ChatGPT and the Robot Dog.
The Robot Dog fascinates me: People conspire silently to be excited by how it resembles and approximates a Real Dog, and enjoy ignoring the ways it fails. Even if Real Dogs are available.
The Robot Dog is meant to resemble a Real Dog. It doesn’t have any aim or purpose independent of semblance to Real Dogs, else it would not be a Robot Dog. (It would just be a robot.) It is meant to resemble a Real Dog physically (though not too closely) and especially in mannerisms: It barks, wags its tail, begs, and walks on four legs. That fact that we know it’s not a Real Dog is essential to the delight we participate in over its approximation to one: If it were simply a Real Dog, we would not feel that same delight (though probably take a different delight in those same mannerisms). For the Robot Dog, our enjoyment requires the tension between it approximating a Real Dog and yet not being one.
The phenomenon of ChatGPT—and let us exclude other “AI” applications for the moment—strikes me as another Robot Dog phenomenon. First, my friends in IT insist that ChatGPT has nothing to do with artificial intelligence; it is a programme for machine learning; there’s no intelligence involved, not even by artifice. Another way of thinking of this point is to recognise that ChatGPT is an elaborate version of a standard Google search: our “questions” trigger Internet searches whose results are compiled and integrated into text-responses. (This in turn ought to remind us that if ChatGPT assembles its responses from Internet content, the “answers” to our questions will consist of as much error [and drivel] as can be found on the Internet broadly. The Internet now contains progressively more “AI”-generated content; so the proportion of dubious results from ChatGPT can only increase over time.) For those of us who remember the introduction of domestic access to the Internet, we saw a search engine called “Ask Jeeves” appear in the late 1990s. It arrived between the first search engines, including Lycos & Magellan, and the later ones like Google. This search engine encouraged the surfer to make queries in the form of questions (rather than non-grammatical lexical strings) under the pretence that “Jeeves” would retrieve an “answer,” as every good manservant should. ChatGPT provides more convincing answers than did Ask Jeeves, but the grammatical competence of the results lures us to regard these results as “answers” and not just extracts from the first page of Google results. Educated colleagues tell me daily that they “asked” ChatGPT something and it “answered” them. Does ChatGPT “understand” what we “ask” it?
On the question of “understanding,” we are reminded of Searle’s famous Chinese Room problem. A man is enclosed in a room without furniture, containing baskets of different Chinese characters. The man does not speak Chinese. But when Chinese characters are passed into the room through a window, he consults a book (in English) that tells him which pattern of Chinese characters he should assemble and feed back out through the window, based on whatever pattern of Chinese characters were just received in through the window. He doesn’t realise that the strings of characters coming in are called “questions,” and the strings that he returns out through the window are called “answers.” Does the man “understand” Chinese? Clearly not. There is an argument that maybe the whole room (including the man and his rule book) speaks Chinese, but we cannot help being unnerved if not horrified by the prospect of confusing that man with another who “understands” Chinese.
The Robot Dog fails to replicate a Real Dog completely, and likewise ChatGPT fails to replicate an intelligent and knowledgable person who can answer our questions. And we know it’s not a real person, right? We say, “ChatGPT told me…” but that’s just a turn of phrase we delight in using. It’s not even C3PO is it? The disturbing element for me is the rush to suspend disbelief (and maybe even to ignore glaring failures) in order to imagine (or speak as though) ChatGPT were such a person. Colleagues have said they have consulted ChatGPT for psychotherapy. But I suspect that ChatGPT probably enjoys the same advantage as Wikipedia insofar as it is used typically by people searching for information they don’t already have; they assign it research tasks they do not carry out themselves afterwards. And it is difficult to evaluate the correctness of an “answer” in an area that one is not already familiar with. When a guest at dinner claims to speak seven languages fluently, chances are there is not another guest present who could verify the claim. Encyclopaediae are seldom used by people looking-up things they already understand well, because that’s not what they are for. Yet the only way to evaluate something like Wikipedia is to look-up things one is already an expert in1. Or to ask ChatGPT something one already knows the answer to. One colleague presented ChatGPT with a copy of an essay she had written, and asked it to extract quotations. It produced sentences within quotation marks, but they weren’t extracts from her essay. It literally presented her with false quotes, moments after she gave it the complete text. When she “told” ChatGPT that these were simply not quotations, ChatGPT “apologised,” with the text, “You’re right. They were not quotes. My bad.” Did it “know” it was lying? Did it not “realise” it was lying until she pointed it out? Is that possible for a machine?
Renowned journalist, Glenn Greenwald, made an excellent demonstration of the failure of ChatGPT even to approximate impartial presentation of information in response to requests. Instead, answers were so politically skewed, ChatGPT literally refused to provide answers for (only) some politically-charged questions (System Update #36). Again, problems with the whole application are easier to recognise when we are asking for information we already have (or don’t need) and can therefore evaluate.
If a person, say a research assistant, returned false quotes in response to a request, or produced work of explicit political leanings, and only acknowledged the fact after having been “caught out,” we would not work with such a person again. It’s hard to call these exchanges with ChatGPT deceitful, when it can only be a failure of the programme to do the (only) thing we need it to do. It’s worse than deceit, because deceit requires knowledge of the truth. Deceit doesn’t belong to computers, though it might belong to persons programming them. And it’s almost lèse-majesté to say AI is poor; we must say at least that it’s on the way of continuous improvement, and it’s just not perfect yet. The Robot Dog is not deceiving us when it only approximates a Real Dog, yet we get (or stay) excited about it. And people remain excited enough about “AI” to continue using ChatGPT for things they could easily (and more competently) do themselves.
Giac Giacomantonio, Ph.D.
Wikipedia claimed that I have two daughters. I’m sure I do not.