News
EU Gains Power to Fine Makers of the Most Powerful AI Models Starting August 2
The transitional period of the EU's AI Act ends on August 2, 2026, giving Brussels the power to demand documentation from developers of the most powerful AI models, impose fines of up to 3 percent of global turnover, and pull models from the market. The Commission has also expanded its AI Office by 38 new staff.
Contents
On Sunday, August 2, 2026, the transitional period of the EU's Artificial Intelligence Act ends for providers of general-purpose AI models. From that day, the European Commission can, for the first time, genuinely enforce the rules against companies such as OpenAI, Anthropic, and Google DeepMind, demanding documentation, conducting independent model evaluations with access to source code, and imposing fines of up to 3 percent of global annual turnover.
Brussels' New Powers
Until now, the AI Act's provisions on general-purpose models applied mostly on paper - companies had to comply, but the Commission did not yet have the full tools to enforce them. That changes on August 2. The office gains the right to interview company employees during investigations, demand access to models and technical documentation, and, in extreme cases, order a model's withdrawal from the EU market.
Analysts tracking the AI Act's rollout describe these powers as among the most far-reaching regulatory authority any government has ever claimed over artificial intelligence technology. Previously, requests for information were voluntary; now they become mandatory for providers of advanced models.
As enforcement begins, we are taking an important step - Henna Virkkunen, Vice-President of the European Commission for digital technology
Strengthening the AI Office
Alongside the activation of its new powers, the Commission expanded the EU's AI Office team in Brussels by 38 additional staff. The office now employs more than 140 people, of whom over 40 work in a dedicated AI safety unit, most holding specialized technical backgrounds, mainly in cybersecurity.
Despite the staffing boost, experts note that current headcount may prove insufficient given the scale of the office's responsibilities. A report by the organization Pour Demain recommends that by 2030 the team overseeing general-purpose models should number at least 160 people. The office also struggles with recruitment - rigid EU salary scales and bureaucratic procedures make it hard to compete for specialists against the private sector, where pay can be several times higher.
Penalties and Enforcement Tools
For violations of the rules on general-purpose models, the Commission can impose a fine of up to 3 percent of a company's global annual turnover or 15 million euros, whichever is higher. That rate is close to the maximum penalties known from the EU's data protection regulation (GDPR), though the AI Act sets separate, higher thresholds for the most serious violations involving prohibited practices.
The Commission has also launched a whistleblower platform and a tool letting users report suspected violations by AI providers. This is meant to offset the office's limited staffing - part of the detection work is expected to come from outside, from company employees and end users.
Poland's Context
For Poland, August 2 carries extra significance - it is also the date from which EU-wide rules on labeling AI-generated content take effect. Poland's national law implementing the AI Act has already been signed by the president, and the new Polish commission overseeing artificial intelligence is set to launch in November. Until then, companies operating in Poland remain directly subject to EU rules and EU oversight, even without a fully formed national body.
The biggest impact of the new powers will likely be felt by global model providers rather than Polish companies - as developers of general-purpose models, they fall under the AI Office's direct oversight. Polish businesses deploying ready-made models in their products remain, for the most part, subject to lower risk thresholds, though industry reports indicate that some companies in the country still do not fully comply with EU requirements even without a fully formed national oversight body yet in place.
The activation of full enforcement powers on August 2 closes a three-year process of implementing the AI Act, which began when the regulation was adopted in 2024. It is the first test of whether the EU's model of AI regulation, built on documentation obligations and independent risk assessments, can actually be enforced against companies that are, for the most part, headquartered outside Europe.
