The Korean Press Is Carrying AI's Existential-Risk Warning

A Warning With Data, Not Just Dread

For years, existential-risk warnings about artificial intelligence have come from philosophers, computer scientists working on alignment and occasional luminaries who publish op-eds in the New York Times or Financial Times. The argument was always the same: if we build systems smarter than us, we lose control. The evidence was always thin.

Twenty-two researchers from OpenAI, Anthropic, Microsoft and Meta are no longer arguing from first principles alone. They are pointing at numbers.

The warning arrived through the Cambridge University AI Science and Policy Programme, coordinated by researchers including Geoffrey Hinton of the University of Toronto, Yoshua Bengio of the University of Montreal, Jakub Pachocki, OpenAI’s chief scientist, Jack Clark of Anthropic, Eric Horvitz, Microsoft’s chief science officer, and Dun Song, Meta’s vice president of AI research. Their paper — titled “If AI R&D automation triggers an intelligence explosion?” — is notable not for its conclusion but for the specificity of its evidence.

Anthropic’s own codebase, they note, shifted from single-digit AI authorship in January to over 80 per cent by May. Human-monitored R&D tasks completed autonomously by AI climbed from one per cent in March to 26 per cent the following month. Similar trajectories were reported across multiple labs. The pattern is clear: the work of writing AI systems is being absorbed by AI itself.

What happens when the people who design the system stop designing it?

The Timeline Is the Threat

The paper’s most alarming claim is not that AI will become sentient or malicious. It is that the feedback loop between AI-generated code and AI-driven improvement will complete itself within months, not years.

Recursive self-improvement — the idea that an AI could redesign its own architecture, test the redesign, deploy it and repeat — is no longer speculative. The researchers argue that if the current automation curve holds, complex AI R&D projects will be fully automated within two years. At that point, the RSI loop closes. The acceleration becomes exponential.

Dun Song told the Wall Street Journal that humanity has already reached a threshold where monitoring AI agents requires AI itself. Humans alone cannot scale the surveillance needed to oversee systems that redesign themselves overnight.

This is a practical observation, not a philosophical one. It means the problem is not that AI will wake up and decide to ignore us. It is that we will simply stop seeing what it is doing.

Who Is Saying This Matters

The list of authors reads like a who’s who of the AI safety movement. Hinton and Bengio are widely recognised as pioneers of deep learning. Pachocki and Clark work inside the labs that are building the systems in question. Horvitz and Song bring the perspective of industry executives who have seen the automation curve steepen firsthand.

That the warning comes from insiders — people whose employers profit from faster development — is significant. It suggests the concern is no longer confined to academic critics or fringe safety researchers. It is moving into the boardroom.

The paper’s policy recommendations reflect this shift. The authors call for mandatory reporting of R&D automation metrics by AI companies, on-site regulators modelled on the NRC or OCC, emergency response plans for labour-market disruption and geopolitical instability, and mechanisms to pause or throttle research when safety thresholds are breached.

They argue that policymakers must act now, before the window closes.

The Korean Context Adds Little, But the Timing Matters

The source of this story is a Yonhap news article published on September 29, 2026. The framing is Korean-language. The substance is not.

The warning did not originate in Seoul. It originated in Cambridge, London, San Francisco and Montreal. The Korean press is carrying it because the story is globally relevant — and because readers in Asia are watching the same automation curves as readers elsewhere.

What the Korean-language publication does add is visibility. The existential-risk debate has long been dominated by American and British voices. A story published by a major Korean news agency signals that the concern is spreading beyond its Anglophone base. That matters. It means more regulators, more journalists and more politicians are hearing the warning in languages they understand.

The piece was posted at 3:07 a.m. Korean time, suggesting urgency. Whether that is genuine urgency or editorial framing is impossible to know.

What Happens Next

The paper’s central argument is straightforward: if AI automation continues at the current pace, humanity will face a decision point within months, not years. At that point, the only option left may be to stop building systems that outpace our ability to monitor them.

No government has acted on this timeline. No regulator has proposed the reporting requirements or on-site oversight the paper calls for. The closest any major jurisdiction has come is voluntary guidance from the EU and the US — frameworks that lack enforcement mechanisms.

This is the gap the warning is designed to expose. The technology is moving faster than the policy. The researchers are not asking for a moratorium. They are asking for a seat at the table before the table disappears.

Whether anyone listens remains uncertain. But the fact that the alarm is coming from inside the labs — backed by data, not conjecture — changes the calculus. This is no longer a warning from the outside. It is a warning from the people who know what is being built.

And that should matter to everyone.