What result should an Article 5 assessment produce?
The first draft of this article was organised around eight prohibition labels. It was accurate enough and still not very useful. A product team reviewing one proposed AI use needs a next step, not a tour of Article 5.
A practical assessment should end with one of five results:
| Assessment result | Operational next step |
|---|---|
| Prohibited practice identified | Pause the proposed use or release, preserve the facts and route the decision for professional legal review. |
| Professional review required | Resolve the open legal concept or claimed exception before treating the practice as allowed. |
| No Article 5 prohibition identified | Continue the wider EU AI Act classification and check other applicable law. This result is not general approval. |
| New prohibition not yet applicable | Record the 2 December 2026 application date for the new synthetic intimate-content branch and assess any law already in force. |
| Assessment incomplete | Collect the missing description, category facts or evidence before relying on the result. |
The first eight Article 5 prohibitions have applied since 2 February 2025. They cover specified manipulation, exploitation of vulnerabilities, social scoring, individual criminal-offence prediction, untargeted facial-image scraping, emotion inference, sensitive biometric categorisation and real-time remote biometric identification. The ninth prohibition added in 2026 starts applying on 2 December 2026.
Run the free Article 5 assessment to test one intended practice and create a Record tied to the published schema version. No account is required.
Which facts decide the result?
Broad labels are weak inputs. “We use emotion AI” or “the system manipulates users” does not establish the legal elements. The assessment needs facts about what the system does, who is affected, how an output is used and why an exception may apply.
Use the closest Article 5 branch, then collect its decision-driving facts:
| Possible practice | Facts the assessment needs |
|---|---|
| Manipulation or deception | The technique, its effect on informed decision-making and whether significant harm is caused or reasonably likely. |
| Exploitation of vulnerability | The relevant age, disability or social or economic situation; how it is exploited; the behavioural distortion; and the harm analysis. |
| Social scoring | What is evaluated over time, the data or characteristics used and how the score leads to unrelated, unjustified or disproportionate treatment. |
| Individual crime prediction | Whether the prediction relies solely on profiling or personality traits, or supports a human assessment already based on objective and verifiable facts linked to criminal activity. |
| Facial-recognition database collection | Whether facial images are scraped without a target from the internet or CCTV to create or expand the database. |
| Workplace or education emotion inference | The context and whether a documented medical or safety reason is claimed. |
| Sensitive biometric categorisation | Which characteristic is inferred from biometric data and whether a narrow dataset or law-enforcement boundary is claimed. |
| Real-time remote biometric identification | Whether the use occurs in a publicly accessible space for law enforcement, the objective relied upon and the procedural basis for an exception. |
| Synthetic intimate or child sexual abuse material | The material, provider or deployer role, intended purpose, foreseeability, reproducibility, safeguards, consent and the 2 December 2026 application date. |
The Commission’s Article 5 guidelines add explanations and examples for the original prohibitions. The Commission describes the guidelines as non-binding; authoritative interpretation remains with the Court of Justice of the European Union.
What does a prohibited-practice result look like?
Take a simplified example already covered by the published assessment: an AI system infers worker emotions from camera feeds for performance management.
The relevant facts are more useful than the product label:
- the system is intended to infer emotions;
- the context is a workplace; and
- no medical or safety reason applies.
On those entered facts, the workflow returns prohibited practice identified. The next step is to pause the proposed use and preserve the assessment for professional review. Adding a generic human-oversight control would not change the Article 5 category being tested.
This is a workflow result, not a legal opinion. A real assessment still depends on an accurate description of the system and use. If the product measures fatigue for a documented safety purpose, for example, the exception and its evidence need professional review rather than an automatic approval.
The same element-by-element approach matters elsewhere. A persuasive interface is not automatically prohibited manipulation. A reputation score is not automatically prohibited social scoring. The result changes only when the facts satisfy the relevant elements or leave a material point unresolved.
