MHRA advances lifecycle AI-device rules; Anthropic proposes a high bar for blocking frontier models
Regulators and institutions are building practical safeguards for AI in healthcare, education and Tanzania’s petroleum sector, emphasizing monitoring, accountability and human review. But Anthropic’s proposed federal bill would allow model suspensions only at an exceptionally high risk threshold, while students and governance researchers argue that public oversight must become stronger and more responsive.
AI legislation and oversight tracker
UK: Medical-device regulator moves toward lifecycle AI rules
Britain’s Medicines and Healthcare products Regulatory Agency (MHRA) is developing rules that would follow AI-enabled medical devices throughout their lifecycle—not just assess them before launch. Its plans include stronger traceability, regulatory sandboxes, a new Centre for AI and Digital Health, and closer cooperation with trusted regulators such as the US Food and Drug Administration (FDA).
In plain English: medical AI can change after it is deployed, so regulators want ongoing checks to ensure it remains safe and performs as expected. The MHRA and FDA may reduce duplicated work, but each agency will continue making its own decisions. The approach is still being built through proposed reforms, the AI Airlock testing programme and NHS partnerships.
US: Anthropic proposes a narrow federal power to block dangerous models
Anthropic has circulated a draft federal bill that would allow the government to suspend a frontier AI model only when there is an imminent biological, cyber or loss-of-control threat capable of causing more than $1 billion in damage or killing or seriously injuring more than 50 people. A federal judge in Washington would have to approve the suspension.
In plain English: the proposal would create a high threshold for stopping a model and would limit state-level frontier-AI rules. Independent evaluators would assess risks, but the draft does not establish detailed standards for those evaluators. Supporters could see the bill as a predictable national framework; critics argue it could make government intervention too difficult and too slow.
UK: Universities urged to prove consequential AI is safe in practice
Coventry University professor Elena Gaura is calling for organisations that use AI in high-impact settings to demonstrate that systems are fit for purpose, assign responsibility, monitor performance and offer human review and remedies. She argues that existing data-protection, equality and sector rules—as well as the EU AI Act—are only a starting point.
In plain English: an AI system should not be treated as safe merely because it meets a legal checklist or has a strong accuracy score. Organisations should test who is missed or wrongly flagged, look for differences between groups, and reassess systems when their use or behaviour changes. The model’s supplier would remain part of the accountability chain, but the organisation deploying it would have to show that it works safely in its real setting.
Massachusetts: Student walkout calls for stronger AI regulation
Students from 10 Massachusetts colleges and universities are scheduled to walk out on October 20 and gather outside the State House in Boston. The action is part of a nationwide protest organised by Irreplaceable, which is calling for a pause on AI development until its risks are better understood.
In plain English: the protest is pressure on elected officials to take a larger role in decisions that organisers say should not be left to technology companies. Students are highlighting concerns about employment, education and public accountability. The immediate question is whether the mobilisation produces concrete legislative action.
Tanzania: Petroleum regulator builds AI expertise for audits
Tanzania’s Petroleum Upstream Regulatory Authority is training staff to use AI for petroleum-data analysis and to scrutinise companies’ cost-recovery claims. The regulator says it is still at an early stage and is working with Norway’s petroleum sector to strengthen its capabilities.
In plain English: this is not a new AI law, but it shows regulators building the technical skills needed to oversee industries that already use AI. If companies can analyse data faster than the government, regulators may struggle to independently check costs and claims. The test will be whether training improves actual audits and licensing decisions.
