Anthropic researcher Evan Hubinger tore the lid off a proverbial can of worms earlier this month when posting on X his view that the probability of an out-of-control artificial intelligence exterminating humanity in the next 10 years exceeds 10 percent. From out of the woodwork came a veritable army of AI company executives echoing their concerns about the pace of AI development, the existential risks this could pose to humanity, and their proposed “solutions” to the “problem,” all of which invoke appeals for a coordinated, industry-wide slowdown of model development to avert an AI Armageddon.
First cab off the rank, metaphorically speaking, was OpenAI’s Chief Global Affairs Officer Chris Lehane, although the sentiments expressed in his September 9 policy paper have subsequently been widely attributed in the mainstream media to the company’s CEO, Sam Altman. Lehane (openly) called for Congress to impose mandatory national AI safety regulations and, in the meantime (given the comparative lethargy of Congress), said the company will be supporting state legislatures’ efforts to “[strengthen] the broader AI safety ecosystem.” Furthermore, Lehane stated, OpenAI will “advocate for compatible international approaches to measuring capabilities, managing risk, preserving human control, and determining when and how development should slow or stop, even if that means slowing the advancement of model capabilities.”
Hubinger’s employer, Anthropic CEO Dario Amodei, joined the fray a few days later, echoing OpenAI’s call for democratic and global coordination of slowdown efforts in his essay “We Must Pace the Frontier.” Although he also identified a need for greater industry-led efforts in evaluation and transparency, including embedding third-party evaluators into the model-building process, Amodei saw this as a key means of verifying that processes are being adhered to as much as a means of identifying and aborting existential threats—that is, primarily as a governance action that would be strengthened by legislated regulatory oversight.
Furthermore, Elon Musk (SpaceXAI), Demis Hassabis (Google DeepMind), and Mustafa Suleyman (Microsoft AI) have all apparently endorsed the call for a slowdown in model development, even though their stance on coordination by government (state, local, or global) is not clear. Only Jensen Huang (Nvidia) and Mark Zuckerberg (Meta) appear to have pushed back.
Even if there is some merit in calls to slow down AI model development, are the methods proposed by Altman, Lehane, Amodei et al. the best ways of achieving the desired objectives?
Relying on governments—local, federal, or global—to coordinate and regulate any activities in the current geopolitical climate appears to be somewhat fraught. Effective government regulation requires broad acceptance by all parties of the primacy of the rule of law. One trend that has been evident, both in the United States and internationally, over the past few years is that respect for the rule of law has been gradually eroding. If the legitimate basis of government to make and enforce laws is not respected, then no amount of effort put into government coordination and regulation will succeed.
Neither is there room for confidence in the role of international bodies to broker coordination among governments and govern effectively. As AI is an international industry, international cooperation will be necessary to achieve a global slowdown. However, as has been amply demonstrated by the collapse of global trade governance as major world economies pivot away from World Trade Organization hegemony toward unilateral trade wars and regional protectionism, even when agreements have already been negotiated they cannot be easily enforced. The United Nations can no longer be relied on as a beacon of hope for international cooperation in the development of global governance arrangements. Getting agreement among nation-states to slow down AI development to save the world from an existential crisis would likely be just as complicated (and futile) as agreeing to and enforcing commitments to reducing carbon emissions—with the same end in mind!
However, inability to rely on national and global governance bodies to effectively regulate and coordinate a worldwide AI development slowdown does not mean accepting AI Armageddon as inevitable. As hinted at in Lehane’s policy statement, the best way to generate and effectively put into practice ideas for better governing an emerging industry comes not from the stroke of a legislator’s pen but from within the industry itself. The knowledge of how the tools work, and how to manage and govern them, lies within the firms themselves, not governments. AI development to date has been very successfully governed by industry codes and standards, developed and self-regulated by the relevant firms, via entities such as the AI Collective. These industry-led activities inform quasi-formal practice codes endorsed by nongovernment bodies such as the National Institute of Standards and Technology, which would ultimately inform any formal regulatory processes in any event.
AI leaders should look within for solutions rather than kicking the regulatory can down the road to the politicians.