When does an exception need professional review?
Some Article 5 branches contain narrow exceptions or boundaries. These include medical or safety reasons for workplace or education emotion inference, specified law-enforcement uses of real-time remote biometric identification and boundaries for lawful biometric-dataset handling.
The workflow should not turn a selected exception into “compliant”. It should record:
- the exact exception claimed;
- the facts and evidence supporting it;
- any authorisation or procedural requirement;
- the responsible reviewer; and
- the event that would trigger reassessment.
The individual crime-prediction branch has a similar boundary. Article 5 distinguishes a prediction based solely on profiling or personality traits from AI supporting a human assessment already based on objective and verifiable facts directly linked to criminal activity. A ceremonial human click does not establish the latter. The Record needs the objective facts, the human decision-maker’s role and the system’s actual influence.
When the evidence cannot resolve one of these points, professional review required is the useful outcome. It names the uncertainty without quietly approving the use or labelling every uncertainty a breach.
What if no Article 5 prohibition is identified?
“No Article 5 prohibition identified” has a narrow meaning: the entered facts did not identify one of the Article 5 prohibitions in the assessment.
It does not mean that:
- the AI Act is out of scope;
- the system is low risk;
- Article 50 transparency duties do not apply;
- provider, deployer or GPAI obligations are satisfied; or
- data-protection, employment, consumer, criminal or sector-specific law allows the practice.
Continue with the EU AI Act scope and classification assessment. That route records territorial scope, operator role and the next specialist track. If the wider route is already known, use the EU AI Act solution family to move directly to the relevant transparency, high-risk, GPAI or AI-literacy workflow.
What changes on 2 December 2026?
Regulation (EU) 2026/1744 added Article 5 prohibitions concerning AI systems that generate or manipulate specified non-consensual intimate material or child sexual abuse material. The new points, together with their provider and deployer conditions, start applying on 2 December 2026.
For providers, the amended text looks at intended purpose or a reasonably foreseeable and reproducible outcome without significant technical modification, together with reasonable and adequate safety measures and safeguards. For deployers, the use prohibition depends on using the system for the purpose of generating or manipulating the covered material.
The date belongs in the logic. On 1 December 2026, the Article 5 assessment should not claim that this new prohibition already applies. It can record new prohibition not yet applicable, preserve the facts and point to the future date. Other criminal, privacy, child-protection and platform rules may already apply, so the result is not permission to proceed.
What should the Record preserve?
Keep enough evidence for another reviewer to reproduce the route:
- system, feature and version assessed;
- intended practice and affected people;
- assessment date and effective legal period;
- Article 5 category tested;
- element-by-element facts;
- exception or boundary claimed;
- supporting design documents, policies and authorisations;
- calculated result and next step;
- unresolved question, owner and reviewer; and
- exact published schema version used.
The result may change when the intended purpose, interface, affected group, safeguard or law changes. Keeping the input facts with the rule version makes that change visible instead of rewriting the earlier decision.
Inspect the published Article 5 schema to see the questions, sources, effective periods and release history behind the Flow.
How can a team run the assessment?
Start with one intended use, not the organisation’s entire AI inventory.
Describe what the system does, who is affected and how its output changes a decision or treatment. Choose the closest possible Article 5 category. Answer the branch-specific questions and attach the rationale. Then act on the result: stop, obtain professional review, continue to the wider classification route or collect the missing facts.
Run the Article 5 prohibited-practice assessment. The completed Record keeps the entered facts, calculated next step, assessment date and published interpretation version together.
Sources
- Regulation (EU) 2024/1689, consolidated version of 27 July 2026
- Regulation (EU) 2024/1689, Official Journal text
- Regulation (EU) 2026/1744
- European Commission guidelines on prohibited AI practices
- European Commission AI Act overview and application timeline
This article is not lawyer-reviewed and is not legal advice. It was checked against the linked primary sources on 26 August 2026.