U.S. labs split on AI liability

OpenAI and Anthropic have split over how much legal protection artificial intelligence labs should get if powerful models cause catastrophic harm. (wired.com) In Illinois, Senate Bill 3444 would shield developers of “frontier” models from liability for “critical harms” if they did not act intentionally or recklessly and if they post a safety protocol and transparency report online. The Illinois General Assembly summary says the bill would also treat compliance with European Union safety requirements or a federal agency agreement as enough to satisfy the law. (ilga.gov) Wired reported on April 10 that OpenAI supported the bill, which defines “critical harms” to include the death or serious injury of more than 100 people, at least $1 billion in property damage, and some chemical, biological, radiological, or nuclear weapon scenarios. Wired then reported on April 15 that Anthropic opposed the same measure and argued that it protects developers too broadly. (wired.com, wired.com) The fight is about who pays when an advanced model causes damage: the company that built it, the person who misused it, or both. Illinois is trying to answer that before Congress has passed a national liability law for frontier artificial intelligence systems. (wired.com, ilga.gov) Europe is moving in the opposite direction on timing and detail. As of August 2, 2025, the European Union’s Artificial Intelligence Act governance system was fully in force, with the European Artificial Intelligence Board operating and member states required to designate national enforcement authorities. (interoperable-europe.ec.europa.eu) Brussels is also filling in how the law applies to newer tools that can act with more independence, often called “agents.” The European Commission said on January 30 that it is drafting 2026 guidance on high-risk classification, transparency duties under Article 50, serious-incident reporting, and other compliance questions under the Artificial Intelligence Act. (interoperable-europe.ec.europa.eu, economistjurist.es) At the same time, the European Commission is pushing public agencies to use more artificial intelligence, not less. On April 9, 2026, it marked one year of its Artificial Intelligence Continent Action Plan by publishing a report on adoption in public administration and saying 19 Artificial Intelligence Factories and 13 antennas are now operational. (digital-strategy.ec.europa.eu) That push builds on the Apply Artificial Intelligence Strategy published in October 2025, which promotes an “AI first” approach, covers 10 industry sectors plus the public sector, and says public bodies should favor European and open-source solutions where possible. The Commission said the strategy has already opened dozens of calls worth up to €1 billion. (interoperable-europe.ec.europa.eu, digital-strategy.ec.europa.eu) Europe’s own data shows that public agencies were already experimenting before this year’s push. A 2024 Commission-backed survey counted 1,617 cases of emerging technologies in European public administrations, with 80 percent of cases led by national and local governments. (interoperable-europe.ec.europa.eu) So the Atlantic split is getting sharper in April 2026: one debate in Springfield is about limiting lawsuits after worst-case failures, while one debate in Brussels is about writing detailed rules and expanding day-to-day use inside government. (wired.com, digital-strategy.ec.europa.eu)