Trust is part of the integration.
Plan consent, governance, review quality, and uncertainty before building around recorded facial-expression signals.
Consent and purpose
Document why media is processed, who sees results, and how participants can ask questions or withdraw.
Data handling
Set retention and deletion rules before launch. Minimize raw media, restrict access, and test the unhappy path.
Quality and uncertainty
Validate across your capture conditions. Surface low quality and uncertainty instead of turning a score into a verdict.
Versioning and service
Confirm supported model and schema versioning, completion and retry behavior, service expectations, and incident communication before launch.
Observable movement is not a person’s inner state.
Mental Edge does not determine emotion, thought, intent, truthfulness, health, engagement, burnout, or suitability. Do not use it for surveillance, deception detection, patient diagnosis, sensitive-attribute inference, hiring, student scoring, or athlete selection.
Turn questions into written responsibilities
Identify who obtains consent, answers participant questions, reviews footage, and handles deletion requests in your organization. Decide what information each role actually needs. Set retention rules before collecting media, rather than keeping footage indefinitely for possible future use. These are implementation responsibilities, not promises of specific API privacy controls.
Use a narrow purpose statement that names the visible observation being reviewed. Mental Edge scores recorded facial expressions frame by frame and associates scores with seven expression categories. Those categories do not establish thoughts, feelings, intent, truthfulness, or suitability. Preserve the source context instead of turning a label into a personal assessment.
Keep a written record of questions that remain unanswered after review. A release decision should not hide unresolved interpretation problems.
Check evidence and legal scope
Barrett and colleagues describe how expressions vary across cultures, situations, and people. Build an evaluation around held-out clips and independent reviewers, not a favorable demonstration. The Observer XT supports Cohen's kappa for inter-rater reliability. That evaluates agreement within a coding scheme, not the truth of an emotional interpretation.
EU AI Act Article 5(1)(f) prohibits emotion-inference AI in workplaces and educational institutions, except for medical or safety reasons. The prohibitions apply from February 2, 2025. A consent form or a different product label does not by itself establish permission. Talk to counsel about your use case. This is not legal advice.