This page helps screen an AI system against the original eight Article 5 categories and the new prohibited-content category before launch, procurement, integration, or material product change.
The focus is practical evidence: what the system does, who it affects, which Article 5 limb is relevant, and whether any narrow exception or condition is actually documented.
Structured answer sets in this page tree.
Cited legal and guidance references.
The original prohibitions have applied since 2 February 2025. Regulation (EU) 2026/1744 enters into force on 27 July 2026 and adds a prohibited-content category that applies from 2 December 2026. These rules ban specified practices rather than treating them as high-risk systems that may proceed with controls. Before launch, procurement, integration, or material change, identify the system and intended use, test every relevant Article 5 limb, document each element of any narrow exception, and preserve the product, data, legal, and approval evidence behind the decision.
The review should start with the original eight prohibited-practice categories and the new category added by Regulation (EU) 2026/1744. Do not collapse them into a generic high-risk assessment: several categories turn on specific facts such as deception, vulnerability, biometric data, criminal-risk prediction, law-enforcement use in a publicly accessible space, provider safeguards, or a deployer's purpose in generating content.
For each category, record whether the system is placed on the EU market, put into service, or used in the Union, what function is enabled, what data is used, and which people or groups can be affected. Record the application date too: the original categories have applied since 2 February 2025, while the new prohibited-content category applies from 2 December 2026.
Some entries include carve-outs or conditions. Treat those as evidence requirements, not as informal risk arguments. The screening record should quote the relevant Article 5 limb, identify the specific exception being relied on, and show why the facts fit that exception.
The strictest evidence burden is for real-time remote biometric identification in publicly accessible spaces for law enforcement. allows it only when strictly necessary for specified objectives and subject to safeguards, national-law conditions, fundamental-rights impact assessment, EU database registration, authorisation, notification, and reporting steps described in the Article.
The new content prohibition uses different tests for providers and deployers. A provider must assess intended capability, reasonably foreseeable and reproducible outcomes, foreseeable misuse and circumvention, and whether reasonable state-of-the-art safeguards demonstrably prevent or sufficiently reduce the output. A deployer is prohibited when it uses the system for the prohibited purpose; accidental output or lawful use for other material is not the same test. Non-consensual intimate material is limited to realistic intimate or sexually explicit depictions of an identifiable person without the required consent, subject to the provision's stated scope and medical exception.
The most useful record is short but factual. It should let a reviewer reconstruct the system purpose, user journey, data sources, affected persons, and exception analysis without interviewing the product team again.
Keep separate evidence for negative conclusions. A statement that does not apply is weak unless it explains, for example, why a scoring feature is not social scoring, why biometric processing is not covered by the prohibited biometric categories, or why behavioural nudging does not meet the manipulation or vulnerability tests.
Use these questions before approving a feature, buying a vendor system, changing a data source, or enabling a biometric or profiling capability. A yes, unclear, or vendor-only answer should stop approval until the record is completed.
The review should be repeated when the same AI model is embedded in a new workflow. can turn on deployment context, especially workplace, education, law enforcement, public accessibility, vulnerable groups, and the consequences of a score or prediction.
A prohibited-practices check is an approval gate. The output should be a dated screening record with category-by-category conclusions, not a general AI-risk memo.
If a category is potentially triggered, pause release or procurement until the legal, product, data, and operational owners have either removed the relevant function, narrowed the use case, or documented why the exception and conditions are satisfied.
No. prohibits specific biometric uses, including certain biometric categorisation to deduce protected characteristics and real-time remote biometric identification in publicly accessible spaces for law enforcement unless strict Article 5 conditions are met. Other biometric systems may still need separate AI Act analysis, but they are not automatically Article 5 prohibited practices.
It should prove which AI system and use case were reviewed, how the original eight categories and the new prohibited-content category were assessed, what data and user-impact evidence supported each conclusion, which application date controls, and whether any stated exception or condition was relied on.
Sorena can help turn this Article 5 screening guide into a structured review record with category conclusions, source citations, owners, and release-blocking evidence gaps.
Ask questions tied to cited sources about Article 5 categories, exceptions, and evidence needed for a prohibited-practices review.
Review an AI system, supplier claim, or biometric use case against the EU AI Act Article 5 screening categories.
"No decision that produces an adverse legal effect"
"All AI systems considered a clear threat"
"laying down harmonised rules on artificial intelligence"