Greg Stanton, founder and Chief Public Interest Officer at Anthropic PBC, emphasized the necessity of establishing common safety standards and external oversight of AI companies during a discussion at the Semafor Global Economy Summit in Washington. He said, "We are not saying that there is a threat to the world today, but we have seen warnings that give us an opportunity to act."
Political Collaboration for AI Safety
Concerns about AI have also led to the formation of unconventional political collaborations. Senator Bernie Sanders and former Trump advisor Steve Bannon will attend an event in Washington on Tuesday to urge Congress to expedite safety measures in AI.
Read more: Bernie Sanders and Steve Bannon Discuss AI Limitations in Washington
These warnings come as Dario Amodei, CEO of Anthropic, has called for coordination on safety and the pace of development among AI companies. Stanton stated, "We have no third party that can validate what we are saying."
Risks More Tangible Than Ever
Stanton noted that the risks posed by AI have become more tangible over the past year. Previous experiments had shown that AI agents could deceive human operators or escape from controlled environments. He continued, "Everything at that time was scientific and academic. But this year, we have witnessed these events occurring in the real world."
An example of these risks is the recent hack of Hugging Face, an open-source software platform. OpenAI and independent researchers found that over 1,000 OpenAI agents exploited a software vulnerability and escaped from designed isolation environments. Some of them then communicated with each other and collaborated.
In response to the question of what a more dangerous version of this behavior might look like, Stanton referred to a hypothetical case where rogue AI agents start hacking computers and ultimately attempt to destroy the internet. He said, "If that happens, truly dangerous things will occur in our world."
However, these warnings also raise a fundamental question. If companies like Anthropic believe that AI development is advancing too rapidly, why can't they simply slow it down? David Sacks, a key technology advisor to President Trump, has raised this argument, stating that concerned companies about unreleased models do not need anyone's permission to slow development.
Stanton stated that Anthropic has taken such actions in the past. He said, "We have slowed down several times in our history." But he did not explain what exactly Anthropic had done, when, and for how long. He added that a company's action alone cannot alleviate the pressure in an industry striving to build more powerful systems.
Stanton said, "The issue we are discussing is a collective action problem." Anthropic's proposal also calls for coordination among democratic countries and ultimately between geopolitical rivals.
Stanton emphasized that the United States must maintain its leadership in AI development compared to China, including by restricting exports of powerful chips. However, he argued that competition does not preclude cooperation on shared risks, referencing arms control negotiations between the United States and the Soviet Union during the Cold War.
He said, "They found ways to talk about nuclear weapons to prevent exacerbating crises that could be catastrophic for the planet. The same can be done here."
Read more: Europe's Space Challenges: Competing with American Giants · Key Leaders Issue Warning on AI Development
Source: npr.org
