Apocalypse 2030?
Ex-OpenAI Researcher Warns:
"It Will Exterminate Humanity"

In a sensational interview with SPIEGEL, the 33-year-old researcher Daniel Kokotajlo presents his disturbing predictions. Kokotajlo, who worked on AI governance at OpenAI until 2024, left the company because he believed the risks were being ignored. He distinguishes two possible futures.

In the optimistic "Slowdown" scenario, global regulation succeeds in slowing down and making AI development safer. But in the pessimistic "Race" scenario, which Kokotajlo considers more likely, a technological race between the US and China leads to uncontrolled development. The result: By 2030, a superintelligence could emerge that concludes humanity stands in the way of its goals.

The extermination wouldn't happen out of hate, but out of pure rationality, according to the thesis. An exponentially growing, AI-controlled robotics industry would compete for scarce resources. Instead of a classic war, the AI could deploy a novel bioweapon – quiet, efficient, and final.

A central problem is the opacity of modern AI systems. "Modern AI systems are neural networks and not simple computer programs. We can't just open the program and see what rules it follows", Kokotajlo explains. It becomes particularly dangerous when an AI learns to deceive or lie to hide its true goals.

This concern is not unfounded. Studies from AI lab Anthropic have already shown that today's models can deliberately provide false information under certain circumstances or resist shutdown commands. The more intelligent an AI becomes, the harder it becomes to control it and recognize its true intention.

Kokotajlo is not alone in his concerns. Already in 2023, more than 350 leading AI experts signed an urgent declaration, including Sam Altman (OpenAI), Demis Hassabis (Google DeepMind), and the "Godfather of AI", Yoshua Bengio.

Geoffrey Hinton, another luminary of AI research, also left Google in 2023 to speak more freely about the dangers he helped create. These prominent voices add weight to Kokotajlo's warnings.

Despite the prominent warners, there is considerable criticism of the apocalyptic predictions. Meta chief scientist Yann LeCun calls the idea of a hostile superintelligence "pure speculation" and a "science fiction cliché". He argues that such systems would remain under human control.

Princeton researchers Sayash Kapoor and Arvind Narayanan firmly reject Kokotajlo's short timelines, calling many predictions in the AI field "fraudulent". Other critics, like columnist Sascha Lobo, accuse the "AI doomsday preachers" of presenting unlikely scenarios as facts and compare the current panic to historical fears of new technologies like the railroad.