A dividing line separating biometric face and voice emotion-recognition signals on the regulated side from a stream of written customer feedback text sorted into themes on the clear side, framed by an EU regulatory border.

Does the EU AI Act's Ban on Emotion Recognition Affect Customer Sentiment Analysis?

The EU AI Act bans certain emotion-recognition systems. Here is why text-based customer sentiment analysis generally falls outside the ban, and where the boundary gets subtle.

Insights
>
>
Does the EU AI Act's Ban on Emotion Recognition Affect Customer Sentiment Analysis?
While you're here

TLDR

The EU AI Act's emotion-recognition prohibition is narrow: it covers emotion inference from biometric data (faces, voiceprints, physiological signals) in workplace and education settings. The European Commission has confirmed that inferring emotions from written text, including content and sentiment analysis, is not based on biometric data and falls outside the ban. Text-based customer sentiment analysis is therefore generally out of scope, though voice-biometric analysis and employee monitoring are boundary cases to watch.

A new clause in the EU AI Act bans certain "emotion recognition" systems, and the phrase has set off alarms on CX and Insights teams that run sentiment analysis on customer feedback. If the law bans inferring emotions, does it ban reading whether customers sound happy or frustrated in a survey?

In almost all cases, no. The EU AI Act's emotion-recognition prohibition is narrow on two axes at once. It applies only to systems that infer emotions from biometric data, such as faces, voiceprints, or physiological signals, and only in the workplace or in education settings. Analyzing the text customers write in surveys, reviews, and support tickets is neither of those things. The European Commission's own guidance says so directly. Thematic analyzes written customer feedback as text, not biometric data, which places this kind of sentiment analysis outside the emotion-recognition rules.

This article explains what the ban actually covers, why text-based customer sentiment analysis sits outside it, and where the boundary gets subtle enough to check. It is informational, not legal advice. For a specific deployment, confirm with your own counsel.

What the EU AI Act actually prohibits

The relevant clause is Article 5(1)(f), one of the Act's prohibited practices. It bans putting on the market or using AI systems "to infer emotions of a natural person in the areas of workplace and education institutions," with a narrow exception for medical or safety reasons.

Two definitions decide what gets caught:

  • It must be an "emotion recognition system." Article 3(39) defines this as an AI system that identifies or infers emotions or intentions of natural persons "on the basis of their biometric data." The biometric basis is the hinge. No biometric data, no emotion recognition system.
  • Biometric data is physical, not textual. Article 3(34) defines biometric data as personal data from technical processing of someone's physical, physiological, or behavioral characteristics, such as facial images or fingerprints. Plain written text is not biometric data.

The Act's Recital 18 adds detail. The emotions in scope are states like happiness, sadness, anger, surprise, disgust, and satisfaction. It explicitly excludes physical states such as pain or fatigue, and it excludes the mere detection of obvious expressions like a smile or a raised voice unless the system uses them to infer an emotion.

The prohibition has applied since February 2, 2025, when the Act's banned-practice rules took effect. The Act itself entered into force on August 1, 2024.

Why text-based customer sentiment analysis is generally outside the ban

Customer sentiment analysis on written feedback misses the prohibition on both required axes.

It is not based on biometric data. Survey verbatims, app reviews, support tickets, and chat messages are text. The European Commission's February 2025 Guidelines on Prohibited AI Practices address this case by name. The guidance states that an AI system inferring emotions from written text, including content and sentiment analysis, "is not based on biometric data and therefore does not fall within the scope of the prohibition." That is about as direct an answer as a regulator gives.

It is not in a workplace or education setting. The ban is scoped to emotion inference about employees and students. Analyzing how customers feel about a product or a service is a commercial context, not the employment or education relationship the clause is written to protect.

Two practical points follow for enterprise teams:

  • Reading sentiment and themes from customer feedback text is generally not the regulated activity, so it is neither banned nor classed as a high-risk emotion-recognition system under the Act's Annex III, which also turns on the biometric definition.
  • This is a carve-out, not an endorsement. Some academics and observers have called the text exemption a gap in the law. Treat it as a reason to keep your analysis governed and explainable, not a reason to stop caring.

