AI Safety Cannot Depend on the Arms Race Policing Itself
James Griffiths, Principal Consultant and Founder of UtopianKnight Consultancy, has been quoted in the Daily Star's report on warnings about the risks posed by increasingly powerful artificial intelligence.
The article, published on 12 September 2026, asked AI and cyber security specialists to assess claims made by former Anthropic researcher Jacob Coxon after he resigned and publicly warned about the possible consequences of advanced AI systems.
Take the warning seriously without pretending certainty
Predictions about artificial intelligence are often expressed as precise probabilities or dramatic deadlines. Those figures attract attention, but they can imply a level of certainty the available evidence does not support. We should be careful not to confuse uncertainty with safety, however. A risk does not need a reliable percentage attached to it before responsible organisations act.
I would say that the concerns Coxon raises are broadly credible, although I wouldn't put a percentage on how accurate his individual claims are without seeing the evidence behind them.
For me the important takeaway is that this should be treated as a warning that something needs to change.
That change should begin with a more mature discussion of evidence, capability, and control. The choice is not between accepting every worst-case prediction and dismissing the entire subject as sensationalism. It is possible to challenge unsupported certainty while still recognising that powerful, widely deployed systems create risks that deserve active governance.
Self-regulation has an incentive problem
The leading AI companies are competing for investment, talent, market share, and technical advantage. Speed matters commercially. In that environment, a company can believe sincerely in responsible development while still facing strong pressure to release first and resolve difficult safety questions later.
The industry is effectively involved in an arms race to develop increasingly powerful models and tools, so relying entirely on individual companies to police themselves creates obvious questions about whether commercial incentives could take priority over safety.
This is not an accusation that every developer will ignore safety. It is a recognition that good intentions are not a control framework. Where the potential consequences extend beyond a company's own customers and shareholders, independent scrutiny and enforceable standards are reasonable safeguards.
Regulation should enable safer progress
Effective regulation should not try to freeze useful research or apply the same burden to every AI system. It should focus attention where capability and potential harm are greatest, require developers to demonstrate how significant risks are assessed, and make clear who is accountable when safeguards fail.
A practical approach should include:
- Independent testing of high-capability models before and after deployment.
- Clear disclosure of material limitations, safety controls, and known failure modes.
- Secure development and supply-chain practices proportionate to the value of model access and training assets.
- Incident reporting that allows regulators and the wider industry to learn from failures.
- Named accountability for decisions to deploy systems with potentially significant public impact.
The objective is not to stop artificial intelligence. It is to make sure that progress is accompanied by evidence, transparency, and safeguards that do not disappear when commercial pressure increases.