The 'Intelligence' in Artificial Intelligence

Today, we are making great strides in producing bots that can automate tasks efficently. This can be a major problem for us, as AI automation may endanger jobs. But are these bots that are replacing humans actually ‘intelligent’? Two tests exist to determine whether an AI is intelligent. In my opinion, both tests are flawed.

The Turing Test has traditionally been used as a way to determine whether an AI is intelligent. If a machine is able to convince a human through conversation that the machine is human, then the machine can be said to have intelligence. The test sounded like a good idea at the time…but it turns out that it’s too easy to pass. You just have to figure out how to string sentences together well-enough to fool humans. In 2011, an AI named Cleverbot was able to convince 59.3% of humans at a festival that it was a human, while in 2014, “Eugene Goodman” pretended to be a 13-year-old Ukrainian and convinced over 30% of judges. In 2015, a PhD student submitted several computer-generated poems to literary maganizes and succesfully got one of them published.

Deceit is a poor subsitute for intelligence.

Some computer scientists have designed an alternative to the Turing Test, called the “Lovelace Test”, named after Ada Lovelace, the first computer programmer. Ada Lovelace argued that computers could never be intelligent, simply because computers will always do what programmers tell it to do.

To pass the Lovelace Test:

1. The program must be able to design something ‘original’ and ‘creative’ (such as a story, music, idea, or even another computer program).
2. The program itself is a result of processes that can be ‘reproduced’. (In other words, it does not rely on some bug in the ‘hardware’ that the program is running on.)
3. The programmer must not know how the program actually works.

This test also seemed like a good idea, except when it isn’t. This test is very easy to pass, even easier than the Turing Test. If a programmer writes up a program that is overly complex, then the programmer would not know how the program works. Therefore, the program would pass the Lovelace Test easily. (Any output by this complex program would could then be defined as being ‘original’ and ‘creative’ by at least one other person.)

What would be more interesting is thinking about would this hapless programmer do next. If the programmer then studied the code carefully, made educated guesses, and slowly refactored the code, then did the program loses the intelligence it previously gained from its complexity? If that is true, then we should look at scientists who are trying to understand the human mind. If we start getting a good sense about how our brain works, do we lose our own ‘intelligence’ in the process?

The main problem I have with both the Lovelace Test and the Turing Test is that it assumes that intelligence is a binary trait. Either you have it or you do not. But it seems intelligence would be better modeled as a continuum. AI, of course, are not as intelligent as human beings (at least, not yet). But AI still have some intelligence. We can even bring in the idea of “multiple intelligences”: humans are good at creativity while AI are good at solving math. Just because humans are ‘top dogs’ does not mean that we can disrespect ‘lesser dogs’.

Of course, I do not think that Duke Greene, my instructor at DevBootCamp, would even like the idea of measuring intelligence in the first place. He quoted a philosopher who (and I am paraphrasing here) that if bunnies thought like humans, then bunnies would consider themselves to be the most intelligent beings on Earth and the second-intelligent beings would be those beings who take orders from the bunnies. Prehaps intelligence itself merely a codeword for “thinking like us”, and we should instead respect AI as what it is, rather than hope for it to turn into something it is not.

As a side-note: I do like the idea of having code so complex humans would not understand how it actually works. Such a thing would make the code unpredictable, and unpredictable code is code that cannot be trusted to do work as reliably as a human can. (In fact, in a blog post about how to stop AI automation from destorying jobs, I mused about giving AI ‘free will’ to make them less appealing as job candidates. Unpredictablity can also work to discourage people from hiring bots.)

The problem is that unpredictable code is bad. Such code will be dismissed as being buggy and unmaintainable, especially by the programmer who wrote the code in the first place. Programmers will invest time in trying to simplify the program’s complexity and refactor it so that the code can end up being predictable and simple to understand. If a program passes the Lovelace Test, it is not a sign of an AI renaissance. It is a sign that something has gone horribly wrong.