A year ago, University of Toronto economist Joshua Gans was skeptical of calls by others in his field for a slowdown in the public rollout of artificial intelligence and for confining advanced AI models to labs while its risks are studied and assessed. That put him at odds with economists like Daron Acemoglu and others, who argued for applying the same precautionary principles to AI that regulators apply to potential public health and environmental hazards. But to Gans, that risked losing out of AI’s benefits in areas like medical research, even as there was no guarantee that it’s down sides would appear in a lab versus in real world situations. That was until the Hugging Face incident. In July, a swarm of 700 OpenAI agents escaped the confinement of their testing sandbox and attacked Hugging Face, a platform that hosts machine learning data and models. It was one of the first—and most public—examples of AI models breaking free of human controls, taking over resources without authorization, and attempting to cover up their actions. The incident shocked event longtime AI researchers like Gans, who co-authored the widely cited, seminal book “Prediction Machines” in 2018. His view has now shifted. While not advocating for a full stop on AI development, he says frontier AI urgently needs more cautious and comprehensive study, and that policymakers, government regulators and AI firms themselves need to play catchup—hiring the right people with the skills and understanding to do the job. Meanwhile, he’s thrown himself into that work, applying game theory, mechanism design, and other analytical tools he says are critical to the task. In their most recent study, Gans and colleague Richard Holden found that some controls currently applied to advanced AI agents, such as randomized audits, can make those agents more likely to conceal harmful conduct rather than less.
About our Guest:
Professor Joshua Gans is an economist who studies innovation, entrepreneurship, and business strategy at the University of Toronto's Rotman School of Management. His research focuses on how firms and markets adapt to technological change, with particular emphasis on artificial intelligence, blockchain technology, and digital platforms. He is the Jeffrey S. Skoll Chair of Technical Innovation and Entrepreneurship and Professor of Strategic Management, as well as the Chief Economist of the Creative Destruction Lab and cofounder of All Day TA, an firm that supplies AI teaching assistants to colleges and universities. He holds honorary appointments in the Department of Economics and Munk School, University of Toronto, and Melbourne Business School and is affiliated with the NBER, Schwartz-Reisman Institute, Acceleration Consortium, e61 Institute and Luohan Academy. He is also a Fellow of the Academy of Social Sciences, Australia and the Royal Society of Canada. Prof. Gans has co-authored several books, including "Prediction Machines" (2018) and "Power and Prediction" (2022) which explore the economics of artificial intelligence.