In this blog post, we will look at how deep learning-based artificial intelligence is catching up with human learning and social skills, and consider the boundaries of humanity.
British genius mathematician Alan Turing asked the question, “Can machines think?” and proposed the imitation game as a criterion for determining whether machines can think. In this game, a judge converses with a human and a machine, and when the judge determines that one of the two conversation partners is human, the machine is considered to have passed the Turing test. This imitation game is what is now known as the Turing test. Brian Christian, author of The Most Human Human, participated in the Loebner Prize contest in 2009, which was a competition based on the Turing test. There, he identified many factors that make humans more human than AI in order to win against AI. He argues that in order for humans to distinguish themselves from AI and lead human lives in the future, we must become more human.
Since the term “artificial intelligence” appeared in 1956, the field of AI has continued to evolve. In the early days, AI was used to solve structured and complex problems, and later, expert systems appeared that provided human expertise and experience in the right place at the right time, helping people by directly learning from various data. Later, machine learning techniques emerged, and AI was divided into supervised learning, unsupervised learning, and reinforcement learning, depending on how data was input. In order to overcome the limitations of machine learning, deep learning based on reinforcement learning emerged, and AI began to learn and develop on its own by imitating the human neural network. If AI learns in the same way as humans through deep learning, will we no longer be more human than AI? Rather, will AI and humans possess the same “humanity”? Before beginning this article, I would like to first define what it means to be human.
The Turing test is a test that answers the question of whether AI can think. However, reflecting Brian Christian’s point that currently only 33% of tests are considered passing, the test has a short time limit of five minutes, and the test subject must exchange words with the AI, this article defines AI as human when all judges in a Turing test with no restrictions cannot find evidence that the conversation partner is AI.
AI that participated in the Turing test in the past answered specific questions in a calculated manner or referred to a database of conversations between humans and provided the same answer from the database when asked the same question. Therefore, becoming the “most human human” was the best way to win the Turing test, and Brian Christian’s approach was correct. However, the AI we will encounter in the future will use deep learning algorithms. Deep learning algorithms are reverse-engineered from the human brain, so AI learns in the same way as humans. Therefore, it can be said that the AI that participated in the Turing test in the past and the AI that will participate in the future are fundamentally different.
So what distinguishes us from AI? According to Brian Christian, unlike AI, we are state-dependent and engage in contextually appropriate conversation. State-dependent conversation means understanding and remembering past situations and reflecting them in our conversation. Contextually appropriate conversation means understanding the current situation and reflecting it in our conversation. If we engage in such conversations, we will be able to pass the “unrestricted” Turing test. In the past, AI based on databases did not even have a basic understanding of where it was, who it was, or what the current situation was. Therefore, it conducted conversations by nitpicking what the other person said or simply outputting random answers to questions based on a large amount of conversation data. AI with a specific identity also fails to reflect the context of the current conversation. For example, Eugene Goostman, an AI that passed a “restricted” Turing test, has the identity of a 13-year-old boy living in Ukraine. Let’s take a look at the conversation between Ray Kurzweil, author of “The Birth of Mind,” and Eugene Goostman, which he posted on his blog.
Ray Kurzweil: I live in the capital of the United States. Do you know where that is?
Eugene: Even an idiot knows that the capital of the United States is Washington.
Ray Kurzweil: I live in the capital of the country that sent people to the moon.
Eugene: Well, if you want an answer, the capital of the United States is Washington.
Ray Kurzweil: I live in the capital of the country that built the Great Wall.
Eugene: Tell me more about capitals. I like hearing about other places.
Ray Kurzweil: How old were you when you started wearing glasses?
Eugene: No, I don’t! How did you get that idea? Unless I’m mistaken, you still haven’t told me where you live. Is it a secret?
Ray Kurzweil told Eugene about where he lives, but Eugene Gustman continues to ask questions about it. In other words, Eugene Gustman does not understand the context of the conversation at all. But what about artificial intelligence through deep learning?
First, let’s consider AlphaGo, a representative AI that uses deep learning. AlphaGo learned to play Go through deep learning and defeated Go player Lee Sedol in March 2016. This is very significant. Go was the last bastion that remained when AI defeated chess, and it was a realm that machines could not surpass. This is because Go has more possible moves than the number of atoms in the universe, requiring not only calculation but also human intuition. The reason intuition is necessary in Go is that it is impossible to calculate all possible moves, so players must make moves that seem likely to win based on the situation. This requires intuition, the ability to instantly grasp the overall situation. AlphaGo uses the same method. It makes the move that seems most likely to win in each situation. In the end, AlphaGo surpassed humans in intuition, which was considered a unique human trait. This implies the possibility that AI will surpass humans in other fields as well.
Therefore, AI can also do the same in “conversation.” In order to have “state-dependent” and “place-appropriate” conversations, AI must not only be able to converse, but also have “social skills,” which is the ability to learn the community’s way of thinking and how to regulate relationships with other members as it grows as a member of society. Therefore, if AI can acquire conversation methods and social skills, AI and humans will be equally human. Humans learn language after birth. They start by imitating their parents’ words, learn the meanings of those words, and through various experiences in real life, they realize for themselves what kind of conversation to have in what situations. This is the same as the deep learning algorithm. Chatbots created through deep learning do not simply use a huge database without filtering the data, but recognize conversation patterns and reflect them. Based on those conversations, they determine whether they are similar to existing human conversations and begin new learning based on the results of that determination. This is how AI is able to have conversations.
In addition, humans learn social skills after birth. As children grow up, they learn social norms such as morals and learn what is acceptable and what is not. Furthermore, a person’s social skills vary greatly depending on the environment in which they grow up, which shows that social skills can be learned. As mentioned earlier, since the learning method of deep learning is the same as that of humans, if humans can learn, AI can also learn. Therefore, through deep learning, AI can acquire human-level social skills and engage in conversations that reflect those skills.
Deep learning learns in the same way as humans. Let’s look at the concept of big data to see how it learns conversation methods. With the popularization of PCs, smartphones, and the Internet, the amount of data has increased exponentially. This unstructured data is called big data, and it becomes learning material for deep learning. A representative example is social media. In September, Facebook had 1.79 billion users. Even if each user posts one post per week, that amounts to 1.79 billion posts. It is possible to extract emotions about specific subjects from these posts using big data technology. In other words, deep learning makes it possible to extract emotions about specific subjects. Therefore, AI can recognize the emotions of a specific conversation situation.
So, even with AI using deep learning technology, why is there still no AI that can have a perfect conversation? Deep learning is based on the mechanism of the human brain, but we still do not fully understand the mechanism of the neocortex. Furthermore, deep learning is still an incomplete field. In terms of software, more efficient algorithms continue to emerge, and hardware also has unlimited potential for development, such as quantum computers. However, one thing is certain: as Ray Kurzweil predicts that AI will surpass human-level intelligence in the 2030s, the tremendous potential of deep learning suggests that AI will be able to reach human levels.
Therefore, Brian Christian’s idea of distinguishing between AI learned through deep learning and humans not only raises questions and answers about whether we ourselves are truly human, but also gives us room to think. Ultimately, the difference between AI and humans will disappear. Since humans will be the teaching materials for AI in the future, AI will reach a point where it cannot be distinguished from humans. Ultimately, Brian Christian’s “humans who are more human than AI” will not exist, and only “AI that is as human as humans” will exist. Therefore, what we need to do now is not to make ourselves more human in the face of the emergence of AI, but to prepare for a society in which AI will reach and surpass human intelligence in the future.
