One year of banned AI practices: where an employer can cross the line without meaning to
· By Bee Flow · 5 min read
Since 2 February 2025 some uses of AI are banned in the EU. For employers, emotion recognition is the ban that comes closest. What the European Commission's guidelines say, with their own examples, and what is simply fine.

Today, 2 February 2026, the prohibited AI practices in Article 5 of the AI Act have applied for exactly one year. On 4 February 2025 the European Commission published guidelines that explain, ban by ban, where the line runs, with a lot of examples. They are not binding: the final word belongs to the EU Court of Justice. The penalty provisions have applied since 2 August 2025, and a breach of Article 5 sits in the heaviest category, according to the same guidelines.
Most of Article 5 has nothing to do with a small business. Real-time face recognition by the police, predicting who will commit a crime: that happens elsewhere. Three of the bans can reach an ordinary employer, though, and usually through software that was simply bought in. They are emotion recognition at work, social scoring and manipulation. This piece goes through them using the Commission's own examples. It is not legal advice.
Emotion recognition is the one closest to HR
Article 5(1)(f) bans AI systems that infer the emotions of people in the workplace and in education, except for medical or safety reasons. Two things decide whether a tool falls under it.
The first is the source. According to the guidelines, the ban only covers emotions inferred from biometric data: face, voice, posture, even the way someone types. A system that judges the tone of a written text is outside this ban.
The second is the place, and it is wide. The Commission reads workplace as any physical or virtual space where someone does their work, so the shop floor, the delivery van and a video call from the kitchen table all count. Job applicants are covered too, because the imbalance of power already starts at recruitment.
The Commission gives these examples. A call centre that uses webcams and voice recognition to track whether staff are angry: prohibited. Software that reads the emotional tone of hybrid teams from the voice and video of their calls, even when the aim is to prevent conflict: prohibited. Emotion recognition during recruitment or the probationary period: prohibited. A call centre that tracks the emotions of its customers to help staff deal with angry callers: not prohibited under this article. Use for personal training is allowed, as long as the results do not reach HR and have no effect on appraisal or promotion.
The exception for medical and safety reasons is narrow. General monitoring of stress at work does not qualify, and nor does a system meant to detect burnout or depression in staff. Measuring a driver's fatigue is not emotion recognition at all, because fatigue is a physical state.
One more point for whoever buys the software. The ban applies to the employer that uses the system, even when the vendor has excluded that use in its terms.
In July 2025 the Dutch data protection authority (AP) called emotion recognition dubious and risky (in Dutch). Many systems translate voice, facial expression or heart rate into emotions without nuance, the regulator wrote. A high heart rate is not always fear, and a raised voice is not always anger, said its chair Aleid Wolfsen.
What is fine
An AI chat that helps write a job advert or a difficult e-mail infers nothing about the emotions of the person using it. A transcription tool that turns a meeting into text and a summary does not either. Even analysing the tone of a written text falls outside the ban, according to the guidelines, because no biometric data is involved. Noting that someone is smiling is not emotion recognition, the Commission writes. Concluding that someone is happy is.
The GDPR still applies to these tools. You need to know where the text and recordings go and who can see them. They do not touch the Article 5 ban.
What deserves a second look
The risk sits mostly in features that come as an extra inside software you already have. A few places to check:
- video interview software that scores candidates on enthusiasm, confidence or engagement from their face or voice
- call quality software for customer service that reads each agent's mood or irritation from their voice
- wellbeing apps and wearables that measure staff stress and report it to the employer
- meeting tools with a sentiment feature, where the question is whether it works from the text or from the voices
Ask the vendor whether the product infers emotions or intentions, from which data it does so, and whether you can switch that feature off. Then check that it really is off in your account.

If your company has a works council, Dutch law already asks something separate from the AI Act. A scheme for facilities suited to monitoring the attendance, conduct or performance of staff needs the works council's consent (Article 27 of the Works Councils Act, in Dutch). The guidelines also point out that employers have to consult workers or their representatives when they introduce this kind of technology.
Social scoring and manipulation are further away
Social scoring sounds like something governments do, and most examples in the guidelines are about them. The ban still applies to companies. It covers AI that evaluates people over a period of time on their social behaviour or personal characteristics, where the score then leads to worse treatment in a context unrelated to where the data came from, or to treatment out of proportion. All conditions have to be met at once.
The Commission explicitly names specific employee evaluations as legitimate and not prohibited as such. The line comes into view when work and private life start to mix. As an example of a score that spills into another context, the guidelines mention job opportunities that depend on someone's social media activity. So keep an appraisal to data about the work itself.
Manipulation is prohibited when AI uses subliminal, purposefully manipulative or deceptive techniques that materially distort someone's behaviour and cause significant harm, or are reasonably likely to. Persuasion is allowed, as long as it happens in the open. The Commission's example: a system that analyses customers' emotions with their knowledge to improve customer contact is persuasion. A system that infers emotions covertly to offer a more expensive product at the moment someone is most likely to buy is manipulation. For a small company this mostly matters for a chatbot on its own website. The guidelines expect whoever uses such a system not to bypass the safety measures its provider built in.
Where Bee Flow fits
Bee Flow is a workspace for AI chat and your own documents: writing, summarising and finding answers in what your company already has. That is the kind of use this piece puts on the calm side of the line. You can run Bee Flow on your own server, so texts and documents stay inside your own environment. The Privacy & Security page explains how that is set up. If you want to know whether it fits the way your team works, get in touch.