Learning about cars. Three stories, one revolution

The first news item appears designed to irritate university vice-chancellors. Andreessen Horowitz, one of Silicon Valley’s icons, has decided that it is no longer enough to wait for universities to produce the talent that tech companies need. It wants to try producing it itself. Thus was born the Horowitz Andreessen Academy, an unaccredited school in San Francisco, initially funded with $42 million and backed by partners that read like the agenda for an AI conference: Anthropic, Google, Meta, Nvidia, OpenAI and Palantir. The first intake, scheduled for 2027, will have around fifty students and will be free of charge. Few traditional exams, few traditional assignments, plenty of project-based learning, work placements in companies, and selection based largely on ‘proof of work’: show me what you can do, not just what marks you’ve achieved. It is not hard to see the cultural provocation here. For decades, Silicon Valley has complained about universities, yet it has gone on to hire their best graduates.

Now try skipping a step. If AI makes the ability to memorise information less valuable, if programming increasingly means managing systems capable of programming, if many skills become obsolete more quickly than the courses that teach them, then the value of a qualification could shift from certifying what you know to demonstrating what you can build. The risk is clear: confusing education with corporate training, producing excellent employees but less well-rounded citizens. But it would be naive to dismiss the experiment as yet another example of Californian arrogance. The question it raises is a serious one: when the labour market is changing at the speed of AI, can universities continue to behave as if their product had a shelf life of twenty years?

The second experiment comes from Taiwan and is almost the opposite. Here, the aim is not to train people for machines, but to train machines to understand people better. The Taiwanese government is building a ‘Sovereign AI Training Corpus’, a vast national archive intended for training models: around five thousand datasets and over 2.2 billion tokens, with publishers, writers and e-book platforms invited to contribute. It is a cultural policy disguised as technological infrastructure. Until recently, defending a language meant funding schools, libraries, translations, cinemas, theatres and newspapers. In future, it will also mean ensuring that the machines millions of people will use to write, translate, search and study know that language well enough not to relegate it to the digital periphery.

And so a small country can become marginalised, not because anyone is banning its culture, but because AI has a poor understanding of it. The idea of ‘sovereign AI’ naturally risks becoming autarkic rhetoric. But the Taiwanese case suggests a more intelligent version of sovereignty: not building walls around the models, but ensuring that the models have enough material to understand who you are. In a few years’ time, language policy might concern not only which books young people read, but which books are read by the machines that will respond to them.

The third article completes the picture. In the United States, Microsoft and the American Federation of Teachers have established new rules on the use of AI in schools. These include external audits; no use of pupil or teacher data to train models, save for limited exceptions related to security; no sale or commercial exploitation of that data; and no features designed to create emotional dependence in young people. OpenAI and Anthropic are currently discussing similar agreements with the union. For three years, we have asked almost one thing and one thing only: will students use ChatGPT to cheat? A legitimate question, but perhaps not the most important one. The more interesting question is the reverse: what will ChatGPT learn from pupils? A pupil who speaks to an artificial tutor does not merely hand in an assignment. They reveal mistakes, hesitations, interests, difficulties, perhaps fears, habits and information about their family. And this is where privacy ceases to be a tedious clause in the terms and conditions and becomes pedagogy.

Taken together, these three stories clearly illustrate where we have ended up. Silicon Valley wants to decide how to shape talent. Governments want to decide what culture to instil in machines. Schools and trade unions want to decide which aspects of pupils’ learning machines must not retain. In short, AI is not merely making its way into education. It is forcing us to grapple once again with the oldest questions in education: who teaches, what is worth passing on, and to whom does what we learn truly belong.