On Google Scholar ("humble" and "AI") yields (About 175,000 results), while ("Humble AI") yields about 34 results.

It seems that the trait of "humbleness" and additionally the trait of "agreeableness" ought to be explored more in the contexts of AI. I am curious if others have heard lectures or seen papers specifically on the topic of humbleness or thought upon the importance of it being built into AI design.

This article delves into humility being a dimension of AI when it comes to validating outputs of AI systems. This paper about trustworthiness does a good job of discussing its necessity as a trait in AIAS and how specifically when talking of AI we need to be careful what definitions we are using when describing AI traits.

I envision "humbleness" will increase in importance in AI design. I am specifically looking for theoretical applications of the humbleness in AI design. Where you could envision it being important and why.

I'll answer my own question below for clarity. humble definition

  • 1
    $\begingroup$ Note that AIXI and AGI are not the same thing. AIXI is an AGI, but Godel machines e.g. could also be considered AGIs. So, you should use the tag agi for questions about AGI. $\endgroup$
    – nbro
    May 20, 2022 at 8:56

1 Answer 1


AI with black box NN type designs can tend to give erroneous output like in cases where specially designed images are tricked into making an AI say a gorilla is an apple...When this happens we could label the AI as "unpredictable" "dysfunctional" or "untrustworthy" if we wanted to mildly anthropomorphize the AI. All "untrustworthy" AI like this need to be "humble" we need to realize their limitations and in future projects where AI interact with humans this is especially the case.

Let's use an extreme example to illustrate. Let's say an AI built to defend people from creatures on an alien world is designed. I would say an AI like this should be designed "humble". Ideally it would have a failsafe to listen for human commands to stop incase it were tricked. This would have the drawback of having the failsafe being prone to abuse but would give people peace of mind knowing if something goes wrong it can be stopped.

In the case of distributed AI systems, it may make sense to create a "humbleness" hierarchy. The most influential/powerful AI in the system have the most resources or perhaps the highest accuracy with their test. If a distributed AI system is doing a diverse array of tasks concurrently, it will occasionally make sense to end groups of tasks deemed less essential. This may look similar to downregulation of certain parts of the brain during moments of stress and upregulation of others.

During high traffic moments, an NLP AI may need to spend less time thinking of the perfect response and more time thinking of a good enough response it keeping with the stress of the situation. If a doctor is in a triage situation it is inappropriate to be perfectly courteous and kind to one patient and let several others die. It is necessary for the essential life saving part of the network to become dominant while the bedside manner is forgotten for the moment.

AI and especially AIXI when released into the world should have humbleness baked in several respects. First the designer of such AI should be very humble creating them acknowledging the tasks they have designed the AI to complete and the reward system implicit in its design is not going to be perfect (not causing any problems). AI released into the world should be humble in that it responds to environmental cues with respect and dignity for the life present. I appreciate this question on general AI seeking to minimize any negative effects on the environment.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .