As world leaders and tech executives gather in London for the UK's first global AI safety summit, a co-founder of Anthropic has sounded a clarion call that cuts through the usual Silicon Valley optimism. Jack Clark, the company's policy chief, warned that artificial intelligence must not be allowed to develop in isolation from human oversight. His message arrives at a critical juncture when the technology is advancing faster than any legal or ethical framework can contain.
Speaking to delegates at Bletchley Park, the historic home of British codebreaking, Clark argued that the current trajectory of AI development risks creating systems that operate beyond meaningful human control. "We are building a new form of intelligence that could reshape every aspect of society," he said. "But if we proceed without embedding human values and oversight into the very architecture of these systems, we risk outcomes that none of us would choose."
The summit, hosted by Prime Minister Rishi Sunak, aims to establish international agreements on AI regulation. It comes amid a flurry of activity from companies like OpenAI, Google, and Meta, all racing to deploy ever more powerful models. Yet Clark's intervention highlights a growing divide between those who see AI as a tool to be managed and those who worry it may soon manage us.
Anthropic, the company Clark co-founded with former OpenAI researchers, has positioned itself as a safety-first alternative. Its flagship model, Claude, is designed with a "constitutional AI" approach that attempts to align machine behaviour with human preferences. But Clark admits that no current method guarantees perfect control. "We are essentially trying to build a superintelligent alien that shares our values," he told the BBC. "That is a hard problem. Maybe the hardest we have ever faced."
The warning is particularly resonant given recent events. Just weeks ago, a Google executive suggested that the company could not fully explain how its latest AI model arrived at certain conclusions. And OpenAI's ChatGPT has been caught generating biased or harmful content despite extensive safeguards. These incidents underscore the opacity of modern AI systems, which are often treated as black boxes by their own creators.
Clark's prescription is both radical and pragmatic. He calls for mandatory "red teaming" where independent experts probe AI systems for vulnerabilities before deployment. He also advocates for "interpretability research" to make machine decisions more transparent. But his most controversial proposal is for a moratorium on training models above a certain computational threshold until safety protocols are proven effective.
Such a pause would slow down the industry's breakneck pace. Critics argue that it would cede advantage to nations like China, which face fewer regulatory constraints. But Clark insists that the alternative is worse. "We are not talking about a few months of delay," he said. "We are talking about ensuring that the technology remains a servant, not a master. That requires hard choices."
The UK government has so far resisted calls for a complete halt, favouring a "light touch" approach that encourages innovation while establishing baseline rules. But Clark's speech may shift the debate. His stature as a co-founder of one of the world's most respected AI labs lends weight to his concerns. And the venue itself, Bletchley Park, serves as a reminder of what happens when intelligence is unmoored from ethics.
As the summit continues, the question hanging over proceedings is whether humanity can keep pace with its own creations. Clark's answer is clear: only by insisting that humans remain at the centre of the story. "The future is not written yet," he concluded. "But if we do not write it carefully, it may write itself."
