01 — Two footprints
Likes describe chosen content; the mouse describes movement dynamics.
A Like is a discrete choice about a cultural object, brand, activity or opinion. Accumulated choices create a preference profile. They are socially meaningful but depend on the platform, historical moment and what users are willing to make visible.
The mouse produces a continuous stream of positions and timestamps. A model can extract initial speed, acceleration, deviation, target overshoot and corrections. This trace strongly depends on task, hardware and motor familiarity.
02 — Youyou et al.
Around 90 to 100 Likes reached the correlation of an average acquaintance’s judgment.
The study, published in the Proceedings of the National Academy of Sciences (PNAS), involved 86,220 volunteers who completed a 100-item personality questionnaire and shared their Facebook Likes. The authors trained models to predict self-rated Big Five scores and compared the resulting correlations with judgments made by participants’ acquaintances.
At the sample average of 227 Likes, computer judgments correlated r = 0.56 with self-ratings, compared with r = 0.49 for an average human judge. The published curve suggests roughly 90 to 100 Likes reached that average human level. This result measures statistical agreement with a personality questionnaire.
03 — Zhao et al.
Cursor features can be associated with personality preferences.
Zhao et al. collected mouse operations from 146 participants and constructed kinematic and adjustment features. They examined preferences from the Myers–Briggs Type Indicator (MBTI), a typology covering preferred ways of gathering information and making decisions.
Reported recognition accuracy ranged from 60.6 to 78.3% across preferences. Participants preferring global information showed faster initial acceleration, while those favoring analytical criteria showed less deviation and target overshoot in this protocol. The MBTI classifies preferences and has less scientific consensus than the continuous Big Five model, so the two frameworks require different interpretations.
04 — What the model measures
Understand what the prediction actually represents.
In both studies, the model target comes from a psychometric instrument. The system therefore learns to partially reproduce variation in that questionnaire from a trace. Performance depends on test reliability, sample composition, data representation and validation design.
Likes and mouse behavior may also encode confounds such as age, culture, platform familiarity, equipment, social context or task goals. Validation with new groups and uses measures the prediction’s actual scope.
05 — Responsible design
Define purpose, consent and limits before deployment.
This work supports legitimate uses: psychological research, explicitly chosen adaptive interfaces, friction detection and transparent personalization. It also enables invisible profiling, microtargeting and decisions that people cannot contest.
Responsible applications should minimize data, explain inference, offer a genuine opt-out, measure errors across populations and prohibit disproportionate uses. The more ordinary a trace appears, the more important it is to remember that combined micro-signals may reveal more than a person intended to share.