The EU AI Act bans manipulation and requires bot disclosure

What happened

Regulation (EU) 2024/1689, the AI Act, was published in the Official Journal on 12 July 2024, after signature by the Parliament and Council presidents on 13 June 2024. Article 113 sets a staggered timetable rather than one start date: the regulation enters into force twenty days after publication, but Chapters I and II, which include Article 5's prohibited practices, already apply from 2 February 2025, while the general application date, covering Article 50's transparency duties, is 2 August 2026. Adoption and application are separate dates, and a claim that the Act 'applies from 2026' misses that some bans have been in force since early 2025.

What the documents show

Article 5(1)(a) prohibits placing on the market or using a system that deploys subliminal techniques beyond a person's consciousness, or manipulative or deceptive techniques, with the effect of materially distorting a person's behaviour by impairing informed decision-making, causing significant harm. Article 5(1)(b) separately prohibits a system that exploits any vulnerability of a natural person or group due to age, disability or a social or economic situation, with the same significant-harm threshold. Article 50(1) requires a provider to design a system that interacts directly with people so that those people are informed they are dealing with an AI system, unless that is already obvious in context. A secondary explainer, artificialintelligenceact.eu, renders the same two paragraphs more plainly and confirms the 2 February 2025 date for Article 5.

The mechanism

The two prohibitions in Article 5 are not a general ban on persuasive design; each requires materially distorted behaviour and significant harm, which the regulation leaves to case-by-case assessment rather than a bright line. Article 50 is a disclosure rule, not a prohibition: a companion app can lawfully use an anthropomorphic persona so long as artificiality is not concealed and is not obvious from context. The two articles target different failure modes, a harmful outcome however achieved under Article 5, and a specific omission regardless of harm under Article 50, so a product can breach one without the other.

What it leaves open

Neither article defines significant harm in the operative text; the recitals point to severity, cumulative effect, scale and the affected person's vulnerability as relevant factors, leaving interpretive work to guidance and enforcement. Article 50's obvious-from-context exception is similarly undefined for a companion product whose entire design goal is to feel like a person. This is an editorial point: a persona-driven companion app sits closer to the disputed edge of both articles than a conventional customer-service bot, without a settled answer on either.

• Does the product disclose its artificial nature somewhere a user is likely to see it, or only in a terms page?
• Could the product's persona or memory features be read as exploiting a specific vulnerability, such as age or isolation?
• Which of the Act's application dates, February 2025, August 2025 or August 2026, governs the provision in question?

Because the AI Act separates a small set of early-applying prohibitions from a much larger body of rules that wait until 2026, a claim about what the Act requires of a companion app needs a date attached, not just an article number.