For a few years, a particular kind of hiring technology sold a remarkable promise: point an algorithm at a candidate's face or voice during an interview, and it would read their emotions, personality and even employability from micro-expressions and vocal tics. It felt like science fiction made real. As of 2 February 2025, in the European Union, much of it is simply illegal.
The EU AI Act's Article 5 prohibits, outright, the use of AI systems to infer the emotions of a person in the workplace and in education, except on narrow medical or safety grounds. This is not a "high-risk" category subject to paperwork and oversight — it is a banned practice, placed alongside the uses the EU considers an unacceptable threat to fundamental rights. Understanding why matters for anyone still relying on these tools.
What the Tools Claimed to Do
Emotion-recognition hiring systems typically analysed a recorded or live interview and generated inferences about the candidate from non-verbal signals: facial movements, tone of voice, speech tempo, word choice. From these, vendors claimed to measure traits such as enthusiasm, conscientiousness, emotional stability or "cultural fit", and to predict how well someone would perform in a role.
The appeal to overloaded recruiters was obvious. Here, apparently, was a way to screen video interviews at scale and quantify the intangibles that human interviewers judge subjectively. The problem is that the underlying science never supported the claims.
The Science Was Never There
The central assumption of emotion recognition — that inner emotional states map reliably and universally onto outward facial or vocal signals — is not well supported by evidence. A frown may signal concentration, not displeasure. A flat vocal tone may reflect culture, nerves, or simply how a person naturally speaks. The relationship between what a face does and what a person feels varies enormously across individuals, cultures and situations.
That variability is not a minor technical wrinkle; it is fatal to the entire premise. A system that reads "low enthusiasm" from a candidate who is simply reserved, or "instability" from someone whose expressions differ because they are neurodivergent or from a different cultural background, is not measuring employability. It is manufacturing a score from noise and attaching consequences to it.
This is why one of the best-known providers of video-interview assessment publicly dropped facial-expression analysis from its products back in January 2021, after sustained criticism. The company acknowledged that the visual component contributed little of value relative to the concerns it raised. The market had begun to reach the conclusion that regulators have now written into law.
Why Regulators Treated It as Unacceptable
The EU AI Act sorts AI uses by risk. Most recruitment AI — CV screening, ranking, matching — falls into the "high-risk" tier, permitted but heavily regulated. Emotion inference in the workplace sits a level above, among the prohibited practices, because it combines three aggravating features.
First, it rests on contested science, so the outputs are unreliable by design. Second, it operates on people in a position of relative powerlessness — candidates and employees rarely feel able to refuse. Third, it is intrusive in a way that reaches into a person's inner life, inferring feelings they never chose to disclose. The combination was judged an unacceptable risk to dignity and fundamental rights, not something that better documentation or a human reviewer could cure.
The penalties reflect that seriousness. Breaching the prohibitions in Article 5 can attract fines up to €35 million or 7% of global annual turnover, whichever is higher — the Act's most severe tier, above even the penalties for high-risk violations.
What Is Banned, and What Is Not
The line matters, so be precise. What is prohibited is using AI to infer emotions or emotional states of a person in workplace or educational settings from biometric data such as face or voice. An emotion-recognition system that scores a candidate's "enthusiasm" from their facial expressions or vocal tone in an interview falls squarely inside the ban.
What is not caught by this specific prohibition is analysis that does not infer emotion from biometric signals. Assessing the content of what a candidate actually said — the substance of their answers — is different from inferring their feelings from how their face moved. Text-based analysis of responses, for instance, is not emotion recognition from biometric data. But caution is warranted: many tools bundle features, and a system that also screens or ranks candidates remains high-risk and subject to the Act's full obligations even where it avoids the outright ban. The safest reading is that any product inferring a candidate's inner emotional state from their face or voice is now off-limits in EU hiring.
Beyond the EU
This is not solely a European concern. In the United States, Illinois requires employers who use AI to analyse video interviews to notify and obtain consent from candidates, explain how the AI works, limit who sees the recordings, and delete videos on request. New York City requires bias audits of automated employment decision tools. The regulatory direction of travel across jurisdictions is consistent: opaque inference about candidates from video and audio is being fenced in or shut down.
What Recruiters Should Do Now
If your hiring stack includes any tool that analyses candidates' faces or voices to infer emotion, personality or "fit", treat it as a live compliance problem, not a future one. Establish exactly what each tool infers and from what data — press vendors for a direct answer, in writing. Where a tool infers emotional or psychological state from biometric signals, the compliant response in the EU is to stop using that function.
More broadly, the demise of emotion recognition is a lesson in the cost of adopting AI on the strength of a compelling demo rather than credible evidence. The organisations now scrambling to rip these tools out are, in many cases, the ones that never asked the vendor to prove the science in the first place. The ones that did ask were spared both the wasted investment and the regulatory exposure.