Two reports out this week describe the same problem from opposite ends. Governments are trying to regulate artificial intelligence, and the technology keeps outrunning them. The interesting part is that writing a law and writing fifty laws seem to fail in strikingly similar ways.
Europe took the comprehensive route. Its AI Act, adopted in June 2024, is the most ambitious attempt anywhere to govern the technology, sorting systems into tiers of risk and phasing in obligations over years. As of this month its general provisions are in force, with the strictest high-risk rules still to come. Yet the people closest to it are already uneasy. The member of the European Parliament who steered it, Valerie Hayer, says the pace of change has outstripped what lawmakers anticipated, and the law's lead author, Gabriele Mazzini, describes a disconnect between the text and the reality it is meant to regulate. The point is not that the Act has failed; it is that a rule negotiated over years can be aimed at a version of the technology that no longer exists by the time agencies apply it. As the UK lawmaker Darren Jones put it, governments work in years while AI moves in weeks.
The United States took the opposite route, which is to say no route at all. Dozens of federal AI bills have been introduced in this Congress and none has reached a floor vote, stalled by lobbying and by a fight over whether federal law should wipe out state law. President Trump's executive order in January 2025 revoked his predecessor's AI rules and told agencies to favour innovation over precaution. Into that vacuum the states have poured.
The result is a genuine patchwork. California's SB 53 took effect this year imposing safety and transparency duties on frontier model developers; New York's RAISE Act follows in 2027. Colorado's broad high-risk AI law went live in June, only to be replaced by a narrower one next year. More than 35 states have proposed chatbot-disclosure bills, 47 have enacted some form of deepfake law, and a new front has opened around algorithmic pricing in housing and retail. For a company operating nationally, that means overlapping and sometimes inconsistent obligations, each with its own definitions and deadlines. Mandatory bias audits for hiring tools alone run 15,000 to 30,000 dollars apiece, with full compliance programs reaching six figures.
Both outcomes carry a warning. A single thorough law risks being precise about the wrong thing; fifty competing laws risk being so tangled that compliance becomes a tax on doing anything at all. Technology companies are now lobbying for a federal statute specifically to erase the state rules, a tidy solution that would also hand the industry a single, friendlier referee. The harder truth sitting under both reports is that passing a law is the easy part. Staffing it, funding the expertise to enforce it, and updating it fast enough to matter is the work almost no government has shown it can do at the speed this technology demands.