For the past two years the story of AI regulation was a story of construction. Legislatures drafted, the European Union passed a sweeping AI Act, and Colorado became the first U.S. state to enact a comprehensive law governing automated decisions. In 2026 the verbs have changed. On May 14, Colorado's governor signed a bill that repeals that very law, the 2024 Colorado AI Act, weeks before it was due to take effect, and replaces it with a narrower framework that strips out the most demanding obligations. The headline is not that a state regulated AI. It is that a state un-regulated it, before the ink had even dried.
Colorado is not an outlier; it is the clearest instance of a pattern. In Brussels, negotiators for the Council and the Parliament agreed on May 6 to a "Digital Omnibus" that simplifies the AI Act itself, shifting industrial and machine-embedded AI toward existing sector-specific safety rules rather than stacking two regulatory regimes on top of each other. In Washington, the Daily Beast reports that President Trump shelved a planned AI executive order just hours before it was due to be signed, after his former AI czar David Sacks phoned to argue it went too far. The order would merely have asked companies to let the government preview new models voluntarily. Even that proved too much.
The reasons rhyme across capitals, and they are mostly about speed and China. Colorado's rewrite is explicitly pitched as reducing the regulatory burden on business. The White House camp around Sacks favours a hands-off approach precisely because Beijing is racing ahead. India set the tone earlier this year, naming "innovation over restraint" as one of seven guiding principles and leaning on existing laws and voluntary codes rather than new statute. The fear of falling behind has become the universal solvent that dissolves draft regulation, and industry lobbying has learned exactly which pressure point to press.
What is striking is how thin the dividing line has become between caution and acceleration. The same Washington order was reportedly triggered in the first place by Anthropic's Mythos, the autonomous vulnerability-hunting tool that defense and treasury hawks worried could be turned on critical infrastructure. One faction read that capability as a reason to regulate faster; another read it as proof that America must not slow down while rivals sprint. India's finance ministry flagged Mythos too, and then kept its light-touch guidelines anyway. The technology that most alarms officials is also the technology cited to justify leaving the field clear.
Not everyone is rowing in the same direction. Amnesty International this week called outright for a prohibition on generative AI systems built on unlawful web scraping, arguing they are incompatible with human rights by design. Analysts at Gartner warn that the rollback could rebound on companies themselves, predicting that two in five enterprises will be forced to decommission their own AI agents by 2027 once governance gaps surface after the fact. That is the open question hanging over the whole retreat: if neither states nor the EU nor the White House is willing to set the guardrails, the job does not disappear. It simply lands, unplanned and after the damage, on whoever has to clean up the first serious incident.