How can one distinguish between an AI and a "sufficiently advanced algorithm"?

Question

Any sufficiently advanced algorithm is indistinguishable from AI.---Michael Paulukonis

According to What are the minimum requirements to call something AI?, there are certain requirements that a program must meet to be called AI.

However, according to that same question, the term AI has become a buzzword that tends to be associated with new technologies, and that certain algorithms may be classified in AI in one era and then dismissed as boring in another era once we understand how the technology works and be able to properly utilize it (example: voice recognition).

Humans are able to build complex algorithms that can engage in behaviors that are not easy to predict (due to emergent complexity). These "sufficiently advanced" algorithms could be mistaken for AI, partly because humans can also engage in behaviors that are not easy to predict. And since AI is a buzzword, humans may be tempted to engage in this self-delusion, in the hopes of taking advantage of the current AI hype.

Eventually, as humanity's understanding of their own "sufficiently advanced algorithms" increase, the temptation to call their algorithms AI diminishes. But this temporary period of mislabeling can still cause damage (in terms of resource misallocation and hype).

What can be done to distinguish a sufficiently advanced algorithm from AI? Is it even possible to do so? Is a sufficiently advanced algorithm, by its very nature, AI?

Answer 1

As you correctly pointed out, people tend to misinterpret the expression AI since they do not know what's behind an AI; it is pretty clear that in AI there is no more than just a bunch of algorithms and flowing bits. Talking about the nature of an AI without talking about the algorithmic paradigm of that AI is pointless and antiscientific.

This point of view is quite cynical; what we call intelligence is just the capability of solving particular problems.

The quote you cited in the title it's a derived version of the quote below, by a science fiction writer.

Any sufficiently advanced technology is indistinguishable from magic. - Arthur C. Clarke

Hence I doubt that a scientific answer exists since science lacks a formal definition of sufficiently advanced algorithm and advanced algorithm.

Answer 2

Intelligence is a quality of behavior, not implementation

Intelligence is a term that primarily applies to behaviors - people, animals or artificial systems can be called intelligent iff they exhibit intelligent behavior or decisions.

While there are many definitions of intelligence - here's a paper that studies 70 of them - it can be summarized to something like "Intelligence measures an agent’s ability to achieve goals in a wide range of environments." (S. Legg and M. Hutter).

Perhaps this definition is the answer to your implied question - while many algorithms can be very effective (often literally superhuman) in their own narrow domain, as of now they are very restricted in the range of environments where they exhibit this effectiveness. This means that they are not fully intelligent, they don't meet the definition/requirements of intelligence, and this also matches our intuitive expectations - we don't call AlphaGo superintelligent, because while it can beat humans in Go, the same humans can beat the same system on pretty much every other task.

However, a truly sufficiently advanced algorithm that can be effective at all or most varied tasks (e.g. a general artificial intelligence) can be reasonably called intelligent in the full meaning of the word.

Answer 3

• Easy: Parse the code! If the code meets your definition of Artificial Intelligence, that's what it is.

Only half kidding there. The question of what defines AI is constantly evolving and there is no exact consensus.

The most reductionist view could regard any decision making algorithm to be a form of intelligence, and the strength of the intelligence (it's level of advancement) is dependent on the frame of reference (typically a subjective entity making that evaluation.)

For this reason, it might be better to regard the issue from the standpoint of utility.

Utility is a central concept in AI because it is the means by which we evaluate "intelligence" (strength in a given task or problem aka "how well does it get the job done?") Does the algorithm fulfill the task(s) with sufficient optimality to meet our definition of Artificial Intelligence?

This leads me to believe the quote itself may be primarily a comment on the definitions of Artificial intelligence, per Lovecraft's answer. Subjectivity is indeed the a central notion of the Clark quote it references.

Answer 4

I am going to answer this questions by stepping away from the previously made insightful comments and academic answers. I am going to offer my opinion only. The problem is as I see it, a bit more complex than the previous answers. For example, why is it that the only measure of intelligence of an AI is when it can "beat" a man at a specific task?

Is not a dog, a bird or a dragonfly sufficiently intelligent and sentient? Yet these creatures as well as many others too, fail to achieve the intellectual challenges we want computers to make.

The mistakes we do often make is by trying to "impute" attributes and characteristics into an AI without it having to "work" for them. Those skills, experiences, memory and knowledge which all sentient beings have to constantly work at, refine and perfect.

You are correct though in your assertion that a lot of what is called AI are just academic or software gimmicks and simulations without offering the real qualities of either sentience or artificial intelligence.

However, I would challenge you by suggesting that your last question:

"What can be done to distinguish a sufficiently advanced algorithm from AI? Is it even possible to do so? Is a sufficiently advanced algorithm, by its very nature, AI?"

Is inherently flawed. Because have you forgotten how we too are free running, biologically self-programming, beings? Where our biological algorithms can often also be critically flawed? Therefore, let me ask you if the reverse is also true? Does an algorithm which seems to make mistakes, any less than one which does not?