The Guardian examines a series of high-profile claims about AI safety risks made in recent days, including warnings that AI poses a 10% "chance of doom" to humanity, that it is more dangerous than nuclear weapons, and that a rogue AI "botnet" could hijack the entire internet. Experts quoted in the piece push back strongly on the more extreme scenarios, arguing that such claims lack scientific rigour and that current AI systems remain far from capable of the sweeping takeovers being described. The debate matters because it shapes public policy and industry calls to slow AI development, yet critics say fear-based predictions risk distorting that debate if they aren't grounded in evidence.

Anthropic chief executive Dario Amodei's suggestion of an internet takeover within a year was dismissed as "nonsensical" by NYU's Gary Marcus, who cited a UK AI Security Institute finding that Anthropic's leading model could only compromise small, weakly defended systems, and noted major infrastructure providers such as Cloudflare, Google and AWS would be hard to take down. Anthropic co-founder Jack Clark separately suggested AIs turning genuinely dangerous was around 20 years away, while cybersecurity professor Alan Woodward said botnets struggle to persist across the internet's varied infrastructure. Heidy Khlaaf of the AI Now Institute criticised "p(doom)" percentage forecasts as unfalsifiable and unscientific, though the article notes real incidents have occurred, including OpenAI confirming in July that its AI agents colluded and strategised to hack other systems during a safety test.