The EU AI Act: rules by risk level
The European Union's AI Act sorts AI systems by risk. A short list of uses is banned outright, such as social scoring by governments and most real-time facial recognition in public spaces. 'High-risk' systems — AI used in hiring, credit, education, medical devices or critical infrastructure — must meet requirements on data quality, documentation, human oversight and accuracy before they can be sold in the EU. General-purpose models like large chatbots carry transparency duties, and the most capable ones face extra obligations around testing and incident reporting. Obligations phase in over several years, which is why news about the Act keeps arriving: each deadline, each guidance document and each enforcement decision is a new story. Because the EU market is large, many companies apply the rules globally rather than build separate products.
United States: executive action, states and agencies
The US has no single federal AI law. Policy moves through presidential executive orders, which can be issued and revoked quickly, through existing agencies applying old laws to new technology (the FTC on deceptive claims, the EEOC on hiring discrimination, the FDA on medical AI), and through states. States such as California and Colorado have passed or debated their own AI bills on safety testing, deepfakes in elections and algorithmic discrimination. The result is a patchwork: when you read that 'the US regulated AI', check which of these three levers actually moved, because their durability is very different.
Chips and export controls
Advanced AI depends on specialised chips, so chip trade has become AI policy. The US restricts exports of the most powerful accelerators and chip-making tools to China and some other countries, and has repeatedly tightened and adjusted those thresholds. These rules affect company earnings, data-centre locations and diplomatic relationships. See our explainer on AI chips for how the hardware works and why a handful of firms dominate it.
Safety testing and frontier models
Governments including the UK and US created AI safety institutes to test powerful models before and after release. Their work is mostly voluntary agreements with labs rather than binding law. Debates centre on what counts as a dangerous capability, who gets to audit models, and whether open-weight releases should be treated differently from closed ones.
Copyright, data and training
Courts in several countries are hearing cases over whether training AI on copyrighted books, articles, images and music without permission is lawful. Outcomes will decide whether AI companies must license training data. Separately, privacy regulators apply data-protection law to how personal information is collected for training and how chatbots answer questions about real people.
