The first binding obligations of the EU AI Act applied on 2 February 2025, six months after the regulation entered into force. From that date, a short list of AI uses that the European Union considers an unacceptable risk, including social scoring, untargeted scraping of facial images and emotion recognition at work or in school, became prohibited, and every organization that builds or uses AI systems gained a duty to ensure AI literacy among its staff. For software teams, the EU AI Act stopped being a future compliance project and started shaping which features can legally ship to European users.

Key Facts#

  • Effective date: 2 February 2025, for the Act's general provisions (including the AI literacy duty in Article 4) and the prohibited practices in Article 5.
  • Legal basis: Regulation (EU) 2024/1689, the AI Act, which entered into force on 1 August 2024 with obligations phased in over several years.
  • Scope: The bans apply to all AI systems, whether they were placed on the market or put into service before or after 2 February 2025.
  • Penalties: Using a prohibited practice can lead to fines of up to EUR 35 million or 7% of worldwide annual turnover, whichever is higher.
  • Guidance: The European Commission published draft guidelines on prohibited AI practices on 4 February 2025. The guidelines are not legally binding.
  • Who is affected: Providers and deployers of AI systems, including companies outside the EU whose systems are used in the Union.

What Happened#

The AI Act uses a risk-based structure, and the prohibitions sit at its top tier. Article 5 lists eight practices that are banned outright rather than regulated. In summary, the Act prohibits AI systems that:

  1. Use subliminal, manipulative or deceptive techniques that materially distort a person's behavior and are likely to cause significant harm.
  2. Exploit vulnerabilities linked to age, disability or a person's social or economic situation.
  3. Evaluate or classify people based on social behavior or personal traits (social scoring) in ways that lead to detrimental or unjustified treatment.
  4. Predict the risk that an individual will commit a crime based solely on profiling or personality traits.
  5. Build or expand facial recognition databases through untargeted scraping of images from the internet or CCTV footage.
  6. Infer emotions of people in the workplace or in educational institutions, except for medical or safety reasons.
  7. Use biometric categorization to deduce sensitive characteristics such as race, political opinions, religious beliefs or sexual orientation.
  8. Perform real-time remote biometric identification in publicly accessible spaces for law enforcement, apart from narrowly defined exceptions such as searching for victims of serious crimes or preventing an imminent threat.

The same date activated Article 4 on AI literacy. Providers and deployers must take measures to ensure that staff who operate AI systems on their behalf sufficiently understand how those systems work and what risks they carry. The Act prescribes no training format, so organizations decide what "sufficient" means in their context and should document it.

Two days after the bans applied, the Commission released its draft guidelines on prohibited practices. They explain how the Commission reads each prohibition, with worked examples of what falls inside and outside the bans, and they are meant to help national market surveillance authorities apply the rules consistently. The Act's definition of an "AI system" also became applicable on 2 February, which matters because software that does not meet the definition is outside the Act entirely. An official version of the prohibited-practices guidelines, dated 29 July 2025, was later published on the Commission's AI Act Service Desk.

Background#

The AI Act was adopted in 2024. Rather than applying everything at once, it staggers obligations: prohibitions and AI literacy first, then obligations for general-purpose AI model providers and the penalty regime, and finally the detailed requirements for high-risk systems. The high-risk rules were originally due to start applying in August 2026, but a 2026 amendment package known as the AI omnibus pushed them back to December 2027 for stand-alone high-risk systems and August 2028 for AI embedded in regulated products.

Putting the bans first was deliberate: these practices were judged incompatible with fundamental rights however well engineered, so they got no long transition period. Most rarely appear in mainstream business software, but the edge cases that do are where developers need to pay attention.

Why It Matters for Developers#

The prohibitions are narrow, but they are absolute, and they reach any product used in the EU regardless of where it was built. For teams shipping .NET applications, a few areas deserve an explicit review.

  • Emotion features in HR, proctoring and e-learning products. The Act defines emotion recognition as inferring emotions or intentions from biometric data such as facial images or voice. A webcam-based "engagement score" for employees or students is squarely in scope. Text sentiment analysis on support tickets is not biometric and is not covered by this ban, but you should be able to explain why. If you rely on third-party vision or voice APIs, check which attributes they actually infer, not just which ones your feature displays.
  • Face data pipelines. Scraping public images to build a face-matching index is prohibited. If a computer-vision feature depends on a face dataset, confirm how that dataset was collected. Our multimodal AI guide covers vision workloads where this question comes up.
  • Scoring systems. Trust or risk scores that combine unrelated behavioral data and penalize people in an unrelated context can drift toward social scoring. Document each scoring model's purpose and inputs.
  • AI literacy. If your team uses AI coding assistants or ships LLM features, keep a record of the training and guidance you provide. Short internal material on model limitations, prompt injection and data handling goes a long way. The Responsible AI and LLM security guide is a practical starting point.

A useful engineering habit is to keep an inventory of every AI feature, the model it calls and the data it processes. Routing model calls through a shared abstraction such as IChatClient from Microsoft.Extensions.AI lets you add logging and telemetry in one place, which keeps that inventory accurate. Feature flags that can disable a capability per region give you a fast response if legal review flags a problem.

What's Next#

The prohibitions were only the first milestone; under the Act's phased timetable, obligations for general-purpose AI model providers and the national penalty regimes came next. Because the Commission's guidelines are not binding, their interpretations can still be tested by national authorities and, ultimately, by the Court of Justice of the European Union. The 2026 omnibus package also added a prohibition covering AI systems used to generate non-consensual intimate imagery and child sexual abuse material, so Article 5 is no longer fixed at eight practices. Re-check the banned list whenever the Act is amended.

Sources#