On April 29 local time, EU member states and members of the European Parliament failed to reach agreement in negotiations on softening the AI Act. The talks ran for 12 hours. An official from Cyprus, which holds the rotating presidency of the Council of the EU, said no common ground had been found with the European Parliament. The negotiations are set to resume next month, with talks possibly held in two weeks.
These talks are one strand of the "digital omnibus" package pushed by the European Commission. The work revises several digital regulations as a single bundle, and along with the AI Act it covers GDPR, the personal data protection regulation, and revisions to the Data Act. The AI Act, which took effect in 2024, has served as a baseline that many countries around the world have referred to when drawing up their own rules. The work of cutting back that baseline itself has now stopped midway.
The sticking point was one issue. Whether to exempt from the AI Act those fields already subject to sector-specific regulation such as product safety rules became the point of contention, and the European Parliament demanded the exemption while the Council of the EU did not accept it. How much to ease transparency obligations for general-purpose AI models, meaning large models where a single model is applied to many uses, was also taken up. On the remaining provisions the two sides had narrowed their views considerably, and this one strand held up the whole thing.
The argument from the side pushing for softening is the double burden. Germany and some other member states and lawmakers have argued that complying with both the AI Act and existing sector-specific regulation drives up compliance costs and delays product launches. No calculation showing how much costs rise and how far launches slip was presented. On the opposing side, Dutch lawmaker Kim van Sparrentak said Big Tech is popping champagne while European companies that have finished their safety work have been left in regulatory confusion.
What that remark points to comes close to the real loss in this case. Whether the regulation is strict or loose, once its content is fixed companies can allocate staff and budget accordingly. On the other hand, when a state of possible rule changes drags on for months, any preparation carries the risk of being sunk. The companies hit hardest are those that have already built audit systems and documentation teams in anticipation of a high-risk designation.
The schedule that stands out immediately is the high-risk AI provisions. These are the rules requiring prior assessment for AI in areas that directly affect people, such as biometrics, healthcare and credit scoring, and they were set to apply in August this year, but no conclusion has been reached on whether to implement them as they stand or revise them. If the negotiations move to next month, the preparation time left until the application date is two or three months. If the rules go ahead unchanged, systems must be built within that window, and if they are softened, part of the money spent will not be recovered.
This is not someone else's problem for Korean companies either. Firms that export AI products to Europe or place their models on European clients' systems need to start by rechecking which version of the AI Act the compliance clauses in their contracts point to. Setting the two scenarios side by side, it wastes less to start filling in the items required in common, training data records, model performance logs, and review procedures with human involvement. These items will be required whichever way the conclusion goes.
The negotiating table opens again next month. In the meantime, development teams in Europe will spend their time poring over the revision history of the legal provisions. How far the EU is taking into account that a few days for the side making the rules spreads into months for the side making the products will be told by the outcome of the next meeting.
