AI analysis of video interviews: is it allowed?

Short answer: AI that analyses video interviews is high-risk because it is used for recruitment and selection (Annex III, point 4). To the extent such a system infers emotions or mood, it also runs into a prohibition. The scientific basis is weak and the GDPR sets strict requirements. Deploying it without solid grounding is legally risky.

A growing number of vendors offer tools that automatically assess a video interview: they measure facial expression, vocal tone, word choice and pace, and infer "competencies" or "personality traits" from them. Under the AI Act and the GDPR several distinct, partly overlapping regimes apply.

Why this is high-risk

Recruitment and selection of staff is expressly listed as high-risk in Annex III, point 4 of the AI Act. A system that assesses, ranks or filters candidates on the basis of a video falls squarely within it. That triggers the full set of high-risk obligations: risk management, data quality and representativeness, technical documentation, logging, transparency to the deployer, meaningful human oversight and accuracy (Articles 9 to 15). The provider carries the bulk, but the employer as deployer also has its own duties, including use in line with the instructions and ensuring human oversight (Article 26).

The hard limit: emotion recognition

Here lies a second, sharper limit. Article 5 of the AI Act prohibits AI systems that infer the emotions of a natural person in the workplace, except for medical or safety reasons. Hiring interviews fall within that context. A tool that infers "enthusiasm", "stress" or "trustworthiness" from facial expression or vocal tone moves towards this prohibition. The ban has applied since 2 February 2025 and cannot be legitimised with consent or a legitimate interest. The line between prohibited emotion inference and (high-risk but permitted) competency assessment is narrow in practice; when in doubt, refraining is the safest route.

Questionable validity

Many claims about "personality analysis" from image and sound lack a solid scientific basis. The assumption that micro-expressions or speech patterns reliably predict suitability is contested. That directly engages the AI Act's accuracy requirement (Article 15) and the GDPR's accuracy and purpose-limitation requirements: a measurement that does not measure what it claims cannot be a lawful basis for a decision.

Bias and accessibility

Image and speech models demonstrably perform unequally across skin tone, accent, language proficiency and speech impairments. Candidates with a disability, a non-native language or an atypical presentation are systematically at risk of a lower score. That is not only an AI Act matter (representative data, Art. 10) but also a discrimination risk under equal-treatment law. See further AI and discrimination in recruitment.

GDPR: special-category data and Article 22

Facial and voice analysis can amount to biometric processing and special-category personal data (Article 9 GDPR), which is in principle prohibited subject to narrow exceptions. An automated rejection without meaningful human involvement engages Article 22 GDPR. Expect to run a DPIA, choose a sound legal basis, be transparent to the candidate and provide a genuine opportunity for human review. The broader framework is set out in AI in recruitment and HR.

What an employer must do

First establish whether the tool infers emotions; if so, do not deploy it. Determine whether the system is high-risk and require the vendor to supply the corresponding documentation, declaration of conformity and CE marking. Run a DPIA, choose a legal basis, inform candidates and ensure human oversight that can genuinely intervene. Test for bias and accessibility and retain the evidence. A tool that reads personality from a face is not an efficiency gain but a liability in the making.