent.
IBM's Personality Insights service (subsequently discontinued) was a commercial implementation of this approach, using the LIWC (Linguistic Inquiry and Word Count) framework and deep learning to infer Big Five profiles from samples of written text. Users could submit a block of text and receive a personality profile in return. The technical performance was roughly in line with academic benchmarks, but the service raised immediate questions about what users actually understood about what was being inferred from their words.
Speech and acoustic features. Voice carries personality information in ways that go beyond word choice. Studies have found that vocal features — speech rate, pitch variability, pause patterns, fluency — are associated with Big Five dimensions, particularly Extraversion and Neuroticism (Cèrcol's Presence and Depth). High-Presence speakers tend to speak faster, with less pausing and greater amplitude variation. High-Depth speakers show elevated pitch and more dysfluencies under conditions of mild stress.
Research by Mairesse et al. (2007) examined the personality recognition performance of acoustic models versus text-based models and found that both carried independent information. Combined models outperformed either modality alone, suggesting that future multimodal systems could achieve meaningfully better accuracy than current single-channel approaches.
Keystroke dynamics. How you type encodes personality information. Keystroke dynamics research examines inter-key intervals, error rates, typing rhythm, and correction behaviour. A study by Epp et al. (2011) found that keystroke features could predict Big Five dimensions at modest but above-chance accuracy, with the strongest signals for Extraversion and Neuroticism. The idea is that habitual typing patterns reflect underlying traits in the motor system — impulsive typists may show different timing profiles from methodical ones.
Smartphone and mobile sensing. Smartphones generate continuous streams of behavioural data: location patterns, call and message frequency, app usage, screen-on duration, movement patterns. Research by Chittaranjan et al. (2013) found that smartphone usage logs could predict Big Five traits at correlations of .22 to .38, with Conscientiousness best predicted by structured usage patterns and Neuroticism by call frequency and evening phone activity. These results parallel, but do not yet rival, the accuracy of validated self-report instruments. For how traditional self-report length affects measurement quality, see why 120 items is better than 10: personality test length.
What Passive Sensing Can and Cannot Predict at the Individual Level
The accuracy picture for passive sensing is nuanced. The correlations reported in academic studies — typically .20 to .40 — are real and replicable for Extraversion and Openness, which have clearer behavioural signatures. For Conscientiousness, Agreeableness, and Neuroticism, accuracy is lower and more variable across studies.
More importantly, these correlations are substantially smaller than those achieved by explicit questionnaire methods, which typically show Big Five internal reliabilities of .80 to .90 and test-retest correlations of .75 to .85 over shorter intervals. Passive sensing can produce usable group-level predictions in large-N research contexts, but at the individual level — the level that matters for any assessment of a specific person — the uncertainty is large.
The distinction between group and individual prediction is crucial and often elided in popular and commercial presentations of this technology. A model that achieves r = .35 between language features and Extraversion across a sample of 10,000 people does not thereby allow you to accurately classify any individual person as introverted or extraverted. The distribution of errors is wide. The model is right, on average, but wrong for many specific individuals. Understanding what "accuracy" means here is directly connected to the concepts of reliability and validity in personality testing.
The Invisible Inference Problem: Privacy and Consent in Passive Sensing
"The most significant risk of passive personality sensing is not that it performs well — it is that it performs well enough to be deployed at scale before the ethical frameworks exist to govern it."
When you complete a personality questionnaire, you know you are being assessed. You can decide what to disclose. You can choose not to participate. When a system infers your personality from your typing patterns, your voice, or your social media posts, none of these choices are available to you unless you have been explicitly told what is happening.
This asymmetry creates a serious informed consent problem. The consent required for passive sensing is more demanding than the consent required for self-report questionnaires, not less — because the inference is less transparent, the subject is less able to understand what is being inferred, and the scope for data repurposing is wider. A dataset of social media posts collected "for research purposes" can be reanalysed to infer personality dimensions that were not disclosed to participants at the time of data collection.
