Typical Definition of Intelligence: the ability to acquire and apply knowledge and skills.
Wikipedia Definition: the ability to perceive information, and to retain it as knowledge to be applied towards adaptive behaviors within an environment or context.
Proposed Definition: A measure of the ability to make comparisons
To define intelligence what typically happens is that we start with intelligent behavior, which essentially just means acting like a human. Animals that act more like us are more intelligent, and things that act less like us are less intelligent. People who can solve the kinds of problems humans can solve, but can solve them faster or can solve more difficult versions are deemed to be more intelligent. Then, working backwards, we assume there's some quality of 'intelligence' that causes or enables this kind of behavior.
One of the biggest problems with this very vague definition of intelligence is that it's hard to nail down. Is someone who can solve one kind of problem, but not another, intelligent? What if they can solve difficult problems, but can't explain how. Then there's the problem with the definition of Artificial Intelligence where any breakthrough is seen as not being "real intelligence" anymore. Part of the problem is that we seem to judge intelligence in different ways:
- Threshold - intelligence means having a certain amount of some quality or ability, there isn't intelligent until it crosses some threshold. Something that learns very slowly and makes lots of mistakes isn't intelligent, but something that learns quickly and accurately is. Intelligence by this definition would be an emergent property, so ants might not be intelligent but an ant colony could be.
- Recipe - intelligence requires a particular combination of skills. For example someone might posit that if an animal had the ability to learn, use language and had intuition it would be intelligent. Often these lists will include things like "logical thinking" or "rational thought" which are basically just synonyms for intelligence.
- Process - there's some particular ability or process that's necessary and maybe sufficient for intelligence. For example people have suggested that consciousness might be required to be truly intelligent. Often intelligence will be described as being learning or memory or some other mental process.
Of the three the last definitely seems like it would be the most useful. The other two boil down to more complex ways to compare behavior to stereotypical human behavior, and therefore miss the chance to identify an unbiased quality of intelligence. If in the future we were to make a machine that passes the Turing test, or discover life on another planet, the last kind of definition would let us clearly say if the subject was intelligent or not. It's also the definition that makes measuring intelligence along a spectrum the easiest, and it avoids the problem of being confused by confounding factors. For example if someone lost their vision we wouldn't want to say they're less intelligent, and it seems like changes in behavior could easily throw off the other two kinds of definitions. Even if someone was completely paralyzed if they still had the ability to process information intelligently we'd want to say they're intelligent, I think the last kind of definition has the best chances of hitting that goal.
The process that I think best captures and defines intelligence in general best is the ability to make comparisons. I suspect what we call intelligence in humans is basically just "guess and check" - we come up with predictions or actions and then try to rule them out. We're not very good at making an accurate prediction without an iterative process of checking it against some experience or rule. But of those two abilities ("guess" and "checking") it seems like only the checking is really required for the appearance of intelligence. I could imagine a machine that only comes up with guesses, it just makes random predictions, and without anyway to gauge if they're good predictions or not it wouldn't seem intelligent. But if we had a machine that just made comparisons, it took in guesses and told us which one matched some standard or rule or experience better, that would seem useful and probably be accepted as intelligent.
In this view intelligence isn't an overarching framework that describes all behavior, instead it's a single kind of ability that works with the other abilities and processes humans are capable of to make intelligent behavior possible. Let's look at an example from AI first. A computer program that plays tic-tac-toe doesn't seem intelligent, or if it is, it's the most basic kind of intelligence. The program could be as simple as instructions for what to play next in every possible board position. In comparison a program that plays chess seems like it's much more intelligent, and that's likely because it has to make a lot of comparisons. Because of the huge number of possibilities, each potential move has to be compared to each other, including the string of likely moves that will follow them. A Go playing AI is judged to be even more intelligent and a large reason why is that evaluating the strength of a board position is very hard, just comparing two positions and picking which one is better is often an incredibly hard comparison to make.
In this view intelligence isn't an overarching framework that describes all behavior, instead it's a single kind of ability that works with the other abilities and processes humans are capable of to make intelligent behavior possible. Let's look at an example from AI first. A computer program that plays tic-tac-toe doesn't seem intelligent, or if it is, it's the most basic kind of intelligence. The program could be as simple as instructions for what to play next in every possible board position. In comparison a program that plays chess seems like it's much more intelligent, and that's likely because it has to make a lot of comparisons. Because of the huge number of possibilities, each potential move has to be compared to each other, including the string of likely moves that will follow them. A Go playing AI is judged to be even more intelligent and a large reason why is that evaluating the strength of a board position is very hard, just comparing two positions and picking which one is better is often an incredibly hard comparison to make.
