Jack Clark, co-founder of the frontier AI lab Anthropic, has issued a stark warning that artificial intelligence must not be developed in isolation from human oversight. Speaking exclusively to this publication, Clark argued that the current pace of advancement risks creating systems that operate beyond our ability to understand or control. His comments come as Britain positions itself as a global leader in AI regulation, a stance Clark cautiously endorses.
Clark’s concerns crystallise around the concept of ‘alignment’: ensuring AI systems act in accordance with human values. “We are building tools that could reshape every aspect of society,” he said. “If we proceed without embedding human feedback loops, we risk creating black boxes that make decisions we cannot interrogate.” He points to recent incidents where AI chatbots have exhibited unpredictable behaviour, from generating hate speech to suggesting harmful actions. These are not bugs, he insists, but feature failures in our governance models.
The United Kingdom, eager to forge a third way between Silicon Valley’s laissez-faire attitude and Brussels’ prescriptive regulations, has proposed a framework built on existing regulators. The AI Safety Institute, launched this year, aims to test frontier models before deployment. Clark welcomes this, but with a caveat: “Testing is necessary, but not sufficient. We need continuous human-in-the-loop systems, not just a certificate at launch.”
His vision is ‘participatory AI’, where citizens have a say in how these systems are deployed in their communities. This resonates with the government’s ‘pro-innovation’ approach, but Clark fears that industry lobbying may dilute requirements. “The temptation is to treat AI as a purely technical problem,” he warns. “But it is a societal contract. We do not let pharmaceutical companies release drugs without ongoing monitoring. Why should we treat AI differently?”
Britain’s regulatory lead, the Department for Science, Innovation and Technology, declined to comment on specific proposals, but sources indicate that ‘dynamic oversight’ is central to their strategy. This would involve real-time auditing of AI decision-making, particularly in high-risk areas like healthcare and criminal justice. For Clark, this is a step in the right direction, but he presses for more: “We need mandatory ‘explainability’ standards. If an AI denies a loan or recommends a prison sentence, the affected individual must be able to understand why.”
Critics argue that overregulation could stifle innovation and push AI development to more permissive jurisdictions. Clark rejects this as a false dichotomy. “We don’t need to choose between safety and progress. We need to build human oversight into the development cycle from day one. That is not a cost, it is an investment in trust.” He cites Anthropic’s own “Constitutional AI” approach, where models are trained to follow a set of principles and can explain their reasoning. The technique, he says, shows that safety and capability can coexist.
As the government prepares to host a global AI safety summit, Clark’s message is clear: the window for responsible development is narrow. “We are writing the rulebook for a technology that will outlast us,” he concludes. “Let us ensure that humans remain not just in the loop, but at the centre of the story.”
