AI Act Article 5: Prohibited AI Practices Checklist — AI Automation article by Rui Codex

Article 5 of the EU AI Act (Regulation (EU) 2024/1689) bans a closed list of AI practices, including manipulative techniques, exploitation of vulnerable people, social scoring, untargeted scraping of facial images and certain emotion recognition uses. Product teams can screen each feature against these categories before design sign-off. Most of the prohibitions have applied since 2 February 2025; the provisions added by Regulation (EU) 2026/1744 (Article 5(1)(ba) and (bb), and Article 5(1a) and (1b)) apply from 2 December 2026.

Key takeaways

  • Article 5 prohibitions apply regardless of sector, and safeguards do not turn a banned practice into a permitted one.
  • Screen every AI feature against the Article 5 list at design stage and record the decision with its reasoning.
  • Manipulation and exploitation bans turn on the effect on people, so review what a feature does, not only its data sources.
  • Some biometric and emotion recognition uses are banned outright, while others are regulated differently, so classify them separately.
  • Screening only works when a named owner signs off the result and the review repeats after every material change.

What does Article 5 of the AI Act prohibit?

Article 5 prohibits a defined set of AI practices outright. The ban does not depend on the sector, and adding a disclaimer or a human review step does not make a prohibited practice lawful. The prohibitions apply under the application rules in Regulation (EU) 2024/1689 on EUR-Lex.

This article is general information and not legal or tax advice. Read the official text of Article 5 and its recitals before you make any decision about a specific system.

Infringements of Article 5 fall in the highest fine tier of the regulation, with fines of up to EUR 35 million or 7% of total worldwide annual turnover, whichever is higher, as set out in Article 99 of the AI Act on EUR-Lex. That exposure is the main reason to screen features early, because removing a banned capability after launch is expensive and public.

How should a product team screen a feature?

Screen each feature against the Article 5 list before design sign-off, and record the answer to every question. The checklist below turns the article into questions a product manager can answer with evidence from the design.

  • Does the feature influence decisions through techniques people cannot notice or understand, in ways that distort behaviour?
  • Does it target people because of age, disability or a difficult social or economic situation?
  • Does it score people based on social behaviour or personal traits, and does the score lead to unrelated or disproportionate treatment?
  • Does it predict whether a person will commit a criminal offence using profiling alone?
  • Does it build or extend a facial recognition database through untargeted scraping of images?
  • Does it infer the emotions of employees or students in a workplace or an education institution?
  • Does it categorise people by race, political opinions, trade union membership, religious beliefs, sex life or sexual orientation?
  • Does it perform real-time remote biometric identification in publicly accessible spaces for law enforcement?

A checklist fails when a single team answers it from the roadmap alone. Each answer should point to a design artefact, such as a data flow, a prompt specification or a user journey, so a reviewer can test it.

Tip: Mark every question with yes, no or unclear. Treat unclear as a blocker until legal review resolves it, because a vague answer usually hides a design choice nobody has written down.

Which manipulation and exploitation practices are banned?

Article 5 bans AI systems that use subliminal, purposefully manipulative or deceptive techniques to materially distort behaviour and cause significant harm. It also bans systems that exploit vulnerabilities linked to age, disability or social or economic situation, as set out in the AI Act text on EUR-Lex.

Where the line blurs

Persuasive design is not banned as such. Urgency messages, personalised offers and recommendations are common in products. The regulatory concern is a combination of personalised targeting, a distortion of choice and a real harm, such as a vulnerable customer being pushed into a costly commitment.

The trade-off is real. Personalisation improves relevance, but every inference about a user's stress, age or financial pressure adds risk. Removing those inferences usually lowers conversion, so the decision belongs to a documented review, not to the growth team alone.

Tip: Ban the use of inferred vulnerability signals as model inputs for persuasion. If a signal is needed for safety, document why it is needed and who may see it.

When does scoring people become a prohibited practice?

Social scoring is prohibited when a system evaluates or classifies people over time based on social behaviour or personal traits, and the resulting score leads to detrimental treatment in contexts unrelated to where the data was collected. Article 5 also bans AI systems that assess the risk of a person committing a criminal offence based solely on profiling or on personality traits, as described in the AI Act on EUR-Lex.

Scores that are not banned

Credit checks, fraud scoring and risk scoring based on objective facts about a transaction are not automatically banned. The difference lies in the basis of the score, the context it is used in and the consequence for the person. A fraud model that flags a suspicious payment is a different design from a score that restricts access to services across unrelated areas of life.

