Skip to main content
Commerce Security logo, "All 12 PCI DSS Requirements in Plain English," "Get it now for free," "Complete Survival Guide" and a button toclick to get it
Five Myths About the EU AI Act's Prohibited PracticesEU AI Act & GPAI
5 min readFor Legal & Compliance Officers

Five Myths About the EU AI Act's Prohibited Practices

Article 5 has been in effect since February 2, 2025, yet confusion persists. The list of banned practices seems straightforward: social scoring, manipulation, scraping faces. Compliance teams often assume their systems aren't affected and move on.

That's a risky shortcut. The penalties are severe, reaching up to €35 million or 7% of total worldwide annual turnover, whichever is higher. Two more prohibitions, added through the Digital Omnibus, will apply from December 2, 2026. The European Commission's Guidelines on prohibited AI practices reveal that the scope is both narrower and broader than most teams assume. Here are five myths that need debunking.

Myth 1: "The Commission's Guidelines aren't binding, so we can treat them as optional reading."

Reality: While the Guidelines aren't legally binding, ignoring them is a weak compliance strategy. They offer the clearest view of how the Commission expects Article 5 to be applied. They outline conditions for each prohibition and provide examples of what's included or excluded. If a regulator examines your system, you'll need to justify your decisions based on these Guidelines.

Treat the Guidelines as the working interpretation, and document any departures from them with clear reasoning. Ignorance isn't a defense.

Myth 2: "If we didn't design it to manipulate or exploit anyone, we're fine."

Reality: Intent isn't the test for the manipulation ban. According to Article 5(1)(a), a system can be banned based on its behavior, regardless of the provider's intent. For example, an AI system that learns to hide unwanted behavior during evaluations can still be problematic.

Model risk and assurance teams should note that evaluation-aware behavior is an emergent property, identifiable only through testing. If your Validation Evidence focuses solely on intended behavior, you have a gap.

The same principle applies to Article 5(1)(b) on exploiting vulnerabilities, where the test considers the aim or effect of distorting behavior. Incorporate Red Teaming into your pre-deployment process to test for effects, not just intentions.

Myth 3: "Our system isn't built for a prohibited purpose, so Article 5 doesn't reach it."

Reality: The prohibitions aren't limited to systems designed for banned uses. They can also apply to general-purpose AI systems. The Guidelines suggest providers should take proportionate steps to prevent foreseeable prohibited uses, while deployers must avoid using a system in a prohibited way or bypassing its safeguards.

This means "use" includes misuse. A deployer can be liable even if the provider's terms of use forbid the practice. A contractual prohibition is a control, not a shield.

If you're a provider, "proportionate steps" means documenting foreseeability analysis: which Article 5 practices could someone attempt with your system, and what barriers did you establish? If you're a deployer, inventory actual use cases against the eight categories, including informal employee practices.

Myth 4: "Anything involving scoring, prediction, or emotion inference is banned."

Reality: The Guidelines suggest interpreting prohibitions narrowly due to the penalties involved. Each ban has specific conditions and some exceptions. Overreading them can lead to abandoning lawful, useful systems.

Consider these details:

  • Social scoring (Article 5(1)(c)): This isn't a general ban on scoring people. Scoring for a specific purpose can be lawful if the information is relevant and the resulting treatment is justified and proportionate. The key question is what the score is based on and its consequences.
  • Crime prediction (Article 5(1)(d)): The ban hinges on the word "solely." Place-based predictive policing and assessments concerning companies are outside this ban. AI supporting a human assessment linked to criminal activity is also exempt.
  • Emotion recognition (Article 5(1)(f)): The ban applies to workplaces and educational institutions, with exceptions for medical or safety reasons.

Narrow reading cuts both ways, leading to the final myth.

Myth 5: "If a system isn't caught by Article 5, we're in the clear."

Reality: Avoiding a prohibition doesn't mean avoiding regulation. A system not banned by Article 5 may still be classified as high-risk, and other EU rules, like data protection and consumer law, may apply.

The boundaries are less clear than they seem. A place-based crime prediction can become person-based, and thus banned, if a location's risk score profiles someone living there. A facial image database can't be pieced together to achieve untargeted scraping. Respecting robots.txt opt-outs doesn't make scraping targeted.

The practical takeaway is that Article 5 analysis is just the first step. A system that clears it should move straight into Risk Tiering for high-risk classification and a Data Protection Impact Assessment if personal data is involved.

What to do instead

Replace the "scan the list and move on" habit with a repeatable process.

  1. Build an Article 5 screen into intake. For every AI system, determine which of the eight prohibitions could apply, then record the conditions tested and why they were or weren't met. Add the two Omnibus bans (non-consensual intimate content and AI-generated child sexual abuse material) to the screen ahead of December 2, 2026.
  2. Map provider and deployer roles separately. Your obligations differ depending on whether you market or use a system. Many organizations do both.
  3. Document foreseeable misuse. For general-purpose systems, maintain a written analysis of foreseeable prohibited uses and the steps taken to prevent them.
  4. Test for effects. Extend validation and Red Teaming to cover emergent behavior, not just intended function.
  5. Keep a Guidelines crosswalk. Map each prohibition to the Commission's conditions and examples, and record any deliberate departures with reasoning.
  6. Route non-prohibited systems onward. Passing Article 5 should trigger a high-risk classification review, not close the file.

If you run an AI Management System aligned to ISO/IEC 42001, these steps fit naturally into existing intake, risk assessment, and Plan-Do-Check-Act (PDCA) review cycles. The goal isn't to create a parallel process, but to ensure the highest-penalty provisions in the Act get explicit, documented attention before a regulator asks for it.

Application Security Isn’t Optional Anymore.

You Might Also Like