No, this sixty-year old rule comes from the era when AI systems were hand-written code developed for each task. Critics argued that any intelligence demonstrated by an AI system was really human intelligence embodied in the code. For decades, under the "Hands off during training" rule, AI systems have been trained by automatice training procedures, such as stochastic gradient descent for neural networks. Now there are many demonstrations of AI systems doing specific tasks as well or better than humans. Narrow Artificial Intelligence or Artificial Narrow Intelligence has been proven.
Human General Intelligence is possible, although it could be improved. Therefore, Cooperative Human + AI >General must be possible if we can learn to cooperate. Is it still necessary to prove that Artificial General Intelligence is possible? Is it even desriable?
Might it be better to concentrate instead on learning to cooperate? If we don't learn how to cooperate with AI systems, how can we teach them to cooperate with us? We also could learn how to cooperate better with other humans.
In addition to all the advantages of cooperation in general, there are specific advantages in having humans and AI systems cooperate in the training of other AI systems. This website starts by presenting a set of ambitious new criteria for the definition of "Machine Intelligence". This website also presents many ideas for actively managing the training process for the AI system in order to achieve the goals set by these criteria. However, in many cases, this active management of the training process has a complexity beyond unaided human capacity. On the other hand, the criteria require human insight and will be very difficult to achieve by fully automatic systems alone. These criteria can best be met by Cooperative Human + AI systems.
The training of these Cooperative Human + AI system needs human insight even more than the systems themselves. The very definitions of Sensibility and Interpretability require human judgment. Therefore, the training for these Cooperative Human + AI systems should be Cooperative Human + AI Learning Management systems, that is, HAT-AI.
by James K Baker