While thinking about AI, this question came into my mind. Could curiosity help in developing a true AI? According to this website (for testing creativity):

Curiosity in this context refers to persistent desire to learn and discover new things and ideas. A curious person

  • always looks for new and original ways of thinking,
  • likes to learn,
  • searches for alternative solutions even when traditional solutions are present and available,
  • enjoys reading books and watching documentaries,
  • wants to know how things work inside out.

Let's take Clarifai, an image/video classification startup, which can classify images and video with the best accuracy (according to them). If I understand correctly, they trained their deep learning system using millions of images with supervised learning. In the same algorithm, what would happen if we somehow added a "curiosity factor" when the AI has difficulty in classifying an image or its objects? It would ask a human for help, just like a curious child.

Curiosity makes a human being learn new things and also helps to generate new original ideas. Could the addition of curiosity change Clarifai into a true AI?

  • 3
    $\begingroup$ It could be possible to add a simulation of "curiosity" to an AI, but doing so will probably require some hard-coded rules that would themselves probably not be creative. And since there is "no right answer to this question", then it's probably not likely that curiosity is the missing factor to AI. For an example where curiosity can be added to an AI, see the Mario Lives! video, where someone built a video game AI that could become "curious". $\endgroup$ Commented Aug 11, 2016 at 14:08

5 Answers 5


Does this addition of curosity changes clarifai into a true AI?

As per my answer to this question, we don't know what the ingredients for a 'true AI' are. Via the Turing Test and its variants, the best we can do is "know one when we see one".

Curiosity certainly appears necessary for intelligence, though it doesn't seem sufficient - a lemming-like creature curious to see what's at the bottom of a steep cliff might not survive long enough to learn caution, even if it had the learning mechanisms to do so.

Here is some work by Schmidhuber on Artificial Curiousity.

Pierre-Yves Oudeyer has also done quite a lot of work on this and Active Learning/Intrinsic motivation.


when the AI has difficulty in classifying a image or its objects it should ask a human for help just like a curious child

It's called active learning, it's already used quite often.


It's a well known concept that's already used

What we call "curiosity" in humans and animals is in effect the chosen level of the "exploit vs explore" tradeoff for any active system. For example, the field of reinforcement learning is one approach that studies implementations of what essentially is the equivalent of curiosity; and we have research on how much curiosity is best e.g. multi-armed bandit concept.

So "using curiosity" is something that we already do as much as we can/should/are able to, but it would usually be called in some other, more specific term to specify the exact meaning instead of the vague word of "curiosity".


Curiosity by itself does not improve intelligence.

It increases the chances of better understanding a given subject, given that curiosity is coupled with actions in that direction.

For example: I am curious about how to make pancakes and decide to find a recipe but stop at the first instance of an answer with steps to follow.

Curiosity needs to be coupled with the desire to improve a given understanding and be followed by a review of current knowledge with the aim of updating the previousely reached conclusions. Provided that the used logic that judges improvements is correct.

Curiosity will not necessarily be benefitial for an improved intelligence. But to allow for an improved intelligence, curiosity is a mandatory requisite.

Curiosity is a symptom of an evolving intelligence.


Curiosity is used successfully with Random Network Distillation (RND). OpenAI has published a detailed article about their approach using this method, which was especially successful with previously unsolved games like Montezuma’s Revenge.

While this does not fully answer your question about curiosity being required to build a true AI, it shows that previously unsolved problems became solvable introducing curiosity in the reward system.


You must log in to answer this question.

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