Why Big Tech Doesn't Like Europe

This morning I have time. Ugo left yesterday for Bavaria and will spend August running through wet grass, because dogs have two houses too and August belongs to the other one. I haven't been outside. I made the coffee in the moka and didn't rush to drink it, and I pulled my father's old Baedekers off the shelf, the red ones, with maps that unfold like accordions and city names set in a typeface nobody uses anymore. Monday I go down toward Trieste. It counts as a holiday, though people like me never quite put the computer away: two things to keep an eye on, no deadline breathing on me. A morning like this produces a longer piece than usual.

On 5 August Demis Hassabis stepped back from running Google DeepMind to become its Chair. The teams working on code are moving from London to Mountain View, and the lab is left without the one non-US executive with oversight of development. Three days earlier, on 2 August, the Commission had switched on enforcement of the AI Act. I put those two things together in half a second, and for a moment I saw the Trade Federation from Star Wars folding its tents and leaving the planet that had just tried to write a rule. A trade blockade rather than an invasion: more elegant, and history repeats itself.

Then I checked, which is worth doing precisely when the hypothesis is a good one. Hassabis had been easing out of the chief executive role for about a year while Google consolidated Gemini around Pichai: an internal reshuffle with a man in it who prefers scientific problems to meetings. And somebody had actually measured the thing I was busy imagining. The Centre for the Governance of AI examined 375 model releases from Meta, Google, OpenAI and Anthropic between 2018 and May 2026: 11% arrived late in the Union or never arrived at all, and of the 68 documented cases, 56 have a regulatory cause. That cause is the GDPR, the handling of personal data in training, with higher barriers for images and audio, where a face and a voice stay recognisable. On the AI Act, the research finds little evidence of any effect on releases. In March, meanwhile, Google was opening an AI centre in Berlin.

So the question moves, and it gets more interesting than the one I started with. If the law isn't what slows the race down, what actually stings?

The answer sits in a public document nobody has read as a confession. On 31 July 2025 xAI signed the GPAI Code of Practice, but only the Safety and Security chapter, declaring that the other parts contain "requirements that are profoundly detrimental to innovation" and that the copyright provisions are "clearly an over-reach." Two chapters were refused: transparency and copyright. Thirteen days earlier Meta had signed nothing at all, with Joel Kaplan calling the measures ones that go "far beyond the scope of the AI Act." Anthropic, Google, Microsoft, OpenAI and Mistral signed the whole thing.

Set the two sets of obligations side by side and the refusal stops looking like a position of principle. The safety chapter asks a company to assess its own models inside a framework it writes itself, to mitigate risks, and to report incidents from here on. The two rejected chapters ask you to write down what you have already done. The Commission's mandatory template for the public summary of training content wants the top 10% of domain names you scraped, how your crawlers behaved and on what dates they collected, whether you used the data your own users generate inside your own products, whether you folded in synthetic data and which models it came from. Refreshed every six months. Article 53(1)(c) additionally wants a written policy on how you treated the opt-outs of everyone who didn't want to be mined, and Recital 106 drops all of it on you regardless of the jurisdiction where the training happened.

Disclosure is discovery. The safety file is read by a regulator that has already put on record that it will not audit content and that it acts on complaints or on a qualified alert from the scientific panel. The training-data summary is read by publishers, collecting societies and class-action lawyers, in every forum where that company can be sued. Two documents, two audiences, and only one of them ends up in a case file.

What that document costs is known, because somebody has already been handed the bill. On 20 July 2026 the settlement in Bartz v. Anthropic was approved at one and a half billion dollars, roughly three thousand a book across half a million works. It came with more than four thousand pages of unsealed filings describing Project Panama, the lossy library built by buying volumes, scanning them and destroying them. A billion and a half is the price of a single discovery event when a court is the one prising it open. Article 53 asks for a weaker version of the same thing, every six months, with nobody suing you. Disclosure is discovery, and the figure that measures it has been public for three weeks.

The strongest objection has to be granted, and on this one I don't have a settled view. The list of domains is a trade secret: data curation is a large part of what separates one model from another, and publishing the map of your sources hands it to whoever is chasing you. Nobody raising that is lying. It would hold if the refusal covered the whole disclosure, and instead xAI signed the chapter with the heaviest operational load and rejected the two that cost least to implement and most to have on the record. On copyright, what the chapter asks is how you treated a reservation of rights the European legislator had already written in 2019. Contesting Recital 106 is legitimate, and lawyers genuinely do contest it, because copyright is territorial as a matter of principle. Contesting territoriality is one thing. Calling it over-reach to declare what you took is another.

And here comes the part that leaves a taste. In the same week it kept the clause that threatens balance sheets, the Union pushed back by sixteen months the one that protects people: the obligations on high-risk systems, hiring, credit, essential services, migration and borders, slip to 2 December 2027, and along the way the Omnibus watered down AI literacy and the public register too. Anyone refused a visa by an automated system waits until 2027; anyone who scraped half the web answers now. Corporate Europe Observatory counts 69% of the Commission's 2025 meetings with business groups against 16% with non-governmental organisations, and it doesn't take much to work out which half of the law had somebody sitting at the table. There it is: the Trade Federation got itself a convenient delay. Legibility before ownership is still the European bet, and it bends at the exact point where it touches capital.

If you write, photograph, translate or publish code, your work is almost certainly inside a training set, and the only instrument that would tell you is that fifteen-page summary two companies refused to sign. It gives you nothing back, that document. It only tells you which room to search in, which is precisely what you need to open a case and precisely what anyone without a thousand lawyers behind them lacks. That is why they attack it at the point where it works.

You can measure the weight of a law by which of its articles gets attacked. Nobody issued a press release against the obligation to label deepfakes.

I'll close with the Baedeker open on the table. On the corner of St Mark's Basilica in Venice, next to the Porta della Carta of the Doge's Palace, stand four Roman emperors in Egyptian red porphyry, a stone reserved in Rome for emperors and effectively nobody else: the Tetrarchs. They embrace in pairs, they wear armour, and they have the same identical face: they are not four portraits, they are the depiction of a power-sharing deal. No individual features, enormous eyes fixed on a point far away, and the nose has crumbled off over the centuries. Look at them and you will see immediately where the Star Wars designers went to find the faces of the Federation's merchants.

The statues are the archetype; the film is the copy, and it is a good copy. Whoever keeps only the copy takes home the image and leaves behind the thing that produced it, which is the same difference that separates commanding a machine and buying one.

If you pass through Venice, go and look at them: they came from Constantinople in 1204, sawn apart to be carried, cemented into a corner as loot nobody ever gave back. If you can't get to Italy, make do with Star Wars.

Have a good Sunday.