Leading AI Doomsayer Rethinks the Timeline for Humanity’s End
Humanity’s ultimate demise at the hands of AI sounds like a sci-fi trope, but today experts debate its very real probability.
For much of the past year, a figure whose name has been spoken time and time again in discussions about artificial general intelligence (AGI) and human extinction is Daniel Kokotajlo.
A prominent voice in the AI safety community and former OpenAI researcher, Kokotajlo helped spark a global conversation last April with “AI 2027,” a speculative scenario that he coauthored in which rapidly advancing AI systems outsmart humanity, seize control of global decision-making, and ultimately wipe out the human race. Attention-grabbing, right?
Well, now, let’s just reel in the panic for a minute, as Kokotajlo is revisiting that prediction. In a recent update, he and his collaborators are backpedaling, saying the path to superintelligent AI appears to be taking longer than they initially expected. And while they argue that the most dramatic risks have not vanished, they say that they may be a bit delayed, rather than following the apocalyptic AI 2027 timeline that captured so much attention.
The recent update revises the expectation for when AI systems may achieve fully autonomous coding, a key aspect of the originally predicted scenario. This would trigger an intelligence explosion, as AI systems would be able to rapidly improve without meaningful human oversight. While they predicted that this would likely occur by 2027, it seems now that that prediction was too optimistic. Or, pessimistic, if you just so happen to enjoy the human race.
“Things seem to be going somewhat slower than the AI 2027 scenario,” Kokotajlo wrote, adding that even at the time of publication, 2027 was only a central estimate, and not the certainty that many assumed. The revised outlook places autonomous AI research in the early 2030s, with “superintelligence” potentially emerging closer to 2034. And if this concerns anyone, unlike the original scenario, the update does not attempt to predict if or when AI might destroy humanity.
The limits of modern AI progress
Currently, there is broader skepticism about predicted timelines for AGI. Since the release of ChatGPT in 2022, predictions about human-level AI have accelerated dramatically, with some experts suggesting it could arrive within years. But recent experience has exposed the uneven nature of AI progress, and what critics describe as jagged intelligence, or systems that can perform cognitive tasks but struggle with things like basic reasoning, long-term planning, or autonomous action.
Malcolm Murray, an AI risk expert and contributor to the International AI Safety Report, says this is pushing many forecasts further into the future. “For a scenario like AI 2027 to happen, you’d need AI systems with a lot more practical, real-world skills,” he said. “There’s enormous inertia in society — institutions, infrastructure, politics — that slows everything down.”
Others question whether the very concept of AGI still makes sense, and argue that even the language used in these debates is becoming outdated. Henry Papadatos, executive director of the French nonprofit SaferAI, questions whether “AGI” still has a clear meaning. “When AI only played chess or Go, the term made sense,” he said. “Now systems are already quite general, but in messy, incomplete ways. The label doesn’t capture that very well.”
Industry goals and policy reality
Still, the race to build AI capable of automating its own development remains a significant objective for leading companies. OpenAI chief executive Sam Altman said in October that building an automated AI researcher by March 2028 is an internal goal, while acknowledging that the company could just fail outright. But if successful, such systems could dramatically accelerate AI progress, even if they fall short of the 2027 science-fiction predictions.
Policy experts caution that intelligence alone does not equal real-world power. Andrea Castagna, a Brussels-based AI policy researcher, notes that advanced AI would still need to be embedded in complex human systems. “You can’t just drop a superintelligent computer into military or government structures that have evolved over decades,” he said. “The world is much more complicated than the scenarios people imagine.”
Why Kokotajlo still matters
Kokotajlo’s influence is earned in part thanks to his track record. He earned his credibility in a 2021 blog post, when he accurately anticipated the explosion of chatbots and the importance of “inference-time scaling,” a technique that is now central to modern AI systems. But at the same time, he has consistently emphasized that AI 2027 is not a prophecy, but one potential future among many possibilities.
The AI 2027 essay itself includes an alternate ending: governments intervene, international agreements limit AI development, and systems are deliberately constrained to remain interpretable and aligned with human values. Thus, in this alternate ending, humanity survives, though difficult political questions remain unresolved.
Competition and the road ahead
Critics like NYU neuroscientist Gary Marcus have dismissed Kokotajlo’s more dire scenarios as speculative fiction, arguing that doom-laden narratives distract from immediate concerns such as bias, misinformation, and job disruption. Others worry that these dramatic timelines could foster fatalism or even encourage reckless acceleration in the name of “winning” the AI race.
But for Kokotajlo, pushing timelines back does not mean resignation. If anything, he argues, it creates a narrow window for action. Whether transformative AI arrives in the late 2020s, the mid-2030s, or even later, the fundamental issue of ensuring that advanced AI systems reliably act in humanity’s interests remains unresolved. That said, the debate Kokotajlo helped ignite is unlikely to simmer down anytime soon.
Also read: AI p (doom) has moved from a niche idea to a mainstream debate as more leaders weigh the odds of AI-driven catastrophe.
The post Leading AI Doomsayer Rethinks the Timeline for Humanity’s End appeared first on eWEEK.