If a robot only acted with efficiency it would solve the same problems the same way every time unless it got new information. Even then, it can’t re-evaluate all of it’s layered processes any time it gets new input to determine if they need an update. It would only change it’s behavior once it did that action again, and only if it was able to relate the new information to the old process.

This would be incredibly inefficient to do in all use cases. Even new related information could have unintended consequences that would cause a re-evaluation of many processes. The design would need to weight incoming information. It would need to only consider new information that was both strongly correlated with the action being taken, and within a sufficient degree of difference from previously obtained information, before being considered.

The interesting implication of this is that we evolved to be random with our actions when we have a lack of information. This is the classic “crawling around grabbing aimlessly for your fallen glasses,” scene. The most efficient way to search would be to move in a sided to side search grid. We don’t do that robotic like action, because we have a counter force. Robots can do the most efficient thing because no one is trying to stop them.

Humans had to evolve with other intelligent creatures watching them. This “counter intelligence” developed our own evolutionary need for intelligence. If every time a human fell down, they always got back up the same way, then a creature could use that pattern to it’s advantage. To survive, we had to be unpredictable all the way down to the split second decision point of the subconscious mind. Or to the point that we will blindly feel around in the dark.

AI could possibly greatly benefit in accuracy, and emergent problem solving, with the use of counter intelligence. Essentially this would involve another AI that instead of trying to improve the results of the first, it tries to stop the first or to eliminate low weight results. This resistance factor could be an extra valve to control in user facing generation.

It would also show the AI in real time the weak points in it’s generation and it could focus more compute in those areas. Theoretically, this would lead to more accurate fingers, more detail, better text in image, and better general reasoning.