Simulating Intelligence vs. Possessing It: Turing and Chinese Room

Introduction

Researchers have been attempting to develop artificial intelligence for more than half a century, but it has become increasingly apparent that intelligence is a multidimensional construct that is incredibly difficult — and perhaps even impossible — to truly simulate. As the review that follows will demonstrate, the assertion that successful simulation of intelligence is essentially equivalent to the possession of intelligence is erroneous, due both to the multidimensionality of intelligence and to its other nebulous qualities that preclude successful simulation from being the essential equivalent of possession (Pogio & Meyers, 2016).

This paper reviews the relevant literature to show that the Turing test and the Chinese Room argument each fail to provide an adequate response to this claim. A discussion concerning whether the successful simulation of human reasoning is in principle possible — and whether the answer to that question is inherently dependent on the answer to the main question — is followed by a summary of important findings in the conclusion.

The Turing Test and the Claim

The test developed by Alan Turing on intelligence, commonly known as the Turing test, was an important development in the ongoing effort to create true artificial intelligence (Shieber, 2016). In sum, the Turing test attempts to produce artificial responses that mimic human communications to the extent that humans are "fooled" into thinking that the machine is another person by holding a conversation (Blackmore, 2011). The manner in which the Turing test tries to achieve this goal, however, is fraught with constraints that prevent it from being the equivalent of possessing intelligence. As Marcus and Rossi point out, "Within the field, the test is widely recognized as a pioneering landmark, but also is now seen as a distraction, designed over half a century ago, and too crude to really measure intelligence" (2016, p. 3). Indeed, given its complex nature, many researchers believe that even the best tests cannot accurately measure human intelligence, and developing computer-based applications that can simulate intelligence are therefore limited by a lack of definitional clarity. Marcus and Rossi conclude that "Intelligence is a multidimensional variable, and no one test could possibly ever be definitive truly to measure it" (2016, p. 4).

While the Turing test represented an innovative approach when it was developed, the test has since become more of an "exercise in deception" rather than a "true measure of anything especially correlated with intelligence" (Marcus & Rossi, 2016, p. 4). As an example, Marcus and Rossi cite the chatbot "Eugene Goostman," which purportedly became the first AI application to pass the Turing test by deceiving one-third of a panel of judges into believing it was a 13-year-old Ukrainian boy with limited English fluency. The program even responded appropriately to a joke from a human judge, an attribute that Hofstadter (1985) maintains is an essential element of intelligence.

According to a report from Aamonth (2014), this seminal event essentially confirmed Turing's original prediction that AI would become sufficiently advanced by 2000 to trick humans into believing they were interacting with other humans at least 30% of the time. The Goostman chatbot, however, succeeded in deceiving human judges by "mainly ducking questions and returning canned one-liners; it cannot see, it cannot think, and it is certainly a long way from genuine artificial general intelligence" (Marcus & Rossi, 2016, p. 4). The ability to "return canned one-liners" is also a characteristic of the Chinese Room argument, which has been used to underscore the fundamental differences between simulating intelligence and possessing true intelligence, as discussed further below.

The Chinese Room Argument

Developed as a thought experiment by John Searle (1980), the Chinese Room argument has frequently been cited as a textbook example of the manner in which intelligence can be simulated but not replicated. The experiment is also evocative because it highlights the basic issues involved in conceptualizing variables such as "intelligence," "understanding," and "thinking." Imagine being in a room surrounded by filing cabinets full of coded responses and guidelines specifying which response to select when presented with requests written on cards by anonymous users. The individual in the Chinese Room does not need to know — or care — what the symbols mean, only which response is most appropriate according to the guidelines.

Given the enormous complexity of the Chinese alphabet and language, it is reasonable to suggest that the individual would require a long time to locate the appropriate response at first. But as proficiency is gained over time, the output process would become far more efficient. This outcome, however, does not mean that the individual has learned anything about Chinese — only that proficiency in answering inquiries has been acquired.

This process is analogous to what takes place when humans request information from computers, and the Chinese Room argument makes it clear that the process has only been facilitated by digital processing rather than by creating truly "thinking computers" (Overskeid, 2005). Likewise, Searle (1990) concluded that "A [computer] program merely manipulates symbols, whereas a brain attaches meaning to them" (p. 26). Moreover, the symbols themselves are completely meaningless to a computer — or to the Chinese Room operator — but understanding what they mean is not required in order to follow the guidelines and generate outputs that may be highly convincing in their authenticity and relevance, yet still do not rise to the possession of true intelligence.

It is important to note, however, that despite the limitations of the Turing test and the constraints identified in the Chinese Room argument, the potential for the successful simulation of human reasoning — at least in principle — remains a possibility for the future, as discussed below.

Conclusion

With computer processing speeds doubling roughly every year or so, it is reasonable to conclude that by the end of the 21st century, human intelligence will be sufficiently simulated by artificial intelligence applications to the point where they interact with humans in a meaningful and seamless fashion — superior to even the HAL 9000 in 2001: A Space Odyssey. This success at simulation, however, will still not imbue these applications with intelligence in the human sense. Because the process is cumulative, the day may well come when computer programs become sentient and acquire what humans consider intelligence; but it will inevitably be of a different sort unless and until the human-machine interface becomes sufficiently interrelated.