Regulatory frameworks — including GDPR in Europe — are beginning to catch up with this reality. Personality inference from behavioural data may constitute processing of personal data about personality characteristics, which attracts specific protections. But regulatory frameworks typically lag technology, and the practical enforcement of consent requirements in systems that infer rather than ask remains largely unresolved. The legal and ethical baseline for conventional personality assessment in employment is explored in personality testing in hiring: what is legal and what is ethical.
Training Data Bias: Why Passive Sensing Fails Underrepresented Groups
Passive sensing models are trained on datasets that reflect the populations from which they were collected. If those populations are not representative — if training data overrepresents certain demographics, languages, or cultural contexts — the resulting models will produce biased predictions for underrepresented groups.
The evidence on this point is concerning. Text-based personality models trained predominantly on English social media data perform substantially worse on text from users whose first language is not English, or who use the platform in culturally distinct ways. Acoustic models trained on Western English speakers do not transfer reliably to other languages and accent groups. The fairness problem in passive sensing is not a future risk — it is a present reality in every system that has been evaluated across demographic subgroups.
This matters for applications in hiring and HR specifically. A passive sensing system that underestimates the Conscientiousness of non-native speakers because their word-choice patterns do not match the training distribution is not just inaccurate — it is discriminatory in effect, even if not in intent. This is one reason the personality science replication crisis matters — researchers working on passive sensing face the same external validity problems that have troubled psychological science more broadly.
Why Cèrcol Chose Explicit Consent Over Passive Inference
Cèrcol uses self-report questionnaires and peer assessment from Witnesses — not passive sensing. Every participant knows what is being assessed, why, and who will see the results. This is not a technological limitation; it is a deliberate design choice that reflects the state of the evidence on passive sensing accuracy and the ethical requirements for meaningful informed consent. The full rationale for anonymity in personality assessment applies equally here: when participants control what is collected and who sees it, the data is both more ethical and more accurate.
The passive sensing research is scientifically interesting and will develop further. But the gap between what current systems can achieve at the individual level and what would be required for accurate, fair, and properly consented personality inference from passive data is substantial. Practitioners evaluating passive sensing tools should scrutinise individual-level accuracy claims, ask for demographic subgroup performance data, and insist on transparency about what is being inferred and how.
| Sensing modality | What it predicts | Typical accuracy (r with self-report) | Primary ethical concern |
|---|---|---|---|
| Language / text (NLP) | Extraversion, Openness most reliably | .30–.40 for E and O; lower for others | Invisible inference; data repurposing |
| Speech / acoustics | Extraversion, Neuroticism | .20–.35 | Continuous ambient data collection |
| Keystroke dynamics | Extraversion, Neuroticism | .15–.30 | Covert deployment; workplace surveillance |
| Smartphone usage logs | Conscientiousness, Neuroticism | .22–.38 | Location and activity tracking |
| Social media engagement | Extraversion, Openness | .25–.40 | Profile scraping without active consent |
How Cèrcol fits the future of personality assessment
As AI-driven inference tools multiply, the value of assessments with transparent, consent-based design will increase rather than decrease. Cèrcol uses the IPIP item pool — public-domain, peer-reviewed, and widely validated — combined with anonymous peer ratings through the Witness instrument. This gives you the accuracy of a validated multi-rater approach without the privacy trade-offs of passive inference. The science behind Cèrcol is fully disclosed, not proprietary. If you are evaluating personality assessment tools for your team and want to understand what the technology can and cannot responsibly do, the free assessment at cercol.team is a good place to start — both as a reference point and as a working example of what consent-respecting, evidence-based assessment actually looks like in practice.
Further reading: Personality testing in hiring: what is legal, what is ethical · What personality science cannot predict
Further reading
- Personality Testing: Open Source vs Commercial
- Why 120 Items Is Better Than 10: Personality Test Length
- Anonymity in Personality Assessment: Why It Matters
- Forced-Choice Assessment: What It Is and Why It Matters
- Personality Science: Evidence-Based HR — Why It Matters
- What Personality Science Cannot Predict