A separate law still applies regardless: the General Data Protection Regulation (GDPR) governs any personal data inside customer feedback, whatever the AI Act says about emotion recognition.

The boundary cases worth knowing

The distinction is clean in the middle and fuzzier at the edges. These are the cases a CX or Insights leader should scope carefully.

  • Voice versus transcript. Analyzing a call-center transcript as text is text analysis. Running emotion AI on the voice recording itself, using vocal biometrics, can make the system biometric and bring it back toward the emotion-recognition definition. The input format matters, not just the goal.
  • Customer versus employee. A tool reading customer feedback is in a commercial context. The same kind of emotion inference pointed at employees, for example monitoring staff during calls, is exactly the workplace use the ban targets if it runs on biometric data.
  • Face and physiology. Facial-expression analysis from video, or inference from physiological signals, is squarely the regulated category. This is the technology the clause was written about.

If a tool only ever reads the words people wrote, it stays on the text side of the line. The moment it processes faces, voiceprints, or body signals to infer feeling, the analysis changes category.

How Thematic stays on the right side of this

Thematic analyzes written customer feedback, and it is built to be inspected. Both facts matter for a team that has to defend its tooling to a security review or a regulator.

It works on text, not biometrics. Thematic analyzes surveys, support tickets, call-center transcripts, app reviews, and CRM notes as written language. It does not run facial-expression analysis or voice-biometric emotion detection. Its sentiment and themes come from the words customers chose, which is the activity the Commission guidance places outside the emotion-recognition prohibition.

Every theme traces back to the words behind it. In Thematic, a theme maps to the specific customer phrases that created it, with an audit trail of how it was identified and refined. An analyst can open any theme and see the source comments. That traceability is what makes the analysis defensible in an executive or compliance review.

A human stays in the loop. Thematic's themes can be reviewed, edited, and validated rather than accepted from a black box. Governance over how feedback is classified stays with your team.

The data posture is documented. Thematic is SOC 2 Type II, GDPR, and CCPA compliant, and customer data is not used to train models. That posture supports a procurement conversation, though it is not a substitute for your own legal assessment.

Atom Bank, the UK app-based bank, is one example of a regulated-sector team using Thematic on multi-channel written feedback. The bank unified feedback across channels into a single governed source and used it to cut call-center volume, all from analyzing what customers wrote, not biometric signals.

A compliance-minded checklist for CX and Insights teams

Before you assume a feedback tool is in or out of scope, ask:

  1. What is the input? Written text is outside the emotion-recognition definition. Faces, voiceprints, and physiological signals are not.
  2. Whose emotions, and in what setting? Customer feedback is commercial. Employee or student monitoring is the workplace and education context the ban targets.
  3. If you analyze calls, are you reading transcripts or processing the voice itself?
  4. Can the vendor trace any sentiment or theme back to the exact words behind it?
  5. Can your team inspect and override how feedback is classified?
  6. Does the vendor document its data handling, including whether your data trains their models?

If the input is text, the subject is customers, and the analysis is traceable, you are almost certainly outside the emotion-recognition rules and on solid governance footing.

The short answer

The EU AI Act's emotion-recognition ban does not generally affect customer sentiment analysis on written feedback. The prohibition targets emotion inference from biometric data in workplace and education settings, and the European Commission has confirmed that inferring emotions from written text is not based on biometric data and falls outside the ban. Thematic analyzes the words customers write, traces every theme back to those words, and keeps a human in the loop, which is the kind of governed, text-based analysis the rules leave alone. This is informational, not legal advice. Confirm any specific deployment with your own counsel, and remember the GDPR still applies to the personal data in your feedback regardless.

1. Guide Analysis
Guides

Build, Buy or Partner? A Layered Guide to AI Feedback Analytics

Transforming customer feedback with AI holds immense potential, but many organizations stumble into unexpected challenges.

/* Top banner pulsating dot */