Could a machine ever use and understand language the way people do? A take on Turing’s question in 2020

Here I analyze the given question in current perspective (2020), by juxtaposing it against Turing’s seminal take and arguments on the question: ‘Can machines think?’ using the ‘imitation game’ in 1950 [7]. Due to the relative subjectiveness and broadness of the notion ‘the way people do’ (hereafter, humanlike), for this write-up, I’ll presume the setup similar to the imitation game, and constrain the discussion as such.

Arguments for ‘Yes’: Of the contrary views posited by Turing, the one most relevant even after 50+ years is the ‘argument of consciousness’ (AoC). It has seen reincarnations in various forms and arguments in AI research and philosophy by many luminaries following Turing, like Dreyfus, Searle, Harnad, Haugeland [2, 6, 4, 5] to name a few. I concur with Turing’s observation that the ‘argument of various disabilities’ (AoD) is often a disguised form of the

AoC. Examples of the AoD are still prevalent (to exemplify, take the various faux pas of voice assistants like Siri). Lady Lovelace’s objection, that machines are unable to do anything novel outside what is programmed, is weak if not debunked in the current perspective. Adding to Turing’s formidable counter to this argument, I’ll refer to the AI research sub-domain of ‘library-learning’ [3] as a counter-example — where machines learn to form new (unknown ex-ante) routines from programmed primitives.

Delving in the AoC, i.e. the lack of ‘intentionality’ [1] or a fortiori ‘human phenomenology’ (emotions, ego, imagination, moods, consciousness etc.), we see that it is largely moot in AI’s humanlike natural language understanding (NLU) debate — especially in the imitation game setup, where all human outward physiological attributes are excluded. A robotic lab assistant’s (textual) response: “I’m terribly ashamed and sorry for burning down the lab” is humanlike irrespective of whether any ‘feelings’ behind are. Turing alludes to this by rhetorically asking whether putting on ‘artificial flesh’ is necessary on a hypothetical ‘thinking’ machine. Here, if NLU entails generating humanlike language responses, then the advancement trajectory of current SOTA models (like GPT3) enables us to answer ‘yes’ to our question.

Contrarians may raise the issue of ‘humanlike’ vs. ‘simulation of humanlike’ in negating such NLU evidence. John Searle’s ‘Chinese room argument’ [6] is an illustrative, well-known e.g. of such takes (which is again, a variant of the AoC). There are many classical responses to this argument, including the systems and solipsistic point of views: although the man in the room doesn’t learn Chinese, the system as a whole does; or you need to be the machine to know whether it feels etc. Regardless, the preceding ‘functionality’ argument suffices against such takes.

Arguments for ‘No’: A stronger argument negating machine’s ability to ever reach humanlike NLU hinges on Chomskyan universal grammar (UG) proposition, and if agreed, whether the aspects of language faculty innate to humans (independent of sensory experience) are required for exhibiting humanlike NLU.

Humanlike NLU, even in our constraint setup, requires humanlike amalgamation of experience and learning from it. Turing alludes to this in the last section with his ‘learning machine’ analogy to teaching language to a human child (birth, education, experience). However, as the poverty of stimulus (PoS) argument stipulates, human children can acquire NLU largely without negative samples – a feat currently impossible for machines. PoS supports UG. Subsequently, introduction of genetic components raises the ‘evolution’ problem, like how do we imbue the hereditary material in machines, and humanlike (requiring innate mechanisms) machine learning.

Thus, even if machines can exhibit close to humanlike NLU trained for specialized domains, it can never achieve humanlike ‘generalized’ NLU and use language the way people do.

How could we resolve the question? Conclusively resolving the question is difficult due to the subjective notion in defining the spectrum ‘humanlike’. Thus, we can constrain the domain and define a baseline threshold of ‘humanlike’ in tests akin to Turing’s. It’s important to stress the omission of outward physiological human attributes in these tests, to not digress from the main question.

A cursory stringent version of the Turing’s test for e.g. could be a blind study with online chatbot(s) that partakes in channel discussions of adequate length with (controlled) group of human participants across various topics. In each channel, the participants are taught a new concept about the topic using chat interface. Later, discussions/questions pertain to this learned concept. The chatbot is marked on its performance in such discourses and passing the baseline threshold of ‘humanlike’ could be an accepted way to resolve the question.

Reference

1. Daniel C Dennett and John Haugeland. Intentionality. In The Oxford companion to the mind. Oxford University Press, 1987.
2. Hubert L Dreyfus, L Hubert, et al. What computers still can’t do: A critique of artificial reason. MIT press, 1992.
3. Kevin Ellis, Catherine Wong, Maxwell Nye, Mathias Sable-Meyer, Luc Cary, Lucas Morales, Luke Hewitt, Armando Solar-Lezama, and Joshua B Tenenbaum. Dreamcoder: Growing generalizable, interpretable knowledge with wake-sleep bayesian program learning. arXiv preprint arXiv:2006.08381, 2020.
4. Stevan Harnad. The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1–3):335–346, 1990.
5. John Haugeland. Artificial intelligence: The very idea. MIT press, 1989.
6. John R Searle. The chinese room revisited. Behavioral and brain sciences, 5(2):345–348, 1982.
7. INTELLIGENCE BY AM TURING. Computing machinery and intelligence-am turing. Mind, 59(236):433, 1950.