“Moving the Finish Line: The Ever-Changing Definition of AI Intelligence”

AI might become smarter than us!” “AI is already smarter than us!”

Today’s media is inundated with claims about AI being smarter than humans or that if it is not yet, it will be soon. It is an odd claim because it depends on, among other things, the very definition of “smart” and “intelligent.” In all our human-centered arrogance, we are changing the definition continuously.

Our human mind has been obsessed with artificial minds for centuries, from Homer’s Iliad to Mary Shelley’s Frankenstein to C-3PO in “Star Wars.” The question of when these artificial minds become intelligent has fascinated us and scared us at the same time. But when do these artificial minds become intelligent?

Imagine running a marathon. You aim to run the 26.2 miles, but the moment you have the finish line in sight, somebody tells you that the organization has just decided to extend the marathon by a few miles. Somewhat annoyed (and exhausted!), you run farther, when suddenly in the distance you see the new finish line. Exhilarated you run farther, but then you are told that the finish line has been moved yet again.

Sounds unfair? Well, in answering the question of whether machines are smarter than humans we do exactly that: We continuously move the finish line.

Since the formal inception of artificial intelligence in 1956, it was held that AI could be considered to match human intelligence when it would be able to play chess. In 1996, IBM’s Deep Blue beat chess grandmaster Garry Kasparov. AI had reached the finish line of intelligence!

But immediately the world argued that while a computer beating a chess master is certainly impressive, this achievement was in hindsight doable. After all, the number of possible chess game variations is large but remains only around 10^120 (a 1 followed by 120 zeros). Handling that many possible chess game variations does not really constitute intelligence. What really does is beating somebody at the game of Go. Go is played on a much larger board than chess, typically 19×19, compared to chess’s 8×8, with simpler rules and far more possible board positions than chess—around 10^170. The complexity of Go would really constitute the finish line of intelligence.

In March 2016, Google-DeepMind’s AphaGo beat professional Go player Lee Sedol in a five-game match. AlphaGo had crossed the finish line of intelligence. After 60 years AI had truly become intelligent.

Now, beating human chess and Go champions is impressive, but 10^120 or 10^170 variations are still manageable. If a computer is powerful enough to compute, it is feasible it can capture these variations. So, perhaps computers are not really intelligent when they beat chess and Go champions. Instead, what would make AI really intelligent is not just making computations based on some moves, but actually interacting with the real world. Self-driving cars! That would constitute intelligence.

When the first self-driving cars hit the highway, we decided to move the finish line yet again. Sure, self-driving cars were intelligent in their own right, but highways did not make them really intelligent. Once they hit the streets in crowded cities, then they would be intelligent.

No, creativity—that would make AI intelligent. If AI could create text or pictures that did not exist, that would really constitute intelligence. Meanwhile, although many of us now rely on AI to generate texts, pictures, and movies, we are still wondering whether AI is smart.

The question is, why do we move the finish line? If an organization moves the finish line during a marathon, that hypothetical organization would simply be mean and heartless. In contrast, we humans likely move the intelligence finish line for AI because of a simple reason: We don’t want to lose the game and hence do not play by the rules.

We don’t play by the rules. Not only because we continuously move the finish line. We are not even true to ourselves.

Undoubtedly, you consider yourself smart—certainly smarter than a machine. Yet, you probably cannot beat a chess grandmaster or a high-level Go player. And even though you can drive a car, you will not be able to do that for long when a self-driving car is better able to do this.

By these previous definitions of intelligence, you are not as smart as AI. But you will quickly come up with a clever solution: You may not be as smart as AI when it comes to chess, Go, or the self-driving car, but at least you are conscious. And you have a soul. And a body. And have five senses.

Will AI ever become as smart as humans? I am reluctant to answer this question with “yes.” Not because AI can never become as smart as humans, but because humans will always eagerly move the finish line called intelligence.