3. Principle of controllability
Developers should pay attention to the controllability of AI systems. [Comment] In order to assess the risks related to the controllability of AI systems, it is encouraged that developers make efforts to conduct verification and validation in advance. One of the conceivable methods of risk assessment is to conduct experiments in a closed space such as in a laboratory or a sandbox in which security is ensured, at a stage before the practical application in society. In addition, in order to ensure the controllability of AI systems, it is encouraged that developers pay attention to whether the supervision (such as monitoring or warnings) and countermeasures (such as system shutdown, cut off from networks, or repairs) by humans or other trustworthy AI systems are effective, to the extent possible in light of the characteristics of the technologies to be adopted. [Note] Verification and validation are methods for evaluating and controlling risks in advance. Generally, the former is used for confirming formal consistency, while the latter is used for confirming substantial validity. (See, e.g., The Future of Life Institute (FLI), Research Priorities for Robust and Beneficial Artificial Intelligence (2015)). [Note] Examples of what to see in the risk assessment are risks of reward hacking in which AI systems formally achieve the goals assigned but substantially do not meet the developer's intents, and risks that AI systems work in ways that the developers have not intended due to the changes of their outputs and programs in the process of the utilization with their learning, etc. For reward hacking, see, e.g., Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman & Dan Mané, Concrete Problems in AI Safety, arXiv: 1606.06565 [cs.AI] (2016).
Published by Ministry of Internal Affairs and Communications (MIC), the Government of Japan in AI R&D Principles, Jul 28, 2017