A Hospital Director Considers AI That Measures Staff Emotions: Development or Surveillance?

Facial expressions and emotions

A camera captures movements in an employee’s face, but it does not directly show what that person is feeling. A review of research on emotional expression challenges the assumption that each emotion has one unmistakable facial pattern. This does not mean faces never convey information: an expression may be meaningful in the context of a conversation and situation. But describing a visible smile is different from treating it as reliable evidence of satisfaction or calm. Likewise, facial tension alone cannot establish the inner state of someone on shift. When a hospital considers automated emotion assessment, this distinction matters in practice: an uncertain inference can easily be mistaken for a measurement.

What is an emotion recognition system?

Before assessing a tool, it is worth asking the supplier what data it analyses and exactly what it infers from them. A system that infers an employee’s emotions from facial expressions, voice or other biometric data must be distinguished from a survey in which the person describes their own experience at work. An ordinary questionnaire should not be called biometric emotion recognition simply because it asks about well-being. How the results are actually used also matters: are they intended to support a development conversation or to assess staff behaviour? The definition of the system matters legally, but classification alone does not establish whether its measurements are valid. A voluntary survey also needs its own assessment, including its purpose and whether participation is genuinely optional.

The boundary set by Article 5

Article 5(1)(f) of the EU AI Act prohibits certain uses of AI systems that infer people’s emotions in workplaces and educational institutions. It provides exceptions for medical or safety purposes. This does not mean every HR tool is illegal, nor does a hospital automatically qualify for the medical exception. Before buying a system, the hospital needs to establish whether the model infers an employee’s internal state from biometric data, how it works in the proposed setting and what its actual purpose is. Only then can it assess how the provision applies and whether an exception may be invoked. The provision sets a legal boundary; it does not establish that the tool works.

The effect of power imbalances

Even when participation is presented as voluntary, the relationship between management and employees can make it difficult to refuse. A hospital should therefore ask directly whether opting out truly has no consequences, who sees the results and why. Data collected for development should not become a covert staff assessment. If a manager interprets behaviour through a presumed mood, an employee may decide it is safer to conceal a problem than to report insufficient resources on a shift. This risk should be considered before implementation, not only after concerns arise within the team. A clear purpose for the tool also matters for trust in conversations about needs.

Emotions and professional work

When working with patients, staff often regulate how they express emotions because the professional situation calls for it. A meta-analysis of research on emotional labour distinguishes between putting on a particular expression and more deeply regulating one’s feelings; however, it covers various professions and primarily examines associations, so it does not establish causes for every hospital shift. A polite smile may therefore be part of professional conduct rather than a straightforward sign of satisfaction. Equally, a tired face should not be taken as evidence of a lack of empathy or poor care. When discussing work quality, it is better to separate what can be observed in someone’s actions from assumptions about their mental state. Facial expressions alone do not provide reliable access to what a person is feeling.

An alternative to digital surveillance

If the hospital’s goal is development, it could start by discussing obstacles encountered on shifts rather than analysing faces. Staff can voluntarily describe their needs and provide feedback, after which management can examine specific working conditions. This approach also needs rules: data collection should be limited to the stated purpose, the interpretation of results should be explained, and individuals should be able to see how their results are interpreted. Before implementation, it is worth establishing who can see individual results, what reaches management and how employees learn the rules for using the tool. A questionnaire is not the same as biometric emotion inference, but its voluntary nature and how it is used must be more than assurances on paper.

Questions to ask before implementation

An implementation decision should begin by documenting the system’s purpose, the types of data it uses and how the model actually works. The hospital must then check its classification, the current wording of the law and the legal basis for the planned use in light of the specific application. Validity is a separate question: is there evidence that the proposed measure works in the population where it will be used? Neither a supplier’s promise nor the mere fact that a solution uses AI is enough. The hospital also needs to determine who will have access to the results, how long they will be stored and how to prevent their reuse for other purposes. Before buying, it is worth comparing the plan with a less intrusive way to understand staff needs that does not involve assessing their expressions.

A staff member’s smile does not prove they are satisfied, nor does a tired face show a lack of empathy. Before buying AI to analyse emotions, a hospital should examine how the system works, its purpose and its effects on staff.

Empatyzer and staff skills development with respect for privacy

A hospital director who wants to encourage staff to develop their communication skills can start with conversations about collaboration rather than measuring facial expressions. In Empatyzer, employees can see their own results in the “About me” view and choose which ones to discuss with Em, for example when preparing for a specific conversation. Management can use an aggregate picture of the team to choose a meeting topic, while team mode supports discussions about shared patterns of collaboration and development needs. Micro-lessons offer short exercises in conversation skills during everyday work. A manager can also use Em to prepare a conversation about employees’ needs and practise responding to a report of excessive workload, without judging the other person’s face. This use still requires a clear explanation to staff of its purpose, who can see the results and the rules for participation. The aggregate picture should not become a pretext for guessing how any individual feels; the starting point remains what employees want to say about their work.