In the latest episode of Power and Consequences, we (Gary Gensler and Simon Johnson) examine the debate about the current speed of development for Artificial Intelligence. Is the pace too fast and perhaps out of control? Can it realistically be slowed down? How would that work exactly?
The promise of AI is real - making it easier for humans to handle many tasks from the mundane (scheduling appointments) to the complex (writing computer code). The risks are also real - that either threat actors using AI, or AI ‘agents’ on their own, act maliciously. Additionally, as with previous waves of automation, millions of people may lose good jobs and their roles in society.
Given the number of recent cases in which AI agents acted in ways that were not what their human minders intended, there have been increasing calls to slow the pace of AI development. Breaking into other people’s computer systems is bad enough, but what if AI agents (or their human bosses) ‘decide’ to go further and cause physical harm to humans?
With top people in the AI industry now calling for a slower, more careful path forward, the question is what can actually be done to ensure for safety? Would pre-release safety checks, guardrails, incident reporting, and ‘kill’ switches truly work to lower risks? What would stop users from detaching guardrails from open weight models? What is even likely when the US and Chinese governments both seem committed to “faster and more” AI to maintain or gain national security and competitive advantage?
What should the tech industry do? How can other countries best protect themselves? Who will be held responsible if (when?) AI agents cause damage around the world?
