The Digital Omnibus on AI came into force on 27 July 2026. This text amends the European regulation on artificial intelligence by postponing certain deadlines and easing several obligations, notably for systems classified as high-risk. In parallel, the European Data Protection Board has opened a consultation on web scraping for generative AI.
We covered the entry into force of transparency obligations. Here is the reverse move, and it deserves an explanation without caricature.
What an omnibus does
The term refers to a text that amends several existing pieces of legislation at once. It does not create a new framework: it adjusts the one that exists.
The changes mainly concern three points. A postponement of deadlines for certain obligations, particularly those concerning high-risk systems. A reduction of documentary burdens, especially for smaller organisations. And clarifications on points where the initial text lent itself to divergent interpretations.
The two readings, both of which hold up
The pragmatic reading. The European regulation was adopted before the uses it claims to govern had actually deployed. Certain obligations have proved costly and ineffective, particularly for modest companies deploying systems bought elsewhere. Adjusting a text in light of its application is normal, and probably desirable, functioning.
There is also an acknowledged concern: the risk that the compliance burden weighs mainly on European players, while large non-European groups have legal teams sized for the task. A regulation that penalised above all those it claims to protect would miss its target.
The critical reading. These easements come amid intense pressure on Europe to reduce its requirements, and at a time when Washington is reportedly considering asking countries to choose their technological camp. A postponement can be pragmatism, or the first step of an erosion.
And the postponed obligations concern precisely the most sensitive uses: those classified as high-risk, including recruitment and access to essential services.
In parallel, a consultation has been opened on the rules applicable to the mass collection of web data to train generative models. This is a topic we have been following since the publishers' lawsuits and the purchase of printed books. The question posed is difficult: under what conditions can publicly accessible personal data be used to train a model? A restrictive answer would make training in Europe very constrained. A permissive answer would strip individuals of any real control over their own data.
What this changes in practice
For a European company deploying AI, three consequences.
Extra time on certain obligations, which is a real relief for teams that were preparing in a hurry.
Less documentation to produce, mainly for modest organisations.
Increased uncertainty. That is the flip side, and it is underestimated. A text amended a year after its adoption sends a signal that is hard to interpret: should you prepare for the requirements as originally planned, or anticipate further adjustments? For a company planning over three years, a stable and demanding framework is often preferable to a shifting and lighter one.
What to take away
This case illustrates a real difficulty with no obvious solution: regulating a technology that changes faster than the legislative process.
A text written too early governs uses that do not yet exist and misses those that emerge. A text written too late arrives after practices have become established. Europe chose to write early, which gave it a head start on the agenda and now forces it to make corrections.
The question that matters is therefore not whether this adjustment is good or bad, but what it signals. A text that adjusts once to account for reality remains credible. A text that is lightened under every bout of pressure ceases to be, and companies and citizens alike end up ignoring it. We will know which of the two this is in a year.