Transformative AI has arrived. It demands urgent global action.
First published in The Boston Globe
Anthropic’s recent announcements regarding its powerful new AI model, Claude Mythos, are not just a wake-up call: They are a fire bell in the night. Anthropic stated that Mythos is too dangerous to be released for public use. Similar breakthroughs will surely soon follow, both from American competitors (OpenAI already announced its version) and from China. The debate over whether transformative AI is here is settled: It has arrived.
Critics may argue that the risks are overstated, which would entirely miss the point. Whatever Mythos can or cannot do today, the next versions and those to follow will be able to do that and much more. The risks are not limited to “rogue actors” — terrorist networks or hostile nations — that will inevitably seek access to these tools. (In fact, Anthropic has confirmed it is investigating a report that unauthorized users have gained access to its Mythos model.) The risks stem equally from the well-intentioned who deploy them and the users who operate them without considering the potential consequences.
The technology seems to be advancing at a pace that outruns our collective ability to comprehend what is being unleashed. The dangers are real. Doing nothing in response is not an option, but neither is panic. This moment demands thoughtful, parallel action on three fronts.
US-China dialogue on existential risk
The most urgent need is a bilateral conversation between Washington and Beijing, focused on the shared dangers these technologies pose to each nation and to global stability. Both governments must work toward agreed guardrails, defining not just how this technology should be used but where it must never be applied. Red lines need to be defined, established, and agreed upon.
This conversation does not yet exist because the requisite mutual trust does not yet exist. Building it will not be easy. Both nations must recognize that racing for AI dominance without regard for proliferation or misuse is an unacceptable risk.
Would China engage? Evidence suggests yes. Beijing wants to be seen as a responsible global power and would recognize the importance of helping establish critical global standards. Furthermore, the leadership in China most likely harbors its own anxieties about AI’s potential to destabilize society. This window of potential engagement may be narrow: If China surges ahead before an agreed-upon framework is established, its leaders may choose to set the global terms of governance alone.
Would the United States engage? I certainly hope so. The risks of not engaging to establish the appropriate global guardrails are much too high.
Ideally, at the right time, this conversation should be expanded to include the G20 nations if not the whole United Nations.
Informal diplomacy: the expert exchange
Before governments can sign treaties, practitioners must build bridges. We need sustained, informal engagement among nongovernment AI experts, such as academic researchers and industry experts, from both sides. Such engagement can surface concrete technical recommendations and build the human relationships that make formal diplomacy possible. While some of this work has begun, it needs to be expanded, supported, focused, and integrated into US national strategy.
Getting the US house in order
Independent of international progress, the United States must move urgently at home. A representative panel — drawn from industry, research laboratories, and academia — should be given a firm mandate and a short (e.g., three months) deadline to deliver actionable recommendations to Congress. The charge is as simple to state as it is difficult to execute: How do we protect the United States, including its economy, military, and society, while preserving its capacity and ability to continue to advance and lead in AI?
Because this technology is continuously accelerating, the players involved in these three tracks should reconvene at least annually to reassess their findings and recommendations.
Time is of the essence.
L. Rafael Reif is president emeritus and the Ray and Maria Stata professor of electrical engineering and computer science at MIT.