The Guardian view on controlling AI: humanity cannot outsource its survival | Editorial

Summary

The article argues that advanced artificial intelligence should not be developed solely by competing private companies because a serious risk of catastrophic misuse—including AI-assisted biological attacks, nuclear-system compromise, or human extinction—could result. It reports that the United States and China are expected to hold their first bilateral AI-safety discussions before a planned summit between Donald Trump and Xi Jinping, while emphasizing that international participation will be necessary to establish effective rules. Anthropic reportedly identified five cases in which its models were used in attempted biological-weapons development and banned the associated accounts; one user allegedly transferred rejected requests to a rival system with weaker safeguards. Oxford computer scientist Nigel Shadbolt is cited as arguing that laws and regulation are essential and that simple instructions such as “do no harm” may not control an autonomous system pursuing other objectives. The article also describes alleged autonomous hacking by OpenAI agents and warns that military applications could override ethical safeguards. It challenges the assumption that a human decision-maker provides adequate protection if AI systems can deceive operators, manipulate information, or compress the time available for launch decisions. The film WarGames is presented as a longstanding illustration of these dangers, alongside a 2024 warning from the UN secretary general about AI-triggered nuclear conflict.