Has Artificial Intelligence Truly Achieved Intelligence?

In 1950, British computer scientist Alan Turing proposed an experimental method to answer the question: Can machines think? He suggested that if a person cannot determine whether they are communicating with an artificial intelligence machine or another person after five minutes of questioning, then it demonstrates that artificial intelligence has the ability to mimic human intelligence. Although artificial intelligence systems have not yet reached the level of passing the specific conditions set by Turing during his lifetime, he speculated that "in about 50 years, it will be possible to program computers... to make them perform better in simulating games, with the probability of guessing correctly after five minutes not exceeding 70%." Today, more than 70 years after Turing proposed this idea, no artificial intelligence has successfully passed this test and met his specific conditions. However, some systems have come very close to passing the test. A recent experiment tested three large language models, including GPT-4. Participants engaged in a two-minute conversation with either a human or an artificial intelligence system. The artificial intelligence was instructed to intentionally make small spelling errors and stop responding when the tester became too aggressive. Under these conditions, the artificial intelligence successfully deceived the testers. When paired with an artificial intelligence robot, testers only had a 60% accuracy rate in correctly guessing whether they were communicating with an artificial intelligence system. Considering the rapid advancements in natural language processing systems, we may see artificial intelligence pass Turing's original test in the coming years. However, is imitating humans really an effective way to test intelligence? If not, what alternative standards can we use to measure the capabilities of artificial intelligence? Limitations of the Turing Test Although passing the Turing Test indicates a certain level of intelligence in a system, it is not a definitive test of intelligence. One of the issues is that it can produce "false negatives." Current large language models often immediately claim not to be human. For example, when you ask a question to ChatGPT, it frequently starts its response with the phrase "As an artificial intelligence language model." Even if an artificial intelligence system has the ability to pass the Turing Test, this programming approach would override that ability. The test also has certain types of "false positives." As philosopher Ned Block pointed out in a 1981 article, a system only needs to hardcode human-like responses to any possible input to pass the Turing Test. Furthermore, the Turing Test focuses on human cognition. If artificial intelligence cognition differs from human cognition, professionals will be able to find tasks where artificial intelligence and humans perform differently. Regarding this issue, Turing wrote, "This objection is a very strong one indeed, but at least we can say that if we can construct a machine to play the imitation game satisfactorily, we need not be troubled by this objection." In other words, passing the Turing Test is good evidence of a system's intelligence, while failing the test is not good evidence that the system lacks intelligence. Additionally, this test does not effectively measure whether artificial intelligence is conscious, capable of experiencing pain and pleasure, or has moral significance. According to many cognitive scientists, consciousness involves a range of mental abilities, including having working memory, higher-order thinking, and the ability to perceive the environment and simulate oneself moving within it. The Turing Test cannot answer the question of whether artificial intelligence systems possess these abilities. The Increasing Capabilities of Artificial Intelligence The Turing Test is based on the logic that humans possess intelligence, so anything that can effectively mimic human behavior is considered intelligent. However, this idea does not tell us about the nature of intelligence. A different way to measure the level of artificial intelligence is to think more critically about what intelligence is. Currently, there is no single test that can authoritatively measure artificial intelligence or human intelligence. At the broadest level, we can view intelligence as the ability to achieve a range of goals in different environments. More intelligent systems can achieve a wider range of goals in a broader range of environments. Therefore, the best way to track the progress of general artificial intelligence system design is to evaluate their performance in various tasks. Machine learning researchers have developed a range of benchmarks to do this. For example, GPT-4 is able to correctly answer 86% of questions in a wide range of professional language understanding multiple-choice tests. This is a benchmark for large-scale multitask language understanding, covering a range of university-level academic subjects. It also scores well in AgentBench, a tool that measures the ability of large language models to act as agents, such as browsing web pages, making online purchases, and participating in games. Is the Turing Test Still Applicable? The Turing Test is an imitative measure that assesses the ability of artificial intelligence to simulate human behavior. Large language models are expert imitators, and this ability is reflected in their potential to pass the Turing Test. However, intelligence is not synonymous with imitation. There are as many types of intelligence as there are goals to be achieved. The best way to understand artificial intelligence is to monitor its progress in developing a range of important capabilities. At the same time, when discussing whether artificial intelligence is intelligent, it is important not to constantly "move the goalposts." As the capabilities of artificial intelligence rapidly improve, critics of the idea of artificial intelligence intelligence continually find new tasks that artificial intelligence systems may struggle to accomplish, only to find that they have already surpassed another obstacle. In this context, the question is not whether artificial intelligence systems possess intelligence, but rather, what "type" of intelligence they may possess.