Objections From An AI Skeptic

 

After subjecting myself to numerous articles on the topic of Artificial Intelligence (AI), on its seemingly unlimited potential and unnerving capacity to bring about a dystopian future for all mankind, I thought I’d take a moment to consider AI’s most vexing limitation: the fact that complex problems only rarely have direct, simple solutions, but direct and simple are the only ways that AI can actually “learn.” This is not to say that AI can’t be used to discover solutions that hadn’t been previously considered, or that humans adapting an AI-generated solution can’t realize disastrous ends – not at all. I’m just saying that the popular view of AI’s “learning” technique may be imparting to it a level of sophisticated solution-providing that it simply doesn’t have, and likely can’t attain.

Consider that, at its very root, AI can only “learn” via trial-and-error. As an example, how would AI specifically arrive at a solution to, say, discovering which single-digit whole numbers add to ten? That algorithm would have to be limited to the numbers one through nine, calculate each of the possibilities, and then store the successful calculations. The pseudo-code would look something like this:

DO  UNTIL ALL OF THE SINGLE DIGIT WHOLE NUMBERS HAVE BEEN ADDED

DO UNTIL THE RESULT IS 10

ADD 1 PLUS 1

IS THE RESULT 10?

YES: STORE THE COMPONENTS

NO:

ADD 2 PLUS 1

IS THE RESULT 10?

YES: STORE THE COMPONENTS

NO:

ADD 3 PLUS 1

(These four lines incrementally repeat until the numbers 1 – 9 and 9 – 1 have been added together.)

END DO

END DO

TURN THE WORLD INTO A DYSTOPIAN NIGHTMARE

(Okay, that last line has nothing to do with ascertaining a solution to the example problem. It was a joke – though no computer would recognize it as such.)

Then, when the AI researcher retrieves the results, he will find that the stored components are 1 + 9, 2 + 8, 3 + 7, 4 + 6, 5 + 5, 6 + 4, etc. Now, compare this whole process to how a third-grader would attack the same problem, and you can begin to see how more complex or layered problems would be far more difficult to solve using only trial-and-error. Of course, even the most basic computers could execute the trial-and-error algorithm very, very quickly, but the problems that present themselves in Management Science space tend to be far more complicated than the example above – otherwise, we PM-types would find ourselves easily replaced by this nascent AI technology.

Note also that the AI researcher would have had to set up the algorithm with the necessary parameters. This is key to the whole AI-creating-dystopia narrative, where the various computers that had been created in order to address some major problem in real-time, like law enforcement or strategic nuclear arms usage, come up with a solution that never would have been selected by responsible executives or high-level decision-makers, but is, nevertheless, implemented before any actual person can slam the brakes on it. In short, the optimal strategies for major issues, like law enforcement or strategic nuclear arms usage, are so complex as to not be discoverable exclusively through trial and error. Past examples can inform the search for the optimal solution in these instances, including past failures, but they can’t serve as the only method for ascertaining such strategies, tactics, and decisions.

Another way of highlighting AI’s complexity problem would be to consider how the above pseudo-code would be modified if the problem moved from “discover each of the single-digit additive combinations result in 10” to “why do you want to know which single-digit combinations result in 10?” (which is, ironically, something that our comparison point third-grader may well ask prior to attacking the problem in the first place). Indeed, AI is likely to be comparatively helpless when enlisted to answer any question that begins with “why.” Why? (snicker) Because causality doesn’t lend itself to discovery via trial-and-error, unless the alternatives are both (1) identifiable and (2) quantifiable. Yes, we all know that the Titanic sank because it hit an iceberg, but that’s the simple answer – we do not need advanced AI to tell us so. However, if one wishes to consider more nuanced causal factors, such as the speed of the vessel, its rudder’s relative size, the alertness of the lookouts, the lack of watertight caps to the watertight doors, the unavailability of binoculars for the lookouts, and dozens of other factors, simply reading history books would be the way to go. Computers can already perform document searches, so AI doesn’t bring anything to the table there.

One more little tidbit – in the above paragraph, I had originally typed “…that the Titanic sand because…”, and the MSWord Review function didn’t find that odd. My advice, then, would be to tread carefully when tapping AI’s assistance in selecting a solution for an even remotely complex problem. You wouldn’t want the Titanic to sand, would you?