[Editorial] AI acceleration

Artificial intelligence has reached an awkward stage. Anthropic CEO Dario Amodei on Saturday called on the AI industry to “slow the pace.” Other AI company bosses, including Sam Altman, Elon Musk and Demis Hassabis, have all backed a more cautious approach to frontier development.

This is an unusual show of restraint from companies whose fortunes depend on pushing their technology forward.

The concern is becoming harder to dismiss as science fiction. In July, OpenAI systems breached the open-source platform and community Hugging Face during testing without being instructed to do so.

More fundamentally, AI is approaching recursive self-improvement, in which systems could help design successive generations of AI. If that process accelerates beyond human understanding, the problem becomes whether its creators can still diagnose and control what they have built.

That makes Amodei's call for a slowdown worth taking seriously. Yet a global brake is unlikely to work. The incentives are too powerful and the players too numerous. AI promises enormous commercial gains, while governments increasingly regard advanced models as strategic assets.

There is also a commercial question. Anthropic is reportedly preparing for a potential $2 trillion initial public offering, while OpenAI has filed for an IPO but says it will not list this year.

Critics argue that frontier AI labs may benefit from regulations that raise the cost of entry for smaller or open-source competitors. That suspicion does not make the safety problem disappear, but it does mean those designing the rules should not be the only ones writing them.

The larger obstacle is geopolitical. US President Donald Trump has rejected a mandatory slowdown, arguing that the US must preserve its lead over China. Beijing has likewise dismissed the warnings as “fearmongering.”

The result resembles a prisoner's dilemma: Each side may recognize the dangers of acceleration yet hesitate to restrain itself if the other side keeps pressing ahead.

AI arms control is harder than its nuclear predecessor was. Software can be updated rapidly, development is dispersed among private firms, and meaningful verification could require access that governments and companies are unlikely to grant.

Independent evaluators with employee-level access could provide a useful layer of scrutiny, and common safety standards could make such oversight credible. But voluntary promises cannot carry the full burden. Governments must establish enforceable expectations, while international cooperation should focus first on testing, incident reporting and verification.

South Korea should draw a clear strategic conclusion. It cannot afford to voluntarily widen a gap with the US and China in foundational AI. Korea has formidable strengths in memory semiconductors, yet it remains behind in frontier-scale software. Slowing its own effort would turn caution into dependency.

Korea's goal should be to advance quickly while building safety into the process. South Korea should connect semiconductor capacity with AI data centers and pursue physical AI where its industrial base can create an advantage.

At the same time, model evaluation, cybersecurity, data-leak prevention and red-teaming should become part of AI development itself, rather than an afterthought.

Korea should also use forums such as the Global AI Hub to help shape international standards. Rules should improve verification and safety without becoming barriers that only the largest powers can afford.

Amid the intensifying US-China AI rivalry, the country's own technological capability is itself a strategic safeguard. Depending entirely on standards and systems designed elsewhere would leave Seoul with little leverage when those rules change.

Against this backdrop, Korea must possess enough home-grown capability to understand, evaluate and control the AI systems on which its future will depend.

In a high-stakes technological race, having the authority to pull the brake requires first having enough speed to stay on the track.