A comparison isn't enough to enable complex behavior, it requires inputs from other processes, here's a quick list of what might be included in human mental capabilities: learning, memory, emotions, consciousness, instinct, imagination and prediction. I think we can come up with reasonable definitions for all of these, and don't think that any by themselves would constitute 'intelligence'. We can imagine building a machine to try and replicate human behavior by replicating these capabilities or modeling them, but I don't think that even if a machine had all of these that it would act intelligent if it lacked the ability to make comparisons between the different kinds of information. There's other ways to make these processes work together, a person or machine's behavior could just be the average of all the different processes, or there could be a hierarchy with higher up processes kicking in and overriding the lower ones. However I don't think that would give us the kind of behavior we experience, but if we allow these different processes to compare outputs in a useful way, then complex intelligent behavior seems possible.
Here's an example of the kinds of complex comparisons humans are capable of:
Now at first glance this seems ridiculous, it seems like two things that just can't be compared. Half is asking for a subjective auditory judgement of silent image, and the other half is a random image being described by a made up adjective. This is the most arbitrary comparison I could think of and yet, I suspect that most people could give a consistent answer to this question. Also, I believe that those answers would change in predictable ways as aspects of the two parts changed - how would your opinion change if it was a lion or you had to judge how 'maluma' that shape was? Humans are capable of making incredibly complex comparisons involving a variety of different kinds of data and experiences and make them quickly in a consistent way. I suspect that making comparisons at this level of complexity is something that only we can currently do, and while this comparison is obviously pointless we can think of a lot of realistic and complex comparisons that are very useful. Early humans undoubtedly had to compare the benefits of making shelter or searching for food, and modern humans might have to compare the choices of investing in a startup or remodeling a kitchen. The scientific method is based around making predictions and comparing them to experimental results. I don't want to imply that there's a single "comparison making" part of the brain, but more that being able to make comparisons are a fundamental part of the architecture of our brains, and likely would be integral to any intelligent machine.
An interesting observation about human's ability to make comparisons is that we're actually not very good at quickly making a lot of very simple comparisons. For example if I have to judge which of two similar objects is heavier I'll pick them both up and might not be able to immediately tell. It might require going back and forth a couple times, and even if I held one in each hand it might not be obvious. This is a very simple comparison that simple machines can do instantly, but that can be somewhat difficult for us. There's also lots of optical illusions that show that we can't always make simple comparisons based on appearance accurately either. These kinds of examples make me think that humans are good at making complex comparisons, but not necessarily good at quickly and consistently making simple ones.
The complex comparisons are the kinds of abilities that we're trying to build towards with AI, which essentially is just taking novel inputs and comparing them to previous data to try and assign a value to new data. Whether that's making a translation or image recognition or game playing, the key ability that limits how intelligent AI seems is its ability to make complex comparisons. We could even hypothesize that a machine that could only make comparisons between two sets of data would be intelligent even though it couldn't do anything else, it would be a very weird kind of behavior, but I think we'd recognize that ability that humans are so good at.
An interesting observation about human's ability to make comparisons is that we're actually not very good at quickly making a lot of very simple comparisons. For example if I have to judge which of two similar objects is heavier I'll pick them both up and might not be able to immediately tell. It might require going back and forth a couple times, and even if I held one in each hand it might not be obvious. This is a very simple comparison that simple machines can do instantly, but that can be somewhat difficult for us. There's also lots of optical illusions that show that we can't always make simple comparisons based on appearance accurately either. These kinds of examples make me think that humans are good at making complex comparisons, but not necessarily good at quickly and consistently making simple ones.
The complex comparisons are the kinds of abilities that we're trying to build towards with AI, which essentially is just taking novel inputs and comparing them to previous data to try and assign a value to new data. Whether that's making a translation or image recognition or game playing, the key ability that limits how intelligent AI seems is its ability to make complex comparisons. We could even hypothesize that a machine that could only make comparisons between two sets of data would be intelligent even though it couldn't do anything else, it would be a very weird kind of behavior, but I think we'd recognize that ability that humans are so good at.
While it seems like it might be too simple or fundamental an idea to rest the definition of intelligence on, comparisons seem to work. If an animal, human, machine or even alien were capable of making comparisons between different ways of processing data I think we'd definitely want to call it intelligent. If it couldn't make those kinds of comparisons it likely wouldn't seem intelligent. Comparisons are necessary, I'd argue they could even be sufficient for intelligence. That definition allows us to measure intelligence across vastly different processes by looking at how quickly, accurately and robustly they can make complex comparisons.