Doug Smith has been a software developer for three decades. He writes extensively about the impact of technology on culture, especially within the Christian church. He is an anti-LLM absolutist, and from what I can tell also an anti-AI absolutist. Large language models, or LLMs, being a subset of artificial intelligence, or AI, it’s strictly speaking possible to be entirely against LLMs and still allow some positive room for AI. But in my reading of Smith’s work, I can discover nothing positive from him about either.
Smith, whose website is thatdougsmith.com, has for some time now been calling attention to destructive effects of digital technologies on people, especially the young. To that end he wrote a book titled [Un]Intentional: How Screens Shape Your Desires and How You Can Break Free. To his credit, he has identified real dangers, especially how digital technologies are often designed to be addictive, and how their addictiveness is often the result of Big Tech intentionally exploiting human weaknesses, everything from negative emotions, especially fear, to brain-reward systems (dopamine hits).
Just as there were teetotalers to whom we can thank the Prohibition of the 1920s, Smith is a teetotaler with regard to AI, wishing to discourage LLMs and AI as far as possible. But the cat is out of the bag, so Smith is not going to be able to enforce a blanket prohibition against AI in the same way that, say, William Jennings Bryan was able to help get all liquor (beer, wine, hard drink) outlawed. But Smith is seeking such a prohibition at least within his circle of influence.
I regard Prohibition of the 1920s as a bad thing. Granted, alcohol is implicated in many of our problems. But prohibiting it created still more problems, not least a society of scofflaws who went on drinking despite the Eighteenth Amendment, which was supposed to make Prohibition the law of the land. Additionally, Prohibition invited organized crime to circumvent it, greatly empowering it.
Of course, there are disanalogies here between prohibiting LLMs and prohibiting AI. A closer analogy might be the Luddites, early-19th-century English textile workers who, fearing the loss of their skilled livelihoods to industrial machinery, organized covertly to destroy the new machines. But this movement was rooted less in technophobia than in economic protest against the disruptive social effects of rapid mechanization.
Smith’s concern is with the moral ill effects of AI technology. He is a student of and advocate for the Bible, and wants to advance a Judeo-Christian morality (as do I). Yet on biblical grounds there’s no good argument for a blanket prohibition of alcohol (Jesus drank wine and not grape juice), though there is against its abuse. Nor for that matter is there any biblical prohibition on mechanization in the textile industry.
As it is, the Bible doesn’t speak to AI, except perhaps indirectly as an idol, which it can be but doesn’t have to be. Indeed, anything can be turned into an idol by giving it inordinate importance. So there’s nothing exceptional here with AI except perhaps that it inspires widespread fascination and so may more readily be turned into an idol than other things.
AI in Education
In an email correspondence between Smith and me that cc’d some colleagues, he raised concerns about a piece that I posted here at Substack titled “Edification vs. Enhancement—The Non-Transhumanist Vision of AI in Education.” In it I addressed the role of AI in education and challenged what I regard as a false dichotomy between AI either being misused by students (as in cheating with LLMs) or else needing to be barred to students (for safety’s sake). I suggest a third option, which as I put it, is “to use AI as a way of honing students’ skills and knowledge, helping them learn more effectively than before.”
Smith is dead-set against this option. He responded to it at MindMatters.ai in a three-part series here, here, and here. Now, in his defense, I wasn’t as clear as I could have been about the specifics of how I’m imagining the use of LLMs and AI in education. Nor did I distinguish the use of LLMs from AI generally. As I note later in this piece, there are places for AI to impact learning that do not employ LLMs.
Smith might thus understand my vision for the use of LLMs/AI in education as sticking students in front of a screen with ChatGPT, say, and simply telling them “now go learn something,” with an adult/teacher occasionally looking over the students’ shoulders and monitoring their activities to make sure they don’t get into too much trouble.
Such an understanding of my vision for the role of AI in education couldn’t be farther from the truth. I want therefore here to elaborate on some of the particulars for AI in education that I have in mind. Before doing so, however, I want to respond to some concerns and criticisms by Smith about my earlier essay, such as can be addressed on their own terms.
Redressing AI’s Seductiveness
It seems that Smith’s greatest concern about digital technologies as created Big Tech is that they are designed to suck in users, getting them to spend as much time and devoting as much attention as possible to these technologies—thus wasting their time, sapping their spirits, and keeping them from their higher good. Smith is largely right about this, as about the resulting human toll.
Smith is doubly concerned about LLMs in this regard because, as he sees it, they are designed to become conversation partners, even therapists and spiritual guides. An irony here is that in the 1960s, when artificial intelligence wasn’t worthy of the name, Joseph Weizenbaum’s ELIZA program, crude as it was in terms of the underlying technology, did fool people into thinking it was a therapist.
That program was essentially a grammatical rewriting tool that took statements by humans and rephrased them in the style of a “Rogerian therapist” (after Carl Rogers). It would mirror back what the patient said in a non-judgmental non-directive way by grammatically manipulating what a human typed at a keyboard, and it offered canned expressions to intersperse into the conversation when the program ran out of things to say.
Thus, in response to “I feel blue today,” ELIZA might respond, “So you feel blue today. Tell me more.” The sentence “So you feel blue today” would be a simple grammatical rewrite of “I feel blue today,” adding the word “so” and then changing the first to the second person. The sentence “Tell me more” would then be a canned expression to keep the conversation going.
