methodology & validity
What these numbers are — and what they are not.
This tool maps facial-analysis output to established psychometric models. Read this before acting on any card.
Which published research sources inform this methodology?
These links provide context for the research areas discussed here. They do not make an individual result definitive and do not override the stated limits on photo-derived personality inference.
Frontiers in Public Health
Identifying Big Five personality traits based on facial behavior analysis
Open published source
ACM Computing Surveys
Face-Based Automatic Personality Perception
Open published source
Nature Scientific Reports
Assessing the Big Five personality traits using real-life static facial images
Open published source
PubMed Central
The Influence of Each Facial Feature on How We Perceive and Interpret Human Faces
Open published sourceHow scores are computed
The normalized profile feeds established models — Big Five, HEXACO, Dark Triad, and EQ-i — via weighted averages of research-derived coefficients. Safety cards are computed on top of those models and the normalized metrics.
From photo to profile: the pipeline
Every image passes through the same fixed sequence: face detection, 68-point landmark mapping, pose-frontalization to a standard viewing angle, geometric feature extraction, and mapping onto established psychometric models. No human judges your photo, and the photo is deleted when processing finishes.
- Photo
- Face detection · 68 landmarks
- Frontalization — pose-normalizeoriginal image discarded here
- Geometric features
- Psychometric mapping
- Claim-bounded reportprobabilistic impression — not a diagnosis
Image quality gating — why we reject some photos
The system only analyzes near-frontal, evenly lit images. In our testing, badly-angled photos degrade the analysis sharply, so we discard them rather than guess. This is a control on input quality — not a judgment about the person.
Off-angle inputs are discarded, not guessed at.
The model stack
Landmark detection uses an ensemble of regression trees; frontalization uses a GAN-based pose model; visual features feed a ResNet-50 network with attention blocks, transferred from a large public face-recognition dataset. These are standard, published computer-vision components — conventional and inspectable engineering.
Psychometric mapping layer
ResNet-50 + attention · transfer learning
GAN frontalization
Landmark ensemble — regression trees
Input · 256×256, normalized
Psychometric mapping layer
ResNet-50 + attention · transfer learning
GAN frontalization
Landmark ensemble — regression trees
Input · 256×256, normalized
What we map onto — the psychometric models
Facial-analysis outputs are only ever expressed through four well-established psychometric frameworks: the Big Five, HEXACO, the Dark Triad, and EQ-i. These are validated by decades of questionnaire research; our contribution is a mapping layer, not a new theory of personality.
What a face can — and cannot — tell you
The link between facial appearance and stable personality is weak and contested in the scientific literature; treat every card as a probabilistic impression from an image, never a measurement or diagnosis. The effect sizes we cite describe trait-to-outcome links measured by questionnaires, not face-to-outcome links. Where the science is thin, we say so.
↑ our tool reports impressions here
Outputs are impressions, mapped through the strong left-hand evidence.
Data ethics & security
Raily uses a photograph as a temporary analysis input. After the live photo input is deleted, Raily retains only non-identifying derived scores and a random technical identifier, subject to the report choices and retention periods in our Privacy Policy. Raily does not verify a person's identity or maintain a mapping from that identifier to a real-world identity.
Safety, stability & predictability cards
Each card cites its evidence and shows a data-coverage indicator. Effect sizes below describe trait→outcome links measured by questionnaires — not face→outcome links.
Aggression & CWB
Berry, Ones & Sackett (2007)
Agreeableness ↔ interpersonal deviance r≈−.46; conscientiousness ↔ organizational deviance r≈−.42.
Integrity
Ones, Viswesvaran & Schmidt (1993)
Integrity ↔ job performance ρ≈.34, ↔ CWB ρ≈.32–.47; HEXACO H ↔ CWB r≈−.42.
Self-Control
Gottfredson & Hirschi (1990); Pratt & Cullen (2000)
Low self-control ↔ crime and deviance r≈.26–.28.
Manipulation (Dark Triad)
O'Boyle et al. (2012)
Mach/narcissism/psychopathy ↔ CWB r≈.23/.35/.32.
Reliability
Schmidt & Hunter (1998)
Conscientiousness ↔ job performance ρ≈.31 — strongest non-cognitive predictor.
Stress Tolerance
EQ-i / trait-EI meta-analyses
Stress management ↔ job performance ρ≈.25.
Scientific limits of inference from a face
Treat every output as a probabilistic impression from an image, not a measurement or diagnosis.
Boundaries of use — EU AI Act & GDPR
Legal classification depends on the feature, market and actual use. Raily does not verify identity, diagnose health, perform biometric identification or make consequential decisions.
This tool must not be used for:
- Clinical diagnosis
- Judicial or criminological conclusions
- Hiring, employment, or promotion decisions
- Individual "threat" assessment
It is intended as behavioural, claim-bounded insight for non-consequential, exploratory use.