Are current AIs really intelligent?
The short answer is “not really”. At least, not in the way a normal person would understand it.
Computers are better at some things than humans. We know this and are not surprised if it takes no time at all for a computer to run through a whole load of complex calculations — after all, numbers are what computers are good at and we don’t consider that to be intelligent.
But computers understanding human language is a different thing, isn’t it? If a computer can understand the words, phrases and sentences that we use, surely they are displaying some sort of intelligence. I mean, if they can read my emails and throw away the spam they must be pretty clever, right?
No, sorry, not true. It’s all a trick. It isn’t that the computers actually understand the emails, it’s that the ingenious engineers and programmers have found a way of transforming a natural language problem into a numerical one.
It’s a case of classification — is an email spam or not? Humans can usually recognise spam when they see it; it’s an unsolicited email and it’s usually trying to get you to part with some money, or information. So how do the spam filters used by your email provider classify an email as spam?
What they do not do is read it and understand it in any meaningful way.
First of all, they have already got thousands of emails that have been classified as spam, or not spam. They break up these email into individual words and create a statistical model of a spam email based on the words it contains. Incoming email is similarly analysed and then compared to the model. The probability of it being spam is calculated — if the probability is above a certain threshold then the email is marked as spam.
Identifying words in a text is not difficult, because, to a computer program, letters and others characters are simply characters which don’t contain these codes.
So far from understanding what a spam email is, the spam filter software deals with strings of numbers (the words) and uses numerical methods to tag an email as spam, or not spam. This is not to denigrate the humble spam filter, it is ingenious, but it is also the sort of thing that computers have always been used for.
Another example is machine translation. While not perfect by any means, this has progressed enormously over the last few years. So this must be intelligent, surely.
When humans translate from one language to another, they understand the meaning of a text and then use their knowledge of another language to express that meaning in that language.
This is not the case with computers.
Computer translation is, these days, usually, Statistical Machine Translation (STM). While multi-lingual dictionaries can be used to look up words, as you might expect, in order to translate sentences and documents, STM relies on having a corpus of thousands (millions, maybe) of documents that are already translated. Translation is achieved by using statistics and probability to find a likely match for the text to be translated and to find the equivalent in the target language.
So computers are not intelligent in the way we are. They are super-fast calculating machines, as they have always been, except that they are now even faster and have more data to work with.
This is the first of a set of three articles that discuss whether current AI is actually intelligent, whether it is even possible for them to be really intelligent and whether Deep Learning provides a basis for real intelligence. You can find the others below.