Would you like a short or a long answer about what AI really is?
We’ve come a long way in just half a century — from the moment Intel came up with the idea to place an entire digital processor on a single silicon chip, to today, when we have digital systems with artificial intelligence. These systems can think, talk to us, generate texts or entire projects, and even create works of art!
Isn’t such a description of artificial intelligence’s capabilities a bit exaggerated? The answer is unequivocal: yes. Not only is it exaggerated, but it’s also inaccurate, even tendentious. Yet many of us would swear that such a system gave accurate answers to our questions and created various original works right before our eyes! So, where is the truth?
It depends on what we consider to be the truth. If, after a magic trick, you were convinced that the magician really sawed a woman in half in a box (and then somehow put her back together and healed her), or that he really pulled a whole menagerie out of a hat and then swallowed ten ping-pong balls, then the AI system you communicated with might also seem omnipotent to you.
Is it really intelligence? Of course not. Neither algorithmic programs nor neural networks are capable of simulating the thinking process of the human brain. Experts estimate that the probability of achieving “human-level machine intelligence” within 50 years is 45%, and only 10% within the next nine years. Of course, there’s also the possibility that it may never happen.
And if we’re more interested in the structure than the function of the things around us, then we’re left with two answers — a short and a long one. The short one is easy to understand: an AI system uses some advanced programming systems to make us believe it understood the question, then uses the vast resources of the internet to generate an answer.
The long answer takes more time. Before we begin, let’s debunk the biggest misconception: artificial intelligence actually has nothing to do with human intelligence — it only imitates it. The system merely simulates that it understood the question (even though we would swear it did) and searches the internet to construct an answer. Let’s be fair and admit that even this is a great achievement.
A Bit of History
The term AI (Artificial Intelligence) was coined by John McCarthy during the 1950s, when he laid the first theoretical foundations for artificial intelligence. In the 1970s, at the beginning of what we call the computer revolution (excluding the relatively slow development period in the first 30 years from ENIAC to ALTAIR), we had a program called Eliza. It actually appeared in 1966 but was used only on business mainframe systems.
In 1977, versions became available for personal computers like the Apple II and TRS-80, and later for Commodore and Amiga. It’s worth noting that Eliza used the very limited resources of the 8-bit computers of the time — just 16 kilobytes of RAM, a good portion of which was taken up by the program itself. When we compare that to today’s gigabytes (one gigabyte is a million kilobytes) and access to the entire internet, it’s clear how much convenience today’s resources provide to program creators.
The system merely simulates that it understood the question (even though we would swear it did) and searches the internet to construct an answer. Let’s be fair and admit that even this is a great achievement.
Here’s what a typical dialogue between a user and Eliza looked like, translated into Serbian:
[Eliza] Let’s talk.
[User] Are you sure that’s a good idea?
[Eliza] Would you prefer if I wasn’t sure it’s a good idea?
[User] You’re clearly not sure; my doubts are confirmed.
[Eliza] Oh… clearly I’m not sure; your doubts are confirmed?
[User] Explain that to me, go on?
[Eliza] Maybe in better times, I will.
[User] Oh, but you already have.
[Eliza] Let’s talk about you, not me.
[User] That’s fair.
[Eliza] Tell me more about that…
We can see that Eliza used fairly transparent tricks to simulate understanding the user. The answer was often, “What do you think about that?”, “How would you explain your position on this?”, or “Let’s change the topic.” Today’s systems are incomparably more powerful and have access to massive repositories of global knowledge, making their responses far more convincing — but much of the principle remains the same.
Now for a Bit of Theory
This part may seem dry, but it’s key to understanding how artificial intelligence works. If it seems too complicated, feel free to skip the next section. But if you want to understand one of the crucial principles behind machine learning — the most important field in what we call artificial intelligence — here’s a chance to grasp a simplified model.
Neural Networks as Non-Algorithmic Systems
An algorithm is a set of instructions for solving a problem or completing a task. These instructions are often connected through mathematical (+, –, ×, ÷) or logical (or, and, not) operations. A key part is “conditional execution,” in the form “if (…condition…) then (…execute…)”. A good example of an algorithm is a cooking recipe or medication instructions.
Computer programs mostly consist of sets of algorithms written by a programmer as long strings of commands (or more bluntly, lines of code). A special program called a compiler, which is part of the programming toolkit, translates those commands into machine instruction sequences — very long strings of ones and zeros that the microprocessor reads from memory and executes one by one.
In certain specific fields such as artificial intelligence, non-algorithmic systems are also used, mostly in the form of neural networks. Interestingly, neural networks are modeled after the human brain — the brain’s neurons are simulated by nodes (in this case, artificial neurons).
It’s not hard to understand how neural networks work, especially if we use a simplified model. Imagine, for example, a series of input data (on the left side of a diagram) based on which the neural network needs to draw some conclusion or provide an answer.
Artificial intelligence will probably never become evil and selfish on its own, but it will certainly be an effective tool used by evil and selfish people.
