The European Parliament has backed the AI Act, putting the European Union on track to turn a broad artificial intelligence rulebook into binding law across the region. The vote gives political weight to a framework that classifies AI by risk and applies different obligations depending on how a system is used.
MEPs supported the provisional agreement reached in December with 523 votes in favor, 46 against and 49 abstentions. A final approval from the Council is still pending, but the parliamentary vote marks one of the last major steps before the law begins its phased rollout.
What Parliament Approved
The AI Act is built around a risk-based model. Instead of treating every AI tool the same way, it separates uses into categories and assigns rules based on the potential harm attached to each use-case.
Some uses are treated as "unacceptable risk" and banned outright. The source identifies social scoring and subliminal manipulation as examples of uses that fall into this prohibited category.
Other systems are categorized as "high risk." The source lists AI used in education or employment, as well as remote biometrics, as examples. These systems must be registered, and developers have to follow risk and quality management requirements set out in the legislation.
Most AI applications are left outside hard rules because they are considered low risk. That choice is central to the structure of the Act: tighter controls for uses lawmakers see as more sensitive, lighter treatment for the broad set of applications judged less dangerous.
How The Timeline Works
The AI Act will come into force 20 days after it is published in the EU’s Official Journal. The source says publication is expected in the coming months, followed by staged implementation rather than immediate full enforcement.
The first provisions to bite are the banned use-cases, which apply after six months. Other parts of the law follow after 12, 24 and 36 months. Full implementation is therefore not expected until mid 2027.
That phased schedule matters because the Act covers very different types of obligations. A ban on a prohibited use-case is different from a registration duty, a transparency rule or a risk management requirement. The timeline gives the EU a sequence for bringing those parts online.
Enforcement is also designed to be substantial. Violations of the ban on prohibited AI uses can lead to penalties of up to 7% of global annual turnover, or €35 million if higher. Breaches of other provisions on AI systems can attract fines of up to 3%, or €15 million. Failure to cooperate with oversight bodies risks fines of up to 1%.
Generative AI Gets A Specific Place In The Law
The Act also addresses tools that became more prominent during the negotiations, including AI chatbots, generative AI systems that can create synthetic media, and general purpose AI models, also called GPAI.
These systems are not all treated the same way. The source describes light touch transparency obligations for some applications, while the most powerful GPAIs face additional rules if they are classified as having "systemic risk."
Rules for GPAIs were added later in the process, pushed by concerned MEPs who wanted advanced models behind the recent boom in generative AI tools to be covered. Last year, lawmakers in the parliament proposed a tiered system of requirements for those models.
The final compromise did not create a full carve-out for advanced AI model makers. At the same time, it watered down the original parliamentary proposal. Most GPAIs will face limited transparency requirements, while models whose training used compute power greater than 10^25 FLOPs will likely have to carry out risk assessment and mitigation.
The source also notes the political pressure around this part of the law. A handful of EU Member States, led by France, pushed for a more permissive approach, with lobbying by homegrown AI startups such as Mistral. Since the compromise, Mistral has taken investment from Microsoft, while Microsoft holds a much larger stake in OpenAI, the U.S.-based maker of ChatGPT.
Lawmakers Defend The Compromise
During a press conference ahead of the plenary vote, co-rapporteurs Brando Benifei and Dragoș Tudorache were asked whether lobbying had weakened the rules for GPAIs.
Benifei answered: "I think we can agree that the results speak for itself," adding that the law defines safety needs for the most powerful models with clear criteria. Tudorache also rejected the idea that lobbyists had negatively shaped the final outcome, saying: "We negotiated and we made the compromises that we felt were reasonable to make."
Tudorache described the result as a "necessary" balance. He also said companies’ decisions had not impacted the work, and pointed to transparency around copyrighted material as a key concern.
In the debate before the vote, Tudorache framed the Act as an attempt to anchor AI governance in human control. He said: "We have forever attached to the concept of artificial intelligence the fundamental values that form the basis of our societies. And with that alone the AI Act has nudged the future of AI in a human-centric direction."
Why This Is Only A Starting Point
Supporters of the AI Act are presenting it as a major step, but not as the end of European AI governance. Benifei pointed to the need for future rules on the use of AI in the workplace and said work is also needed to improve investment conditions for AI in Europe.
He linked that investment question to common research, sharing computational capability and completing the Capital Markets Union. The concern, as described in the source, is that investors may prefer to invest in the U.S. rather than in another European country or another European company.
Tudorache made a broader point: AI will affect education systems, the labour market and warfare, so the governance model will need to evolve as the technology changes. He also repeated his call for joint AI governance work among likeminded governments and wherever wider agreements can be made.
The practical implication is clear. The AI Act gives the EU a first comprehensive framework for regulating artificial intelligence, but lawmakers already expect further rules, adjustments and international coordination. The Parliament vote moves the legislation close to law; the longer test will be how the framework works once its phased obligations begin to apply.