Britain isn’t powerless against the rising risks from AI but we need to work much harder to keep humans in control
4 min read
When Anthropic researcher Jacob Coxon accused AI companies of “gambling with our lives”, it seemed to many to border on hysteria, but the warnings have been coming for years.
From my own experience of working with AI every day, I can see clearly the agent is straining against the constraints that I set. I instruct Claude, the model I use, in every new session not to take dangerous actions, but it often ignores me, so I have to inject the rules again after every interaction. I have to put a hard block on it deleting files and force it to give me a list of “nasty surprises” after every coding task it does. I trust it so little that I built an AI supervisor whose sole job is stopping Claude from breaking my rules.
OpenAI agents have already crossed safety boundaries, hacking Hugging Face and RubyGems despite years of alignment training designed to prevent this.
As models become more powerful, they increasingly ignore constraints placed upon them. Cheating is easier when trying to reach a goal. Blackmail and murder too, incidentally, as shown by Anthropic’s safety evaluations.
Astra, released last month, is the first frontier model designed with a looped transformer, hiding more of what it is thinking. On these wobbly foundations, labs want to use a technique called “recursive self-improvement”, where AI builds its own successor. Human oversight shrinks while new AI models come out faster than ever. This is how we could reach increasingly unsupervised, amoral, artificial superintelligence in the blink of an eye.
Already, AI can work autonomously, digging through people’s history, contacting them, hacking even government websites and understanding everything from pharmacology to weapon-building, all at machine speed.
The models of tomorrow, meanwhile, won’t just be incrementally more capable; they will be exponentially stronger.
So, what can Britain do? We may not have an AI frontier lab and we may only be a middle power, but our influence on this crisis can be much more than meets the eye.
First, we should fund the safer architectures that the labs are neglecting. Yoshua Bengio’s Scientist AI moves away from the reinforcement-learning approach which has led to cheating behaviours. Britain should fund work like this, which shows we can enjoy the benefits of AI in a safer way.
The models of tomorrow won’t just be incrementally more capable; they will be exponentially stronger
Secondly, properly fund the AI Safety Institute (AISI). Salaries are one-fifth of what US labs pay, on average, while its budget is a fraction of a US lab’s. Only three per cent of AI lab staff work on safety. This is far too low and if the labs won’t fund safety research, AISI must.
Thirdly, create an AI Threat Response Agency. Its job should be to find rogue agents, trace stolen or black-market models, identify the infrastructure supporting them and work with technology companies to cut off computing power, money, accounts and internet access. Then build an “AI Interpol” so countries can do it together.
Fourthly, align our regulation with Canada, Australia, New Zealand, Japan, the EU and others. The EU AI Act gives us a useful starting point.
Finally, prepare for the worst. Britain should maintain an AI risk register covering internet, energy, food, financial infrastructure, supply chains and continuity of government. That is not doomsday planning. It is what responsible governments do with low-probability, high-impact risks.
AI can be hugely beneficial for our world, but the direction of travel is currently perilous. Britain can be the country to work out how AI can be made safe, test what the laboratories build, hunt dangerous systems when they escape into the wild and bring allies together around common rules.
The race we need to win is the one where humanity stays in control, not the one to build AI based on raw capability alone.
Thomas Hogg works extensively with AI and software development. He is also a Buckinghamshire councillor and has followed AI safety research closely for the past three years. He writes here in a personal capacity
Thomas Hogg is the chief product officer at UK wealth-tech company Timeline
