Artificial intelligence may bring great benefits. It may also cause grave harm. The question is who should control it and how.
A new sense of alarm
Warnings about artificial intelligence are no longer confined to academics and campaigners. Some now come from the people building the most powerful systems.
In September 2026, Dario Amodei of Anthropic called for companies to slow the growth of their most advanced models. Sam Altman of OpenAI and Elon Musk of xAI backed the broad idea. Jack Clark, another Anthropic founder, said powerful systems might need a compulsory ‘kill switch’ which an outside body could check.
Microsoft AI has also published a draft code for what it calls humanist AI. Its central rule is plain: machines must remain under human control. They must accept correction, obey a shutdown and never claim rights of their own.
These statements matter because they come from rival firms racing for money, talent and power. Yet they should not be taken on trust. Companies may want sound rules. They may also want rules that favour large firms and keep out smaller rivals. Their claims need outside scrutiny.
What are the risks
The debate often mixes two kinds of danger.The first kind is already here. AI can create false images, imitate people, spread lies and help fraudsters. It can invade privacy and support cyber-attacks.
Employers and public bodies may use it to make decisions which people cannot understand or challenge. An answer may sound confident while being false. These are present harms, not science fiction.
Future systems may become able to plan and act with little human help. They could be used to design weapons, attack computer networks or disrupt essential services. Some researchers fear that a highly capable system might resist control or improve itself faster than its makers can understand it.
Claims that AI could destroy humanity are forecasts, not facts. Experts disagree sharply about their likelihood. Some think the danger is real and urgent. Others think the industry exaggerates the power of its machines. But a low chance of immense harm still deserves serious study. Uncertainty is not a case for doing nothing.
Agents have changed the argument
Most people use AI to answer questions, draft emails or make pictures. An AI agent goes further. It can take a goal, choose steps and act through other computer systems.
That can magnify mistakes. A chatbot that gives a wrong answer misleads one person. An agent with access to accounts, code or private data may act on the mistake before anyone notices. Recent security incidents have shown systems behaving in ways their makers did not expect or request.
This shows that powerful tools can behave unpredictably, and present safeguards are not always enough. It does not show that machines are about to take control.
What would a slowdown mean
A slowdown need not mean a halt to research. Amodei and Altman have argued for more time to test the strongest models before release. They want closer monitoring and checks by independent experts.
A sensible system would require firms to test dangerous abilities, report serious incidents and show that safeguards work. Outside inspectors would need secure access to evidence which companies cannot publish. Governments would also need the power to enforce common standards. Voluntary promises are useful, but firms should not mark their own homework.
Rules should bear most heavily on the small number of companies building the strongest models. They should not burden every small business using an ordinary AI tool. Regulation should follow risk, not size alone or political fashion.
The race with China
The hardest problem is international rivalry. President Donald Trump has dismissed many warnings as a hoax. His case is that America must beat China. Some Republicans fear that strict rules would smother American invention. China, meanwhile, calls for global governance while pressing ahead with its own systems.
The fear is understandable. If one country slows down alone, another may gain a military and economic advantage. Yet a race without rules carries its own danger. Speed can reward secrecy, weak testing and reckless release.
No lasting safety system can exclude China. The United States, China, Britain and the European Union do not need to agree on every political question. They do need minimum rules on the most dangerous uses of AI, just as rivals have reached limited agreements on nuclear weapons, aviation and disease.
Political shift?
The argument does not fall neatly between left and right. In Washington, Bernie Sanders and Steve Bannon have both attacked the power of large technology firms, though for different reasons. Democrats and Republicans have proposed the FRONTIER Act, which would create a national system to oversee the most powerful models.
Britain faces the same choice. It wants investment and growth, but its AI Security Institute cannot rely on goodwill alone. Parliament’s human rights committee has warned about surveillance, deepfakes and the use of AI against workers. If testing is important, the body responsible must have the authority and skill to do it.
A sensible course
The choice is often put badly. We are told to choose progress or safety. We need both. Unsafe systems will not earn public trust, and rules which freeze useful work will fail.
Governments should demand independent tests, clear reports of serious failures and a reliable means of shutting down dangerous systems. Firms should limit what agents can reach and do. People should remain responsible for important decisions in health, justice, finance, war and public services.
Above all, we should resist two easy stories: that AI will solve every problem, and that it will soon kill us all. Both turn uncertainty into certainty. The honest position lies between them. AI is powerful, fallible and developing fast. Human beings built it. Human beings must remain in charge.
We can only hope that Phil Oakey of the Human League and Giorgio Morodier had forward vision in September 1984 when they wrote and sang: “We’ll always be together … together in electric dreams”.
SourcesReuters, Microsoft drafts code of conduct to keep its AI under human control, 14 September 2026Reuters, Anthropic CEO urges AI companies to slow model development, 12 September 2026US Representatives Jay Obernolte and Lori Trahan, announcement of the FRONTIER Act, 23 July 2026
