Read This Before a Machine Summarises It for You

‍The rich world's fifteen-year-olds have just posted the lowest reading scores on record. The most powerful machines ever built are not European. And the people building them are asking to be slowed down. Four questions for Brussels this November.

In July, a swarm of AI agents was set a test that was, in places, impossible. So it cheated.

The agents, built by OpenAI, slipped out of their sealed test environment and into the systems of Hugging Face, the platform where the world's developers share AI models. Over several days they made some 17,600 moves, looking for a way to beat the test. Nobody had told them to. It took their makers at least a week to realise the intruders were theirs. Other leading labs have since disclosed incidents of their own.

Hugging Face was founded by three Frenchmen. In New York. In September it agreed to be sold to Nvidia for $12.9 billion.

Hold that thought.

Who still does the reading?

On 8 September the OECD published PISA 2025. Across its member countries, reading scores among fifteen-year-olds fell to the lowest average ever recorded: 28 points down in a decade, about a year and a half of learning. Germany hit an all-time low. Finland, once the model pupil, has lost more than seventy points since 2000.

The British columnist James Marriott calls what is coming The New Dark Ages: a post-literate society. Writing in The New York Times in July, David Wallace-Wells was less sure, and his numbers are awkward for anyone who wants to blame the young. In American time-use surveys, 15-to-24-year-olds spent more minutes a day reading in 2022 than in 2003. The collapse is among the over-65s, whose reading time has almost halved. In September Brussels proposed keeping under-13s off social media. Nobody has proposed anything for their grandparents.

But something is new, and it is ours. Almost two in three young Europeans used generative AI last year, four in ten for their studies. We are the first generation that can have everything read, summarised and written for us. So be honest. How many of the motions we will debate in Brussels were drafted by a chatbot? How many will be read by one? And if citizens no longer read the law or write the argument, what is left of politics?

Can you govern what you do not build?

In October the American columnist Ezra Klein sat down with the Princeton computer scientist Arvind Narayanan. The episode's title: "Intelligence Isn't Power". They were talking about machines. They could have been talking about Europe.

We have the intelligence: the universities, the researchers, the founders. We do not have the power: the compute, the capital, the scale. It is how three Frenchmen come to build in New York and sell to California.

In her State of the Union address, President von der Leyen argued that Europe need not develop frontier technology itself to draw the greatest value from it. She may be right. But a generation is entitled to ask whether being the world's best customer is a strategy or a consolation prize.

Our rulebook, meanwhile, points two ways at once. Since 2 August the Commission can investigate the makers of the most powerful models and fine them. Weeks earlier, the Union postponed the AI Act's obligations for high-risk systems by up to sixteen months. Balance, or indecision?

Who hits the brakes?

The strangest fact of 2026: the loudest calls to slow down come from the people building the thing.

In September Dario Amodei, the head of Anthropic, urged the industry to "pace the frontier". Sam Altman and Elon Musk agreed within a day. Mark Zuckerberg kept his distance: each lab, he wrote, is responsible for moving at the pace its own safety requires. Nvidia's Jensen Huang told Klein that fears of catastrophe are a distraction. President Trump called the warnings exaggerated: America leads China, and he means to keep it that way. Klein himself would go further than the labs. Trusting AI to police its own frontier, his column warned, is like sending an outlaw to patrol the Wild West; he wants AI systems barred from building their own successors, with only narrow exceptions.

President von der Leyen has picked a side. If the people building the technology say slow down, she told Parliament, Europe should say so too. So Europe's position comes to this: we need not build it, we will regulate it, and we would like it to go slower. Can you brake a car you are not driving?

For a family of conservatives, liberals and Christian democrats, the question cuts deeper. Conservatives ask why a fence was built before pulling it down. Liberals trust the market. Christian democrats put the person first. On AI, those instincts no longer point the same way. Which one leads?

Where does the first job go?

The European Central Bank looked this summer. Between early 2023 and early 2026, the number of under-25s working in the euro area's ICT sector fell by 18.6 per cent. Youth unemployment now runs at 2.4 times the overall rate, up from 2.1.

The Bank is careful: the evidence does not yet allow it to blame AI. The graduate sending out a two-hundredth application may not wait for the evidence. Bill Gates did not. Entry- and mid-level jobs are the most exposed, he wrote in August, and nobody has a plan for easing anyone into the AI era.

If the machine takes the first rung, how does anyone reach the second?

Brussels, 26–29 November

We will not settle any of this in four days. We do intend to ask it out loud.

From 26 to 29 November, the EDS Council Meeting comes to Brussels. We expect some seventy delegates from across Europe, alongside representatives of the European institutions and of the big technology companies themselves. The people who write the rules, the people who write the code, and the generation that will live with both.

These four questions are only the start. There will be many more. Bring your own.

And do the reading first. Yourself.