CTRL

In the 2020 paper that saw her forced out of Google, Timnit Gebru and her colleagues argued large language models (LLMs) were "stochastic parrots" – fluent without understanding. The paper was prescient: two years later, ChatGPT reached 100 million users in about two months, and the world encountered a lot of parrots at a scale they'd scarcely imagined. Chatbots started inventing court cases, citations, biographies and news events, all while doing so in the confident register of an encyclopaedia.

The gap between the reality and the presentation was down to how LLMs work. They predict plausible next words, not true ones. In that sense, a hallucination isn't a failure of the system. It's the tech working as designed on a question it can't answer. Treat fluency and truth as different things… and always verify what you're seeing.

Live in China and algorithmic systems score citizens for welfare fraud risk. Want to work in the west and you might see your job applications triaged by an AI product. Buy insurance anywhere in the world and your price will be dictated by an algorithm, often unseen and frequently unappealable. Social media feeds are the most visible case of algorithms – computer code running complex calculations to come to decisions. They're an automated editor for billions of users.

The EU's AI Act – the world's first comprehensive AI law – grades systems by risk, with fines up to €35 million or 7% of global turnover. But one key unanswered question is whether audit-and-transparency regimes can genuinely see inside systems that change regularly, or just document what they do after the fact.