It’s obvious that the “intelligence” of a neural network comes from its weight coefficients. How are these determined? At first, they are assigned randomly, which means the network is completely “dumb,” as its outputs are also random and therefore useless. That’s why the network undergoes a “training” process: the camera is shown various numbers and the network gives answers, many of which will be incorrect, though some will be correct by chance. For every correct answer, the coefficients are automatically increased a little, and for every incorrect one, they are reduced (toward zero, meaning negative ones will increase). After some time, we notice the network makes fewer and fewer mistakes, until it eventually becomes infallible.
Neural networks are the core of every artificial intelligence system. Their training can be a very long and arduous process, but the results are impressive. If you use ChatGPT, you’ll see thumbs-up and thumbs-down icons next to the text, allowing you to rate the response. After that, you can give detailed feedback on parts of the response. Every click on one of these icons and every detailed comment helps automatically train the network. Don’t be lazy—do this small favor for this excellent system!
If you used Google Translate about ten years ago, you probably remember how many mistakes it made in the beginning. Mistakes were common even after the traditional software was replaced by a deep learning system, but over time it slowly improved. Thanks to its flexible topology and well-tuned weight coefficients, today it is a well-“trained” system whose translation results are approaching those of a solid polyglot.
Is it intelligence?
Of course not. Neither algorithmic programs nor neural networks can simulate the thinking process of the human brain. We barely have a glimpse of that process, and it’s hard to say when we’ll know enough to begin simulating it. There are statistical estimates by experts suggesting a 45% chance that we’ll create “human-level machine intelligence” in 50 years, and only a 10% chance in nine years. Of course, there is also the possibility that it will never happen.
Until then, we can at least be aware of how far we’ve come in these efforts, and avoid falling under the influence of various misconceptions. It’s hard to say whether the term “artificial intelligence” was coined by marketers, journalists, writers, or all of them to some extent, but it’s certainly the result of a conscious mystification of modern technologies.
Artificial intelligence is a Chinese room. It acts all-knowing, but is actually like a zombie with no soul.
Some may disagree with the refusal to treat artificial intelligence as true intelligence, but this is really a matter of interpretation. At one time, there were popular philosophical discussions about the “philosophical zombie” (p-zombie), which aimed to determine whether a zombie has a soul. By all definitions, a zombie is a being without a soul, but if it behaves as if it has one, then it should be considered as having one—regardless of whether it actually does.
These discussions are now outdated and have been replaced by debates about the so-called “Chinese room,” a concept introduced by John Searle in 1980. This is a hypothetical room that is closed off and into which no one can see. Inside sits a person at a desk surrounded by compartments filled with papers containing Mandarin symbols. These symbols are unintelligible to those who do not know the language, and we (but not the visitors) know that the person inside doesn’t understand Mandarin either. The papers contain instructions: if the symbol is (1), the answer is (2); if it’s (3), the answer is (4), and so on. Thus, the person inside can give correct answers to any question without understanding any of them.
Now imagine visitors arriving with questions written on slips of paper, which they slide under the door. The person finds the matching question (symbol) in the archive and “copies” the appropriate answer, then returns the paper under the same door. Each visitor gets a response and thinks a genius is inside the room, though the person doesn’t understand any of the questions.
The analogy is clear: artificial intelligence is the Chinese room. It acts like it knows everything, but is actually like a soulless zombie.
Should we be afraid?
It’s clear that we don’t need to fear today’s artificial intelligence systems, because they will never be able to rule the world or start treating us as their slaves. They won’t even come up with such an idea—let alone want to implement it—because they are incapable of forming ideas or desiring anything, much less world domination.
But what if the zombie one day gains a soul—that is, if we manage to solve the thinking process and create true artificial intelligence? Some scientists believe we should fear that possibility, while others claim it’s just another mystification of the topic.
It’s clear we don’t need to fear today’s artificial intelligence systems, because they will never be able to rule the world or treat us as their slaves. They won’t even conceive such an idea, let alone want to act on it, because they are incapable of forming ideas or desires—especially not for world domination.
Each of us will have our own answer to this question, and many will be so confident in their judgment that it will be hard to persuade them otherwise. For example, Professor Stephen Hawking said that efforts to create thinking machines represent a threat to our existence. In a statement to the BBC, he said: “The development of full artificial intelligence could spell the end of the human race.”
Others are less pessimistic. Rollo Carpenter, creator of Cleverbot—a program that learns from its past conversations—achieved high scores in tests, fooling many people into believing they were chatting with a human.
Carpenter said: “I believe we will remain in control of the technology for a long time and that its potential to solve many of the world’s problems will soon be realized. We can’t know what will happen if machines surpass our intelligence. But I would bet that artificial intelligence will always remain a positive force.”
And maybe the real answer lies somewhere in the middle?
Maybe the pessimistic scenario should be left to science fiction writers, where it originally came from. Artificial intelligence will probably never become evil and selfish on its own, but it will certainly be an effective tool used by evil and selfish people. That’s why we should develop systems to defend against the misuse of artificial intelligence—not from AI itself, because it will always be only what we make it to be.
About the Author
