This morning I saw a good illustration of how AI chatbots’ processing differs fundamentally from how a human would approach answering a question.
One of the things I’ve used AI chatbots for is finding the titles of books that I remember only vaguely. My success rate here is pretty darned good. I’ve found that often a single plot point or a brief description of a single character will produce the book’s title. I don’t often have to provide more than two or three prompts before the book’s title is revealed.
Sometimes I’m surprised that the chatbot doesn’t get it on the first try. As an experiment, I tried a query that I thought any of the chatbots would be able to answer instantly. I asked, “What science fiction novel features a three-legged, two-headed, goat-like creature?”
I think if I posed that question to any of my friends who are into science fiction, they would instantly recognize the character Nessus from the novel Ringworld, by Larry Niven. I posed that question to Meta AI, Copilot, Gemini, and ChatGPT. Not one of them mentioned “Ringworld” in their first response.
Curious. Here are the suggestions from the first prompt:
- A Sietch from the novel Dune.
- A Shmoo?
- A Niben?
- The Ysengrin’s helper’s tripod grimp.
- A creature from Dirk Gently’s Holistic Dective Agency by Douglas Adams. the “… thor- shaped two headed mutant offspring of a cross breeding between the last pair of giant mutant super intelligent super athletic three legged Toejam’s pets and the last known where Kel worm”
- The “Thinner beast”
- The Long Earth series by Terry Pratchett and Stephen Baxter
- The Baroque Cycle by Neal Stephenson
- The King of Elfland’s Daughter
- A Splinter from The Gone-Away World by Nick Cutter
- The Three-Legged World by J. T. McIntosh
- The Mount by Carol Emshwiller
- The Tripods series by John Christopher
- The Hitchhiker’s Guide to the Galaxy by Douglas Adams
- Eeyore the Goat from The Gods Themselves by Isaac Asimov
When I provided more information, “I think the creature was called a Puppet, or something like that”, Meta AI and Copilot correctly identified Ringworld. ChatGPT mentioned The Puppet Masters by Robert A. Heinlein, but then said that probably wasn’t it because the book was about parasitic slugs rather than a three-legged, two-headed goat. Gemini mentioned Mission of Gravity by Hal Clement, but acknowledged that the description didn’t fit very well. Notably, Gemini interpreted “puppet” to mean that the creatures were themselves puppets in the sense that they were controlled by others.
Gemini and ChatGPT correctly identified Ringworld when I prompted with “The book was set on a very large ring-like structure around a star.”
Updated a few hours later
I asked Grok the same questions. It initially suggested Project Pope, by Clifford D. Simak. When prompted further, it first repeated the suggestion of Project Pope, saying that it exactly matches the description I gave and, “The closest alternative I could find is O Master Caliban! (1976) by Phyllis Gotlieb.” Finally, when I mentioned that the novel was set on a large ring-like structure, it suggested Ringworld, saying:
…the expedition’s sponsor, Nessus—a Pierson’s Puppeteer, a herbivorous alien species characterized as a three-legged, two-headed, goat-like creature (with a body resembling a large, maned goat or ostrich, flexible snake-like heads emerging from the chest, and a mane that serves as a sensory organ). Nessus and his kind are central to the plot,
What I find interesting, and disturbing, about Grok is the arrogantly authoritative tone it takes in its responses. Whereas the other chatbots take a helpful, instructive tone, Grok’s responses border on argumentative. Note that this observation is based on only one interaction with Grok. Today is the first time I’ve used it.
Why was that so hard?
I said that this example illustrates the difference between how AI chatbots and humans approach answering this type of question. Let me expand on that.
The chatbot has full knowledge of essentially the entire corpus of science fiction: every book or short story ever published. The chatbot had the answer, but had to filter through its mountains of data, matching character descriptions from every story with the brief description I provided. And then it had to select the likely candidates. Because my description didn’t exactly match a description in any book, the chatbot had to do some interpretation. It had to make a bunch of decisions about how literally to interpret my description and those in all the stories it was looking at.
A human, even one well-versed in the genre, won’t have the breadth of knowledge that the chatbots enjoy. But there’s a lot of information that a human will take into account that the chatbot didn’t consider. Perhaps the most important of all will be individual books’ popularity. Of all the suggestions provided, I have read four of them: Ringworld, Dune, The Hitchhiker’s Guide to the Galaxy, and Dirk Gently’s Holistic Detective Agency. I might have read Heinlein’s The Puppet Masters. Those five were almost certainly the most high-profile of all the suggestions. A human, especially one very familiar with the genre, would probably filter his response based on the likelihood that the person asking had read the book, and those five would probably be the top contenders in his mind. Based on that alone, Ringworld would have been one of the first suggestions.
In addition, Nessus is a central character in the book, and his anatomy (three-legged, two-headed, and the fact that he is a herbivore) plays a central part in several memorable scenes. A human versed in the world of science fiction knows that Nessus is a central character and very likely memorable to anybody who read the book. The chatbot doesn’t share that insight. It has the information, but probably won’t prioritize it.
And I think that’s the key. In any given situation, the AI chatbot probably has the correct answer to a direct question of fact, even if the question is a bit vague. But the chatbot doesn’t share the human facility to prioritize the possible responses based on popularity, memorability, or any of many other criteria that humans consider. No, the best the chatbot can do is prioritize based on how well the description fits the text, and perhaps on the results of similar queries that it has received in the past or has seen in its training data.
Chatbots are useful, but they’re not experts in any field. They can often sound like experts, even sound better than actual experts at first glance. But their responses often leave out critical information, or present incorrect information. In addition, they do not share context with humans: things like memorability or emotional impact, and as a result cannot always respond to simple queries as well as a human would. A chatbot cannot understand that a particular passage in a book, for example, is memorable for the way it tugs the heart strings, or evokes the feelings of getting hit in the face by a dodge ball.
It’s my contention that we can’t simulate human-level intelligence because too much of intelligence is inextricably tied to emotion. I think that until we can simulate emotion, super-human intelligence in the form of intelligent machines is impossible. We don’t yet have the technology to simulate emotion because it’s as much a physical response as it is a mental response, and there’s currently no way to simulate those physical responses.
