Where Recommender Systems and Dark Patterns Sit Against Article 5

Recommender systems and dark patterns are related to the prohibited practices under Article 5, but their legal status is not automatically clear. Understanding where these design practices sit against the prohibition helps organisations know what they can and cannot do.

A recommender system is an algorithm that suggests content, products or services to a user based on their behaviour, preferences, demographic similarity to other users or engagement patterns. Recommender systems are used by streaming services to suggest films and shows, by e-commerce platforms to suggest products, by social media platforms to suggest posts to follow, and by news and publishing sites to suggest articles. They are among the most widely deployed AI systems globally. Most users encounter dozens of recommender systems every day.

Recommender systems are not automatically prohibited under Article 5. They are not subliminal; the user can see that recommendations are being made. They are not inherently manipulative; the user can choose to ignore them or dismiss them. However, a recommender system could become prohibited if it is deliberately designed to work below conscious awareness or to manipulate the user against their interests. A recommender system that uses psychological manipulation to increase time on platform, even if the user would prefer to leave, steps into prohibited territory.

Dark patterns are user interface and experience designs that manipulate people into taking actions against their interests. Common dark patterns include hidden cancellation options that require finding a buried link to unsubscribe, pre-ticked boxes that enrol the user in a service unless they uncheck, misdirecting language that obscures negative consequences, artificial urgency that claims an offer expires soon, and asymmetric defaults that make the chosen path different from the optimal one. Dark patterns coupled with AI could become prohibited practices. For example, an AI system that learns to time its dark pattern intervention to moments when it knows a user is tired, stressed or vulnerable could constitute subliminal manipulation.

The key distinction is agency and awareness. If a user is aware that they are being influenced and retains meaningful choice to refuse or ignore the influence, it is not a prohibited practice. If the system is deliberately designed to bypass awareness or artificially constrain choice, it approaches prohibition.

The EU has recently published additional guidance on dark patterns as a distinct problem. Article 5 prohibition of manipulative techniques applies specifically to dark patterns that use AI. Notably, other EU legislation, particularly the Digital Services Act, also restricts dark patterns regardless of whether AI is involved in their deployment.

For recommender systems, the implications are clear: optimise for relevance and user satisfaction, not just for engagement metrics or time spent on platform. If your recommender system is designed primarily to keep users on your platform even when they would prefer to leave and do something else, it crosses into manipulative territory.

For organisations using or building recommender systems, ask whether the system respects user agency. Can users understand why specific recommendations are being made to them? Can they easily override recommendations or tell the system they dislike a recommendation? Are recommendations made at times and in ways that respect the user’s attention and autonomy? If the answer to any of these is no, the system may be approaching prohibited manipulation. Examples of manipulation in recommender systems include: timing notification of recommendations to times when the user is known to be fatigued; using language designed to trigger emotion; personalising recommendations to exploit known psychological vulnerabilities; progressively shifting recommendations to pull users toward extreme content against their stated preference.

The grey area between permitted engagement optimisation and prohibited manipulation is genuine and organisations often need compliance advice on specific systems. Some systems are clearly permitted (suggesting movies similar to ones you enjoyed), while others are clearly prohibited (timing notifications to exploit fatigue). Many fall in between and require analysis by your compliance team. This is where the Commission’s official guidelines become critical for interpretation. If your system is in the grey zone, consulting those guidelines before deployment protects you.

Your organisation should establish a process for escalating recommender system designs to your compliance team when there is any doubt. Ask your engineers: Is timing of notifications optimised for user engagement or for user vulnerability? Is the algorithm designed to present relevant information or to maximise dwell time? These questions drive the legal analysis.