What Changed on 2 February 2025 Under the EU AI Act?

Why Does 2 February 2025 Matter?

The AI Act entered into force on 1 August 2024, but entry into force is not the same as application. The Act's obligations apply in stages, and 2 February 2025 was the first stage at which substantive duties became operational for providers and deployers. On that date, the general provisions and definitions, the AI literacy requirement and the prohibited practices began applying, which meant that organisations could no longer treat the Act as a future planning exercise for those matters. The date matters as a boundary: work done before it was preparation, work required from it onwards was compliance with applicable law, and records should show which side of the boundary each decision fell on. Later milestones built on this first stage, adding general-purpose AI, transparency and high-risk duties, but none of them replaced the duties that began in February 2025.

What General Provisions and Definitions Began Applying?

The general provisions that began applying include the subject matter, scope and definitions in Chapter I, including the definition of an AI system, the definitions of provider, deployer, importer, distributor and authorised representative, and the definitions of intended purpose, placing on the market and putting into service. Those definitions are the foundation for everything that follows: classification, operator duties, transparency and enforcement all depend on them. An organisation assessing a system after February 2025 therefore had to work from the Act's definitions rather than from marketing labels, product categories or internal terminology. A tool described as an assistant, a copilot or a model is assessed by what it does and how it is supplied, against the definitions, and the assessment record should cite the definition applied and the facts that support it. Getting the definitions right at this stage prevented errors that would have compounded at every later milestone.

For the broader regulatory framework, see EU AI Act timeline.

What Did AI Literacy Require From That Date?

Article 4 required providers and deployers to take measures to ensure a sufficient level of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf, taking into account their technical knowledge, experience, education and training and the context in which the systems are to be used, and considering the persons or groups on whom the systems are to be used. The duty is context-sensitive rather than a fixed training course: a developer, a procurement officer and an operator using AI output in a consequential workflow need different forms of literacy. From February 2025, an organisation should have been able to show what measures it took, for whom, on what basis and when they were reviewed. Regulation (EU) 2026/1744 later rewrote Article 4, and the current consolidated wording should be checked before relying on the original formulation, but the February 2025 start of the literacy duty itself was not moved.

Which Prohibited Practices Began Applying?

The original prohibited practices in Article 5 began applying on 2 February 2025. In outline, they cover manipulative and deceptive practices, exploitation of vulnerability, social scoring, specified biometric identification, categorisation and emotion inference practices, and other uses the provision defines, each subject to its own conditions and exceptions in the consolidated text. Prohibition screening became an immediate use-case question: before a system was placed on the market, put into service or used, the organisation had to test whether its intended and actual use fell within a prohibition. That screening comes before high-risk classification, because a prohibited use must not proceed to conformity work. Later amendments added new prohibitions applying from 2 December 2026, in Article 5(1)(ba) and Article 5(1)(bb), but those additions did not change the February 2025 start of the original prohibitions, and the two sets should be screened together in any current review.

What Did Not Start on 2 February 2025?

Several major parts of the Act did not start on 2 February 2025, and treating the date as the start of the whole Act would misstate the timeline. General-purpose AI model obligations under Chapter V followed on 2 August 2025, Article 50 transparency requirements followed on 2 August 2026, and the main high-risk requirements followed later still: 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems, as amended by Regulation (EU) 2026/1744. Enforcement powers for applicable rules also began later, on 2 August 2026. An organisation that assumed in February 2025 that every AI Act duty was already applicable would have over-planned in some areas and under-planned in others, while an organisation that assumed nothing was applicable until a later date would have missed the literacy and prohibition duties that were already live. The correct approach was, and remains, provision by provision.

Continue with what changed on 2 August 2025.

What Should Providers and Deployers Have Put in Place?

Providers and deployers should have put in place, by February 2025, an inventory of AI systems in use or development, a record of provider and deployer roles for each system, a prohibited-practice screen for the relevant use cases, and AI literacy measures for the people who needed them, with evidence retained for each step. The inventory should have recorded the system, its intended purpose, its users and the people affected, so the later milestones could be assessed against a known portfolio rather than discovered from scratch. The prohibition screen should have been recorded with the provision considered, the facts examined and the reviewer, so a later question about a use could be answered from the record. The literacy measures should have been proportionate to the roles and contexts concerned, with participation and review recorded. Those four elements, inventory, roles, screening and literacy, formed the foundation on which the August 2025, August 2026 and later work was built.

How Does the February 2025 Milestone Connect to Later Dates?

The February 2025 milestone connects to later dates as the first layer of a staged framework. The definitions and operator roles established then are used in the general-purpose AI stage of August 2025, the transparency and enforcement stage of August 2026, and the high-risk stages of December 2027 and August 2028. A classification made in 2025 on the basis of the definitions should therefore still be visible in a 2027 high-risk file, and a prohibition screen carried out in 2025 should be updated, not repeated from scratch, when new prohibitions arrive in December 2026. The staged design also means that a change in a system's purpose, users or data can move it between stages, and the record should show when each assessment was made and what triggered its review. Keeping the February 2025 work as the base layer, rather than discarding it when a later date arrives, is what makes phased implementation manageable across a portfolio.

Frequently asked questions

Continue through this AI Act cluster

• EU AI Act timeline
• what changed on 2 August 2025
• which AI Act rules to check now in late 2026
• AI literacy in practice
• prohibited practices under Article 5
• EU AI Act overview