Trusq

factual analysis · traceable to primary sources

Analysis

AI sentiment analysis of employees: the thin line to the emotion ban

Adopted 2026-06-22 ยท ≈ 2 min read ยท Dirk Baaijen

AI inferring employee mood from email, chat, surveys or speech brushes against the emotion-recognition ban (Art. 5 AI Act) and the GDPR. Aggregated and anonymous is sometimes possible; individual monitoring almost never.

Short answer: AI that infers the sentiment or mood of employees from email, chat, survey text or speech moves along a thin line. If it targets individual emotions in the workplace, it touches the emotion-recognition ban in the AI Act. And the GDPR always applies: proportionality, confidentiality of communications and possibly special-category data. Aggregated and anonymous is sometimes possible; individual monitoring almost never.

Sentiment is not the same as emotion โ€” but the line blurs

Sentiment analysis classifies language as positive, neutral or negative. Emotion recognition infers affective states (angry, stressed, unhappy). The AI Act prohibits the second in the workplace. Many "sentiment" tools, however, shift towards emotion as soon as they interpret vocal intonation, stress signals or individual moods. Then the "sentiment" label is cosmetic and the application falls under the ban.

The emotion ban (Art. 5 AI Act)

In the workplace, inferring people's emotions through AI is prohibited, with narrow exceptions for medical or safety purposes. A tool that measures per employee how stressed or dissatisfied someone sounds is at its core caught by this โ€” regardless of the good intention of "wellbeing" or "engagement". The ban is a hard limit, not a balancing test.

GDPR: confidentiality and proportionality

Email and chat contain confidential communications. Having them structurally scanned for mood is a far-reaching intrusion that is rarely proportionate: the aim (a pleasanter working climate) is out of proportion to permanently reading along. Speech and text analysis can moreover reveal health or belief data โ€” special-category data under Art. 9 GDPR, with a prohibition regime. After-the-fact transparency does not cure that.

What is allowed: aggregated and anonymous

An anonymous, aggregated pulse โ€” for instance a voluntary survey of which only team totals are visible, with no traceability and no individual scores โ€” can be legitimate. Conditions: genuinely anonymous (sufficient group size), voluntary, transparent purpose, no link to persons or decisions. As soon as you can trace it to the individual or attach consequences, it tips over.

Co-determination is not a formality

Monitoring employees and processing personal data engages the works council's right of consent. Introducing sentiment or mood analysis without consent is not only unwise but often unlawful. The works council is the first, not the last, step here. See also AI employee monitoring.

What to do

  • Do not track individual mood from communications or speech โ€” this touches the emotion ban.
  • Limit yourself to voluntary, anonymous, aggregated measurements without traceability.
  • Test in advance whether the tool in fact infers emotions; the "sentiment" label offers no protection.
  • Seek works-council consent and be fully transparent about purpose and method.

An organisation that wants to know how its people are doing can ask โ€” voluntarily, anonymously, in dialogue. An algorithm that reads everyone's mood from their words unasked takes exactly the route the legislator has closed off.

Sources

  1. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
    Regulation (EU) 2024/1689 (AI Act): Art. 5 prohibits emotion recognition in the workplace, save medical/safety purposes.
  2. https://eur-lex.europa.eu/eli/reg/2016/679/oj
    Regulation (EU) 2016/679 (GDPR): proportionality, confidentiality of communications and the regime for special-category data (Art. 9).

Share on LinkedIn

Read next

A

AI analysis of video interviews: is it allowed?

Video AI that scores face, voice or "personality" is high-risk under the AI Act. Inferring emotions in the workplace is in addition prohibited. Its validity is questionable and the GDPR bar is high.

U

AI, trade secrets and confidentiality

Feeding confidential information into an external AI model can undermine trade-secret status and breach confidentiality or GDPR obligations. Protection depends on secrecy measures; uncontrolled sharing erodes them. Manage it with policy, contract and access rules.

U

AI predicting sickness absence: almost always a legal no-go

AI predicting which workers will fall ill or be absent touches health data โ€” special-category personal data with a strict prohibition regime (Art. 9 GDPR). The risk of discriminating against sick or disabled workers is high. In most cases such prediction is not lawful.

Dirk Baaijen

About this knowledge base

Compiled and maintained by YRproject โ€” programme and project direction at the intersection of digital transformation, AI and regulation. Every factual claim is traceable to its primary source. YRproject is led by Dirk Baaijen About & method โ†’

A project or programme? Work with YRproject โ†’

The monthly briefing

AI regulation in five minutes: what changed, what is coming and what it means. No spam, unsubscribe anytime.

Your address is used for this only and stored on our own servers.