Alan Turing believed psychic mind-readers could undermine the Turing Test! Yes, you read that right; and no, I’m not joking. He spends about a page of “Computing Machinery and Intelligence” (CMI) discussing this problem. CMI is the paper that first introduced the Turing Test; it was published in the journal Mind in 1950.
What’s really interesting about CMI are its insights into artificial intelligence. Many of them get overlooked by modern surveys, and I want to look at three of them here. First, Turing offers several tests related to the classic Turing Test that have been largely ignored. Second, we’ll look at Turing’s arguments for and against strong AI. (This is where the psychics come in.) Finally, we’ll look at some of the predictions CMI made about the development of artificial intelligence and evaluate how much progress we’ve really made.
Testing for AI
What we now call the Turing Test, Turing called the “imitation game.” But first he considers a different game. Instead of deciding which opponent is a man and which is a computer, the interrogator must decide which is a man and which is a woman.
For example, we can easily imagine a man who could convince the interrogator that he is a woman (or vice-versa). He may be a good actor, but he is nonetheless still a man. Gender and intelligence are both fundamental aspects of being a human, so why should intelligence be purely functional whereas gender is not?
There is a lot room room to explore how non-traditional LGBT genders might broaden this argument, but I think it would just complicate the question without offering any new insight. Conversely, this style of reasoning may shed some light on the philosophy of non-traditional sexuality. Maybe in another post.
The other test that Turing proposes he calls “viva voce.” In this test, the interrogator asks a machine questions and determines if the machine’s answers exhibit a quality of understanding.
The main advantage of this test is that it should be more “accurate” because there are fewer variables. Christian’s book The Most Human Human provides a behind the scenes account of performing the Turing test. In it he confirms that not all humans are equally capable of convincing the interrogators that they are human, and this leads to high variability in the results. Furthermore, a human’s ability to perform well is not constant. We can be affected by mood, how recently we’ve eaten or slept, and infinitely many other factors.
This test has one main drawback: it is hard to quantify. The interrogator can make a subjective decision that the computer exhibits strong AI (or fails to, or is getting closer, etc.), but he cannot explicitly determine that the machine is 10% complete. This is in major contrast to the standard Turing Test, where we evaluate a machine’s performance by saying it was able to fool the interrogator X% of the time. This difficulty in quantification makes it difficult to determine if progress is being made. In fact, if there exists no algorithm to determine if progress is made, it is probably the case that strong AI cannot exist.
I think the positives outway the negatives and will personally try to apply this test when performing thought experiments on AI.
Turing’s arguments and counterarguments for strong AI
What makes this section of CMI so interesting is that the modern reader will never have heard some of these arguments before.
The first is politically incorrect. Many people disbelieve in strong AI because the idea just seems so weird – Would computers develop emotions? Could humans and computers become friends? Will humans become obsolete?
Turing makes a great attempt to explain away this weirdness with a racial analogy. According to Turing, we humans will never be able to understand what emotions might mean to an artificial intelligence, but who cares? A man can never understand a woman’s emotions; a Caucasian can never understand an African’s; and a man can never understand a dog’s. Nonetheless, we are all able to relate to each other. This argument becomes even more poignant when we consider the racial segregation of the 1950s when the paper was written. At that time, you actually could find people who would argue that non-white intelligence was qualitatively inferior.
Turing spends several pages disputing the claim that intelligence is somehow a property of electricity. To us, the claim seems absurd; but it is understandable that people of the time might have believed it to be the case. The brain’s neurons send electric signals to each other just like a digital computer’s transistors. You still hear this argument occasionally on the internet, but never in more educated circles.
Now for the psychics! Turing considered the implications of paranormal activity on the Turing Test. He concludes (without irony) that “this argument to my mind is quite a strong one.” He concludes this section by claiming “to put the competitors into a ‘telepathy-proof room’ would satisfy the requirements [of the Turing Test].” It’s really hard to say anything more without sounding sensational. Go read it for yourself.
Turing’s claims made me curious about the history of scientific opinion on telepathy. According to a 1991 study, about 4% of members of the National Academy of Science believe in ESP. Even this low number is much higher than I would have guessed. Although there is no hard data on the consensus of the 1950s, it is clear that belief in ESP was much more widespread.
I think there are two important lessons to be learned from this. First, scientific consensus can shift quite rapidly and quickly forget any potentially embarassing history. Second, we must be bold when we publish science. History has clearly forgiven Turing for this “mistake.” His willingness to confront psychics may have even enhanced his reputation and added to his popularity.
How accurate were Turing’s predictions?
Not very. Turing was quite the optimist, but the progress of AI research has slowed considerably since Turing’s day. That Turing’s predictions were wrong do not reflect negatively on him, but rather reflects the optimism of his generation of AI researchers.
Here’s a short list:
- Turing believed that “the problem [of artificial intelligence] is mainly one of programming.” He suggests that all we have to do is find the right (very complicated) algorithm, and then BAM! AI appears. But modern research suggests this is not the case. Data mining and machine learning techniques have shown that large quantities of data are far more important than complicated algorithms. In On Intelligence, Hawkins suggests one of the simplest algorithms for the brain imaginable. He proposes that pattern recognition is the fundamental algorithm of the brain, and that every piece of the brain performs it in the same way.
- “Estimates of the storage capacity of the brain vary from 10^10 to 10^15 binary digits. I incline to the lower values and believe that only a very small fraction is used for the higher types of thinking.” That’s about how many neurons the brain has, without even considering the number of dendritic connections or the potential for a connection to encode multiple bits. After decades of research, we can still only establish a lower bound on the number of bits our brains might use, and that number is still very very high.
- Turing believed vision to be one of the easiest problems in AI. The problem has proven so difficult, however, that it has become its own subfield of computer science. It is rarely even considered by someone who studies “AI.” In On Intelligence, Hawkins argues that our brain’s algorithm for vision is exactly the same as its algorithm for higher order thought. It’s all pattern matching. All that matters is what data gets fed into it.
At least we can still agree with Turing’s conclusion: “We can only see a short distance ahead, but we can see plenty there that needs to be done.”
CMI is full of even more little gems that I haven’t mentioned, and is very readable — even without a computer science background. If you’ve made it this far, you’ll probably enjoy reading the actual text.