What Changed on 2 February 2025 Under the EU AI Act

2 February 2025 was the AI Act’s first application date and it is routinely described as the point at which the Act “came into force”. It did not — the Regulation entered into force on 1 August 2024 — and what became applicable on that day was two chapters, one of which is a single article.

What switched on

Article 113(a) of Regulation (EU) 2024/1689 provides that Chapters I and II apply from 2 February 2025. That is the entire content of the milestone, and it is worth enumerating because the chapters are short:

• Chapter I — general provisions. Article 1 (subject matter), Article 2 (scope), Article 3 (definitions) and Article 4 (AI literacy). Only Article 4 imposes an obligation; the other three make the rest of the Regulation legible.
• Chapter II — Article 5. The prohibited AI practices, in a single article.

Nothing about high-risk classification, technical documentation, conformity assessment, CE marking, registration, transparency to users, or general-purpose models applied on that date. The general-purpose model obligations arrived on 2 August 2025; the Article 50 transparency duties arrive on 2 August 2026; and the high-risk regime, originally set for 2 August 2026, was moved to 2 December 2027 for stand-alone Annex III systems and 2 August 2028 for AI embedded in Annex I products by Regulation (EU) 2026/1744, in force since 27 July 2026. The frequent claim that companies had to “classify their AI systems” by February 2025 has no basis in Article 113 — although working out whether anything in your estate falls under Article 5 obviously requires knowing what your estate contains.

The prohibitions

Article 5 bans a defined set of practices outright, with no risk assessment or mitigation route. In outline, and each with its own carefully drawn conditions and exceptions in the article text: purposefully manipulative or deceptive techniques that materially distort behaviour and cause or are likely to cause significant harm; exploitation of vulnerabilities due to age, disability or a specific social or economic situation; social scoring leading to detrimental treatment that is unjustified or disproportionate or in unrelated contexts; risk assessment predicting criminal offences based solely on profiling or personality traits; untargeted scraping of facial images from the internet or CCTV to build facial recognition databases; inferring emotions in the workplace and education, save for medical or safety reasons; biometric categorisation to deduce certain protected characteristics; and real-time remote biometric identification in publicly accessible spaces for law enforcement, subject to narrowly defined exceptions and authorisation requirements.

Each of those has its own page in this cluster, because each has conditions that decide real cases — social scoring, emotion recognition at work and facial scraping are the three that most often catch commercial products rather than state systems. Reading the summary above as the rule would be a mistake; the qualifiers in the article are where the answers live.

The AI literacy duty

Article 4 is one sentence long and applies to providers and deployers alike. They must take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff and other persons dealing with the operation and use of AI systems on their behalf — taking into account those persons’ technical knowledge, experience, education and training, the context in which the systems are to be used, and the persons or groups on whom the systems are to be used.

Three features make it unlike the rest of the Act. It is not limited to high-risk systems: a deployer of any AI system is caught. It is an obligation of effort — “to their best extent” — rather than a specified outcome, with no prescribed curriculum, certification or record format. And it is not listed in Article 99(4), so it does not attach directly to the 3% fine tier, though a Member State’s national penalty rules under Article 99(1) are a separate question. See what the AI literacy obligation requires in practice.

The six-month enforcement gap

This is the fact that makes the February date genuinely interesting, and it is the one most often missed. Chapter XII — penalties — did not apply until 2 August 2025 under Article 113(b), and Article 99(1) gave Member States until that same date to lay down their rules on penalties and notify them to the Commission. Member States also had until 2 August 2025 to designate their national competent authorities under Article 70.

So for six months the prohibitions were legally applicable while the national machinery for penalising a breach of them, and in many Member States the authority that would do so, did not yet exist. That is not a grace period — the prohibition applied from 2 February 2025, conduct in that window was unlawful, and other legal consequences including data protection enforcement, contractual exposure and civil claims were unaffected. But it does explain why the date passed without visible enforcement, and why the absence of enforcement in early 2025 says nothing about how the prohibitions will be applied.

The guidelines published around the date

The Commission published two sets of non-binding guidelines in the days following the milestone: guidelines on prohibited AI practices under Article 5, and guidelines on the definition of an AI system in Article 3(1). Both were published in February 2025 and both are available through the Commission’s AI Act pages. They matter for two reasons: the definition guidelines address the threshold question of what is an AI system at all, and the prohibited practices guidelines work through the conditions in Article 5 with examples.

Their status should be stated precisely. Commission guidelines are not binding law. They indicate how the Commission reads the provisions and are persuasive in practice, but only the Court of Justice of the European Union can give an authoritative interpretation of the Regulation, and a national authority is not bound to follow a guideline. Where a guideline and the article text appear to diverge on your facts, that is a question for advice, not a resolved point.

For what came next, and why it is a different kind of milestone, see what changed on 2 August 2025.