Use AI to watch AI
In addition to the four forces of regulation identified by Lessig—law, norms, architecture and markets, we have a new tool at our disposal to help us regulate AI. And that is AI itself.
Recently, sociologist Amitai Etzioni and his son, computer scientist Oren Etzioni, suggested a new class of algorithms, called oversight programs. These programs “monitor, audit, and hold operational AI programs accountable.”
Using an AI to police other AIs offers many advantages. First, it allows us to scale our ability to monitor AI systems substantially. Installing humans observers to monitor AI behavior could easily become impractical and prohibitively expensive. When we use AI to do monitoring and audit of other AIs, humans can focus on more novel, insidious machine misbehavior.
Second, an oversight program can monitor AI algorithms at much higher speeds than humans can. This may be necessary for certain domains, such as high-frequency algorithmic trading.
Third, oversight programs allow us to audit AI algorithms while respecting the privacy of those affected by them. For example, suppose we want to understand whether online advertising algorithms use sensitive user data to target ads, or whether an AI is scanning the personal contents and attachments of our private communications to infer our intimate traits. It may be possible to allow independent auditing of such systems without perpetuating the same privacy violation by accessing people’s private information.
But there are limits to using oversight programs. When we use one AI oversight program to audit another AI, we must have confidence in the oversight program. And to acquire such confidence, we may use a third AI to audit the oversight program itself. But the problem continues ad infinitum. At some point, we run out of computing power. As shown in my own research, led by Manuel Alfonseca, we cannot in principle verify that a superintelligence is safe. So at some level, it has to be just a matter of faith.
References
Etzioni, A. & Etzioni, O. AI assisted ethics. Ethics Inf. Technol. (2016)
Aldridge, I. High-Frequency Trading: A Practical Guide to Algorithmic Strategies and Trading Systems. (John Wiley & Sons, 2013).
Cabañas, J. G., Cuevas, Á., Arrate, A. & Cuevas, R. Does Facebook use sensitive data for advertising purposes? Commun. ACM 64, 62–69 (2020).
Winder, D. Google Confirms New AI Tool Scans 300 Billion Gmail Attachments Every Week. Forbes (2020).
Alfonseca, M. et al. Superintelligence Cannot be Contained: Lessons from Computability Theory. Journal of Artificial Intelligence Research vol. 70 65–76 (2021).