Two weeks ago, EU governments and the European Parliament struck a deal to simplify the bloc's landmark AI Act, narrowing the definition of "high-risk" systems, pushing the toughest compliance deadlines back to 2027 and 2028, and carving machinery out of the regime entirely. Last week, the Colorado governor signed a revised version of that state's first-in-the-nation AI law, repealing its algorithmic discrimination provisions and replacing "high-risk AI systems" with the gentler term "automated decision-making technology." On the same days, Vietnam unveiled its own AI law, the UK's House of Lords scheduled a debate on whether to legislate at all, and an NBC New York investigation found that AI lobbyists are now active in every American statehouse. The picture is not a global consensus. It is a hundred separate consensuses being shaped by the same handful of companies.
The reversals in Europe and Colorado are the most striking part. The EU AI Act was, when it passed in 2024, the world's strictest framework. The amendments agreed on 7 May give SMEs simplified documentation, narrow the obligations that automatically trigger the heaviest regime, and remove parallel assessments where sector-specific laws already apply. Supporters call it a pragmatic fix to cut red tape. Critics, including Green MEP Sergey Lagodinsky, call it the first step into fragmentation. "By having excluded machinery, we're making a first step into fragmenting AI regulation," he told Euronews, pointing to the United States as a cautionary tale of conflicting state rules with no coherent federal framework.
Colorado's revision goes further. The original 2024 AI Act was challenged by X.AI in April, the federal government intervened (the first time it had done so over a state AI law), and the state's attorney general agreed to suspend enforcement. The replacement law, SB26-189, drops the duty of care to mitigate algorithmic discrimination, removes the requirement for annual impact assessments, and abandons the term "artificial intelligence" altogether. Colorado characterised the new bill as protecting consumers "while not being onerous on developers." The original was supposed to take effect in February 2026. The replacement begins in January 2027, eleven months later, and demands less.
Behind the reversals sits a lobbying operation now operating at a scale that would have seemed implausible two years ago. TechNet, the industry group whose members include OpenAI, Anthropic, Amazon, Meta and Nvidia, says it has "engaged on 808 bills across 50 states, Washington D.C., and Puerto Rico, with our policy position prevailing 87% of the time." NBC New York's analysis of ethics filings found TechNet opposed a Wisconsin chatbot safety bill, a Nebraska bill on AI-generated child sex abuse material, a Rhode Island bill creating a right to sue over AI harms, and restrictions on AI data centres in Colorado and New Jersey. TechNet says it supports the aims of each bill but objected to specific wording, a position legislators have grown tired of hearing. "Last year, when we tried to move this bill, they actually killed the bill," New York Assembly Member Michaelle Solages said of the New York AI Act.
Research published this week by Trinity College Dublin, in collaboration with the University of Edinburgh, gives the pattern a name. The team identified 27 established patterns of "corporate capture" in tech policymaking, then applied them to coverage of four critical events between 2023 and 2025: the EU AI Act trilogues and the global AI summits in the UK, South Korea and France. They found 249 separate instances. The most common were "narrative capture," dominated by the framing that "regulation stifles innovation," and "elusion of law," which the researchers describe as contested interpretations of antitrust, privacy, copyright and labour rules. The paper's lead author, Dr Abeba Birhane, notes the same arguments now being used to justify the EU's simplification drive.
The tension across these stories is whether what is happening adds up to a workable governance landscape or a deliberately fragmented one. The Senate Judiciary Committee did unanimously advance the GUARD Act this month, a federal child-safety bill that draws on state laws and could establish a baseline. Eight states have enacted chatbot rules between 2023 and 2025. The EU still has a law, even if a thinner one. But each retreat raises the same open question. If a coherent federal or supranational framework is not coming, and state-level rules are being shaped by the same lobbyists in every capitol, then "the patchwork" is not a transitional state on the way to consensus. It is the equilibrium.