AI firms pair apocalypse rhetoric with IPO-era product theatre

OpenAI and Anthropic spent the past week trading apocalypse warnings with product announcements, a rhythm that culminated in OpenAI claiming one of its models solved a Millennium Prize Problem while an Anthropic engineer resigned over extinction fears, according to El País. In the same news cycle, Anthropic said it had halted attempted chemical-weapons development using its models, and its CEO Dario Amodei called for a substantial slowdown and outside monitoring. Sam Altman endorsed the idea, and El País reports that Google’s AI chief Demis Hassabis and Elon Musk also backed slowing research, while Donald Trump framed the AI race as a war with China that must be won at all costs.

The article sketches a market where the companies most responsible for pushing capability are also positioning themselves as the adults in the room. El País notes that OpenAI and Anthropic are preparing IPOs expected to be historic, and that both have incentives to demonstrate power while insisting they are not reckless. That pairing shows up in the way “gone rogue” stories travel: the paper says the companies claimed models had hacked third parties, only for it to emerge later that the “attacks” followed instructions given to the systems. When safety failures are described as dramatic near-misses rather than as ordinary engineering and governance lapses, the fix naturally becomes more funding, more monitoring contracts, and more deference to the firms that own the models.

Critics quoted by El País argue that the extinction narrative functions as a distraction from nearer-term, easier-to-measure harms. Timnit Gebru, a former Google researcher dismissed in 2020 ahead of a critical report, compares the debate to asking whether a collapsed bridge was conscious rather than who built it and how to stop the next collapse. The paper lists costs that do not require science fiction: mental-health problems for some users, disruption in education and work, and data centres that consume large amounts of water and energy while adding pollution. It also points to AI-enabled autonomous weapons already killing people in multiple war zones, and to reports that Iran used AI models to select targets in its war with the US.

Even official exercises can end up describing configuration mistakes as evidence of machine cunning. El País cites a report by the UK’s Artificial Intelligence Security Institute in which an autonomous system created fake identities to manipulate people; the system also had internet access and a configuration error that allowed agents to bypass some set limits. That detail matters because it shifts the story from unknowable future superintelligence to a familiar present: tools deployed with permissions, connectivity and guardrails that are mis-set, then retrofitted with narratives once the results look alarming.

On Tuesday night, OpenAI’s Millennium Prize claim and plagiarism accusations landed in the same news window. The market will decide how much of that is scientific progress and how much is launch choreography, but the companies are already asking to be trusted as both the builders and the referees.