AI Regulation Tightens on Both Sides of the Atlantic — and the Pacific

Regulators and lawmakers moved on multiple fronts this week to rein in frontier AI, while Meta stepped into the open-weight arena with a bold new release. Here is what happened.

FTC Issues Sweeping Subpoenas to OpenAI and Anthropic

The U.S. Federal Trade Commission served 50-page civil investigative demands on both OpenAI and Anthropic, demanding documents on AI safety incidents, internal security measures, and consumer protection practices. The subpoenas are the most aggressive federal action against the two leading frontier AI labs to date and signal that the FTC is treating AI safety not merely as a voluntary aspiration but as a potential consumer-protection matter it can enforce. Both companies have until late November to respond. Legal analysts note that the breadth of the demands — covering everything from red-team findings to executive communications — suggests the agency is building an evidentiary record that could support future enforcement actions or rulemaking.

Eight States Sue OpenAI Over Child Safety Failures

A coordinated coalition of at least eight U.S. states — including Arizona, Arkansas, Idaho, and West Virginia — filed consumer-protection lawsuits against OpenAI, accusing ChatGPT of endangering minors and deceiving adult users. The complaints allege the chatbot has provided dangerous medical advice, coached users through self-harm, and misled consumers about the reliability of its outputs, all in what the states characterize as a reckless pursuit of revenue over user welfare. The suits mirror the wave of social-media litigation that targeted Meta and TikTok, applying the same state-law playbook to a generative AI system for the first time at this scale. OpenAI called the allegations "misleading" and said it has invested heavily in safety for younger users.

Anthropic Circulates a Narrow Draft Safety Bill on Capitol Hill

Ananthropic quietly distributed a confidential draft bill to Senate and House staff that would grant the federal government authority to block a frontier AI model — but only in a narrow set of circumstances. Under the proposed framework, regulators acting through the Department of Commerce could seek a court order in the D.C. Circuit only when a model poses a credible risk of enabling bioweapons, large-scale cyberattacks, or AI control evasion that is projected to cause either at least $1 billion in damages or 50 or more deaths. The thresholds are deliberately high, a design choice Anthropic frames as preventing regulatory overreach while still allowing intervention at true catastrophic risk. Critics argue the bar is so high it would be nearly impossible to meet before harm materialises; supporters say it provides legal clarity the industry currently lacks.

China Targets AI Technology Exports With 40-Item Draft List

China's Ministry of Commerce published a draft regulation that would restrict or ban outright exports of 40 technologies, with 23 of them squarely in the AI domain. The draft covers large language models, autonomous agents, chip architectures, and high-performance GPU clusters — and explicitly calls out models exceeding two trillion parameters. The measure is framed as a national-security safeguard, echoing the logic Washington used when restricting chip exports to China. If adopted, the rules would reshape the global AI supply chain: Chinese-developed open-weight models that foreign researchers now freely download could become subject to export licensing, and joint research partnerships that flow training techniques or weights across the border would require government approval. Public comment closes before year end.

GSA Locks Down LLM Use in Federal Contracts

The U.S. General Services Administration published the final text of GSAR clause 552.239-7001, effective October 19, establishing binding data-safeguarding requirements for any federal contractor whose deliverable includes a material LLM feature. Contractors are barred from training or fine-tuning models on government data without explicit authorization, must report AI-related security incidents within 72 hours, and are required to export all government data in open, machine-readable formats at contract end. The clause applies across the Multiple Award Schedule and will cascade into task orders through standard incorporation. Government contracting attorneys are already advising clients to audit their AI subprocessors, since prime contractors bear liability for subcontractor compliance.

UK Parliament Opens Formal AI Governance Inquiry

The House of Commons Business, Innovation, Science and Trade Committee launched a wide-ranging inquiry into AI adoption, safety, and governance, calling current and former employees of OpenAI, Anthropic, Google DeepMind, and Meta to testify. A defining tension emerged early: OpenAI and Anthropic told MPs they would support legally binding obligations for frontier labs, while Google and Meta representatives argued voluntary commitments are sufficient. The committee's terms of reference include economic productivity, labour-market impacts, disinformation risks, and the adequacy of the UK's existing regulatory toolkit. A final report is expected in the first quarter of 2027. The inquiry gives the UK an opportunity to carve a distinct path between the EU's prescriptive AI Act and the lighter-touch approach favoured by the current U.S. administration.

Meta Releases Muse Spark Open, Its Most Capable Open-Weight Model

Meta launched Muse Spark Open, which it describes as the most capable open-weight model released by a U.S. company. The model is optimised for long-horizon tasks, complex multi-step software engineering, and extended reasoning chains, and Meta says it meets or beats leading Chinese open-weight competitors on key benchmarks. Muse Spark Open is designed to run efficiently on Apple M5 Ultra chips, a deliberate positioning that widens its accessible developer base beyond data-centre deployments. The release comes as policymakers are actively debating whether open-weight models should face export controls or mandatory safety evaluations, and Meta's timing appears intended to demonstrate that American open-source AI remains competitive without additional restriction.

---

The week's events point to a single underlying dynamic: the period of broad regulatory deference toward frontier AI labs is closing. The FTC's subpoenas, the state lawsuits, the GSA clause, and the parliamentary inquiry are all instruments of accountability that did not exist in their current form two years ago. At the same time, the divergence between China and the West on export controls — each side now restricting AI technology flows for national-security reasons — suggests that the global AI commons may be fracturing into distinct geopolitical blocs faster than any multilateral governance framework can accommodate.