When AI Thinks, Do We?
AI seems to be changing verbal repertoires, and maybe not for the better.
The real question is not whether machines think but whether men do.
B. F. Skinner
Have you ever wondered what would happen if everyone on earth jumped at the same time? Or how many times you can return items to Amazon before they ban you? Or have you ever wished you could have someone create a Spotify playlist for you of songs from the 90s with the names of cities or people in the title?
Well, it turns out there’s an app for that. But you probably already know that. If you’re under 30, odds are you are using ChatGPT or something like it. If you’re over 30, odds are that you have at least used it or something like it, even if you didn’t know it.
ChatGPT, Claude, Gemini, and other programs like them, are AI chatbots that can reply to emails for you, screen your calls, plan your next vacation, and validate your worst ideas. (According to OpenAI, who makes ChatGPT, about half of the time people ask for information, and about 40% of the time they want to perform a task—usually writing something.)1
The hope is that AI will make us smarter and better at what we do. We’ll be more accurate, insightful, reliable, and efficient. But there are reasons to be skeptical, if not worried. The consequences of AI could be significant, upheaving our economy and eroding our society. And the machines might rise up and kill us.
But even if AI doesn’t kill us, it will make us dumber.
How? By turning us into readers.
As a college professor, I warn my students that reading is not studying. Studying is practicing. If you want to learn to hit a baseball, you don’t just watch baseball games. You have to stand up and swing a bat at a ball. Over and over. If you want to do well on an exam, you can’t just read the book and your notes. You should sit down and practice answering the questions that might be asked. Reading is important, but it will only get you so far, unless your professor tests you by asking you to read out loud.2
What does this have to do with AI chatbots?
We might read the answer an AI chatbot gives us but do little more after that than cut and paste it into a new place. But reading is not thinking. We can read the answer, slowly and carefully, but that doesn’t mean we can now answer the original question or any new ones. I mean, can you read the following sentence?
Neverland damn the tam about the pig around the room under the definitely upward and bound.
Great. Can you explain it to me? Didn’t think so.
But you probably think I am setting a strawman. That was gibberish. How about this one?
We investigated ultrafast defect-lattice dynamics in diamond using the Ns:H−C0 defect, an analog of bond-centered hydrogen in semiconductors.
I am sure you can read it, say it, and copy it down. But I doubt you can explain it, even though some people can.
My point is that reading is not the same as thinking. Reading an answer to a question is not the same thing as knowing something.
From the standpoint of a radical behaviorist, my primary concern is that AI is changing verbal repertoires, and maybe not for the better. The new verbal sequence looks like this: A question is asked, dictation is taken, then ChatGPT responds. At this point, the listener engages in simple textual behavior. They read the answer.
What’s the problem?
This is where it gets technical and aimed at behavioral psychologists. (But here’s a cheat sheet.)
Questions should evoke echoic (repeating) and intraverbal (saying different but related things) behavior. We repeat the question to ourselves, then respond verbally in ways that are controlled by the verbal stimulus but that have no point-to-point correspondence with it. When someone asks, “How do I,” we don’t just respond, “How do I?” That would be an echoic relation, because each part of our response corresponds (point-to-point) to the question. Instead, we answer, “You…” Here, the form of what we say is determined by what was asked, but there is no point-to-point correspondence—the answer is something different from the question.
Questions, being verbal stimuli, start to boil the cauldron of our behavioral repertoire, to invoke Dave Palmer, bringing various members of various response classes to strength and, ultimately, one or a few responses percolate and jump past the prepotency threshold to be emitted.
If we’ve never answered a question like the one we were just asked, there might be no responses to come to strength. Then we have a problem to solve. The question evokes verbal responses that are not themselves answers to the question, but that might alter the environment in such a way as to evoke the answer.
But with AI, the question is mostly likely to evoke dictation—we type the question into a chatbot. Then the chatbot responds with an answer and we engage in textual, not intraverbal, behavior. That is, we read the answer provided by the chatbot. Probably the strongest response evoked in the future (by the question) will be the echoic and a fragmentary textual response.
Or maybe not even that. Recent research suggests that students who use ChatGPT to respond to writing prompts cannot recall a single sentence they “wrote.” Probably because they never said, wrote, or typed anything—they struck some key combination to cut and paste the text from one place to another. They never engaged in behavior relevant to the question, so there is no relevant behavior to be evoked later by the same question or some other prompt. Students who did not use ChatGPT could recall some or many sentences, presumably because they had actually written them.
This illustrates the crux of the problem. Using AI, our repertoires are not changed in ways that allow us to behave more effectively with respect to the subject matter absent the continued use of AI. But, even then, the actual “answer” to the question is not part of our verbal repertoire in any meaningful sense. It exists only in the physical memorandum of the AI text. And reading cannot occur without a text, meaning the behavior cannot occur without environmental supports.
AI is a crutch. Great if you need it, but better to do without it, if you can.
The stimulus control established in the textual relation is not very useful in real-time interactions, provided anybody actually still has those. To behave effectively, you need an additional, and substantial, source of stimulation—a text or audio or video clip generated by AI that you can see and repeat in some way. The relevant circumstances alone will not occasion behavior appropriate to the situation absent such a technological middleman.
So I worry that the more we let machines think, the less we will.
Funny observation. I used an em dash there, which, it turns out, is an odd signature of writing generated by ChatGPT. AI chatbots, especially ChatGPT seem to love the em dash, more so than the human population. But rest assured, no AI was used in the composition of this essay. I just like em dashes. And I used them way before I knew what ChatGPT was.
Of course, it is important that you read and listen (to your teacher), especially at first. Just like you might want to watch a baseball game before you first try to hit a baseball. It’s just that it takes more than that to learn to do well the things you actually need to do, like hit baseballs and answer questions.
Great post, Matt. Very insightful. I have also been worrying about this, especially the bypassing of the hard work needed to generate VB, as opposed to just reading it. I like the way you put it that the repertoire is not changed in a meaningful way. Good stuff!
Do you think there are legitimate educational uses for LLMs?
Your analysis makes good sense, especially the reading of outputs without subsequent active (collateral) responding. The risk seems notable for learners encountering new subject areas. My oldest son is 10, and the moment he witnesses the output from a chatbot on a difficult academic question, he will be hooked. The immediacy, seemingly accurate, and confident response you receive from frontier models is striking. On a related note, over the last 8 months I mostly talk to my computer using Wisper Flow (at this moment, my count of dictated words is 419,266). This likely has behavioral implications as well. Thanks for sharing your insights, Matt.
As an aside, when you expand the access of frontier models (or as others such as Ethan Mollick call the "harness"), such as moving from a more harnessed chatbot to a less-harnessed AI that has access to your computer and the web in form of Claude Code, it alters how you work and think about working. I am concerned about the negative implications here as well.