Are emotion recognition and biometric uses banned?

Some are. Article 5 prohibits emotion recognition in workplaces and education institutions, except for medical or safety reasons. It also prohibits untargeted scraping of facial images from the internet or CCTV to create or extend facial recognition databases, and biometric categorisation that infers sensitive characteristics. The exact wording and exceptions are in the official AI Act text on EUR-Lex.

Workplace and classroom checks

Engagement tools that read facial expressions to judge staff or students are the most common failure point. Teams often present them as wellbeing features. The safer design uses a voluntary survey or a manager conversation, which avoids inferring emotions from biometric signals altogether.

Other biometric uses, such as device unlocking with consent, are not covered by the same ban, but they carry their own requirements. Classify each biometric use separately rather than applying one answer to all of them.

Hypothetical examples of screening decisions

Here are two hypothetical scenarios: first, a sales chat feature that times urgency prompts according to an inferred stress level of the visitor; second, an HR dashboard that flags staff mood from webcam images during video calls.

In the first scenario, the screening question is whether the feature exploits an inferred vulnerability to distort a decision. The team should remove the stress inference and test whether the urgency prompts still work on general page signals, which is a design change rather than a legal argument. In the second scenario, the feature infers emotions in a workplace, which Article 5 prohibits outside narrow exceptions. The team should replace it with a voluntary survey and record why the biometric version was dropped.

Where do screening programmes fail?

Screening programmes usually fail in three places: they happen after launch, they cover the model but not the surrounding workflow, and nobody owns the final decision. A model that is harmless on its own can become prohibited once a sales or HR process starts using its outputs.

The trade-off is speed. A lightweight screen at the start of each sprint costs little, but a full review of every change slows delivery. A practical compromise is a short screen for every feature and a deeper review only when a feature uses biometric data, personal traits or inferred emotional states.

Teams that build or integrate AI workflows can use the same checklist during architecture reviews. Rui Codex offers technical implementation support for AI workflows through its AI automation services and its AI consulting services. Questions about legal classification remain with the provider and qualified advisers. For the data protection side of these designs, the security and GDPR engineering articles cover related technical controls.

Next steps for your product team

Start by listing every AI feature in your current roadmap and running the checklist against each one. Record the answers, name an owner for each decision and repeat the screen whenever a data source, a model or a user group changes. Where a feature is unclear, pause it until qualified counsel has reviewed the Article 5 text against your design. Rui Codex is led by a founder with 200+ projects delivered, and its team can help you translate the checklist into technical design controls. To discuss your project, request a project consultation or email info@ruicodex.com. You can also see delivered projects at https://ruicodex.com/CaseStudies.

Frequently Asked Questions

What happens if a product is found to breach Article 5?

Breaches of Article 5 fall in the highest fine tier of the AI Act, with fines of up to EUR 35 million or 7% of total worldwide annual turnover, whichever is higher. The penalty rules are in Article 99 on EUR-Lex. Treat screening as a design-stage control, because withdrawing a banned feature after launch is costly.

Does Article 5 apply to internal tools that only analyse company data?

Yes, if the tool performs a prohibited practice. Article 5 covers the practices it describes, not the internal origin of the data, so an internal analytics tool can still fall under a ban if it manipulates people, scores individuals in a prohibited way or infers protected characteristics. Check the outputs and their downstream use, not only the data source.

Can a company use emotion recognition at all?

Emotion recognition is prohibited in workplaces and education institutions, except for medical or safety reasons. Outside those settings, Article 5 does not ban it, but other parts of the AI Act may still impose duties. Confirm the exact exceptions in the official text on EUR-Lex with qualified counsel before relying on any of them.

Does Rui Codex certify that our AI system complies with the AI Act?

No. Rui Codex offers technical implementation support, not legal advice, audits or certification. Its team can help map a feature to the technical controls a screening process needs, such as data minimisation, logging and human review. Legal classification and formal sign-off remain with the provider and qualified advisers.

Does Article 5 apply only to companies that build their own AI?

No. The prohibitions can apply to both providers and deployers within the scope of the regulation, so a business that only buys an AI tool can also be affected. Teams that integrate third-party models should check how the vendor's feature is configured and used in their own processes.

Sources

Tags: AI Act Article 5 prohibited AI practices EU AI regulation AI compliance product teams AI automation GDPR compliant software

Need Help Implementing This?

Our team can help you put these insights into practice. From AI automation to custom software development, we build solutions that deliver real results.

Send a project request