AI analyzes candidates' emotions — true or false?

Emotion recognition AI relies on fragile correlations between facial expressions or vocal parameters and inner states — challenged by cross-cultural research (context, gender, disability).

In hiring, inferring joy, anger, or nervousness to score candidates is both unreliable and legally exposed. Responsible vendors explicitly exclude this in technical documentation.

The AI Act bans AI intended to detect or infer emotions in the workplace and educational settings — except narrow medical or safety exceptions.

Pre-screening for employment falls in this logic: exploiting vocal stress to eliminate candidates exposes deployer and provider. Confirm contract and product docs explicitly exclude this functionality.

• Banned: scoring on inferred stress, confidence, enthusiasm.
• Allowed: factual content assessment from transcribed answers.
• Gray zone to avoid: prosody or pace analysis as personality proxy.
• Require written vendor declaration of signals used.

HiLucy transcribes the voice interview and compares answers to recruiter-defined criteria: years of experience on a technology, project examples (STAR method), availability, mobility, stated salary expectations.

The score reflects coverage of job criteria in the transcript — not voice tone, accent, pace, or supposed emotion. If a candidate clearly states a skill, it is captured; if not mentioned, the criterion remains unmet.

Recruiters can read source excerpts and challenge the score — traceability is semantic, not psychological.

Speech-to-text converts spoken words to text. Analysis then targets that text: job keywords, example structure, answer completeness.

What HiLucy scoring excludes: vocal timbre, hesitations as lack of confidence, lie detection, Big Five personality typing, vocal attractiveness.

Before signing, send this questionnaire to any vendor — including when comparing HiLucy to alternatives.

• 1. What exact signals feed the score (text, raw audio, video)?
• 2. Do you use emotion inference or behavioral biometrics?
• 3. Can you provide the exhaustive list of variables excluded from scoring?
• 4. Can the candidate obtain a understandable explanation of their evaluation?
• 5. Where is audio hosted and how long is it retained?
• 6. Can a human modify or cancel the score — and is it logged?
• 7. Do you have AI Act conformity assessment / Annex III documentation?
• 8. What happens if we disable certain analysis features?

Request a demo with a test interview: confirm the summary cites text excerpts, not emotion labels (nervous candidate). Have your DPO validate the sheet.

If the vendor answer is vague, treat it as a red flag — not a commercial detail.