Britain’s Business, Innovation, Science and Trade Committee has set out a long-term AI safety inquiry after taking evidence on 13 October. Its 14 October announcement puts independent testing and the powers to intervene when systems become unsafe under continuing parliamentary scrutiny.
The inquiry covers mandatory incident reporting and responsibility for losses when safeguards fail. Its scope also asks whether evaluators have enough access to training processes and whether their findings can be reported without interference. It will examine whether contract exclusions leave consumers and small businesses without meaningful remedies, alongside the question of whether voluntary company commitments offer sufficient protection.
A ban or enforceable safeguards?
In written evidence dated 13 October, ControlAI’s Andrea Miotti and Connor Leahy argue for prohibiting the development of superintelligent AI and pursuing an international agreement. They distinguish those systems, intended to outperform people across tasks and act autonomously, from specialised applications used in science or public services. Their submission acknowledges that superintelligent systems do not yet exist. This is a recommendation, not an enacted ban.
ControlAI also challenges two narrower approaches: compulsory testing before deployment and agreements among developers to slow progress. It argues that testing needs enforceable boundaries on development and deployment, because identifying danger alone does not prevent it. The authors say slowing development leaves companies pursuing the same endpoint and allows participants to abandon voluntary cooperation under commercial pressure. Their approach would restrict dangerous capabilities before they emerge, going beyond disclosure obligations.
Its proposed boundary is whether a system could undermine human authority, rather than a fixed technical specification. The submission would also give ministers oversight of precursor capabilities and computing resources. It pairs domestic restrictions with negotiations abroad, annual reporting to Parliament and preparedness plans for national infrastructure.
Anthropic’s 9 October submission supports binding obligations for frontier developers above a certain size. It proposes published safety frameworks, model documentation and regular risk reports, alongside independent evaluation and secure development environments. The company says voluntary frameworks apply only to organisations that choose to adopt them.
It argues that governments must set industry-wide requirements and negotiate international arrangements. The company wants common safety standards while continuing AI development. It says Mythos Preview was restricted to vetted cyber-defence and critical-infrastructure organisations instead of being released generally.
The submission also identifies responsibility inside Anthropic. Decisions on model releases rest with its chief executive and Responsible Scaling Officer, with oversight from the board and its independent Long-Term Benefit Trust. Its voluntary Responsible Scaling Policy links assessed risks to protections during training, deployment and subsequent monitoring.
Anthropic says it publishes system cards for major releases explaining evaluations, external tests and safeguards. Anthropic says it has voluntarily worked with UK and US government evaluators on pre-deployment testing since 2024. Its proposals seek legal requirements beyond that voluntary cooperation.
Duties and development controls
OpenAI’s submission dated 5 October, published by Parliament on 8 October, backs mandatory UK requirements tied to the capabilities of frontier systems. It wants obligations concentrated on the small group developing the most advanced models and proportionate to their risks.
The submission prioritises national-security concerns, including advanced cyber capabilities, biological risks and loss of control. It also supports compatible international safety frameworks, information-sharing channels and coordinated responses to serious incidents. It says the requirements should build on voluntary cooperation between frontier laboratories and government.
OpenAI describes stronger isolation of frontier research workloads, expanded monitoring during tool-enabled training and evaluations, and clearer escalation procedures. For its Astra model, it says it introduced activity monitoring and an alignment-evaluation gate before broader internal deployment.
It also says it has previously paused some frontier training while strengthening safeguards. Following reported cybersecurity incidents, it says it sought independent review and began examining past model activity. Those statements describe its claimed response; they do not amount to an independent finding that the underlying problems have been resolved.
Dalimss News previously covered the separate UK government cyber-defence and near-miss reporting plans.
Questions about evidence and practical deployment also feature in Dalimss News’s report on the India-hosted discussion on agricultural AI standards.
Targeted evidence calls will follow, rather than one immediate general submission deadline. The inquiry will remain open to revisit unresolved issues. The announcement described the opening hearing’s transcript as forthcoming.