When I sat in on an AI course in 1981 taught by Laurent Siklossy, a well-known AI professor at the time and something of a wag (LISP was the AI programming language of choice at the time, and his book on the subject was titled Let’s Talk LISP—get it?), one of his assignments was for us to write an ELIZA program. He made fun of it acting as a therapist and said that among the canned expressions we should have ready to go, especially when the program didn’t know what to say, was “I understand.” The irony was not lost on us students that this program could have nothing like real understanding.
Today’s LLMs are, of course, far more sophisticated and powerful than the early ELIZA programs. They don’t just reflect back to the human what the human is saying. They can hold a full-fledged conversation and, in effect, pass the Turing test (i.e., fooling people into believing that they are dealing with a human at the other end of the line if we didn’t know that we were dealing with a chatbot).
I share Smith’s concerns about LLMs serving as conversation partners for humans. With LLMs, there’s no person at the other end of the line. The top two commandments are to love God and to love other humans, and that requires connections with God and humans—the real God and real humans. Connecting to machines, as one would with other humans, is perverse and, in my view, soul destroying.
But just because humans engage in such behaviors with machines—and even if Sam Altman at Open AI wants to use ChapGPT to create perverse conversational dependencies between humans and machines—that doesn’t mean there aren’t legitimate uses of LLMs or AI in general. Nor that they can’t be suitably restricted.
Phones can be equipped with shields that enable parents to keep their kids from, via their smartphones, accessing pornography and other negative material that parents don’t want them to access. Parents have this right to prevention in virtue of paying for their kids to use the phones and, more importantly, in virtue of being their parents, guardians, and advocates.
There’s no reason that LLMs can’t have a third-party layer imposed on them that prevents intimate conversations and, indeed, exchanges on any topic that the party that legitimately controls access to the digital device (such as parents, teachers, and employers) chooses to restrict.
Technology as the Problem
Given Smith’s absolutism against LLMs/AI, I doubt he would find such a “fix” appealing. Smith’s anti-AI absolutism derives from certain deeper considerations that leave him completely opposed to the technology. It’s not that the technology can be abused. It’s that the technology is inherently problematic.
He cites with approval Neil Postman, Marshall McLuhan, and Jacques Ellul, all critics of how technology can warp content and mislead users. But that’s been a problem with technology from the get-go. When writing was introduced, people were concerned that this would undercut an oral culture, weakening memory and a shared universe of meaning.
With Gutenberg, pamphleteering became the rage, destroying the Catholic church’s monopoly on the written word. Newspaper, radio, and television technologies all introduced their own promises and perils. Postman’s Amusing Ourselves to Death wished for the complete eradication of television.
We face similar challenges with today’s AI technologies. It is a truism that all technologies can be used for ill or well (with the possible exception of certain weaponized technologies, as in biological or chemical warfare, whose only use seems ill).
With LLMs/AI, Smith would stuff the genie back in the bottle. But doing so is impracticable. Nor do I see a principled reason to do so. My own preference is to take a pragmatic line, namely, to accept the technology, harnessing it as much as possible for good and restricting but not barring it, especially in an educational context, from causing evident harm.
But can we know where these technologies do indeed cause harm, especially since we are unaware of long-term studies on their beneficial or ill effects? Smith cites what he regards as obvious ill effects, but I find none of them as deal-killers for the use of LLMs/AI.
The Escapable Dilemma
In his series of articles for MindMatters.ai where he responds to me, Smith writes, “Dr. Dembski’s article is filled with the assumption that AI chatbots like ChatGPT are trustworthy sources of information that students can depend on for learning.” But I’ve never written anything of the sort, nor do I privately think that.
I’ve found LLMs reliable for some tasks, less so for others. For rewriting and summarizing text, I’ve found ChatGPT, for instance, unproblematic. For providing references and quotes, I’ve found LLMs notoriously unreliable. Yet even when their output is unreliable, I’ve found LLMs can nonetheless be useful and even insightful. Not all mistakes are created equal. Some can be helpful.
Doug Smith, citing another Smith (Gary Smith), sees LLMs as collapsing on themselves because of an “inescapable dilemma,” which they take to be the following: “If you know the answer, you don’t need to ask an LLM and, if you don’t know the answer, you can’t trust an LLM.”
This is too simplistic. What if I don’t know the answer? Even then, the LLM may provide an answer that jars something in my memory confirming it is correct (recognition memory is always greater than recall memory). And if I can’t see the truth of the answer, I can do further checking to confirm it. Unlike the LLMs of 2022, the LLMs of 2026 have access to the web and can provide links to websites that confirm or disconfirm its claims. If the links are bogus, that’s often readily confirmed as well.
According to Doug Smith, LLMs don’t know anything because they are not agents that truly understand language in the sense of the classical correspondence theory of truth (to which I subscribe), according to which we as humans (and not algorithms) are capable of intuiting the match between linguistic statements and truths about the world.
Fair enough, LLMs are metaphysically challenged and don’t have knowledge in the same way humans do. They can say or otherwise output that snow is white, but they don’t know deep down what it really means for snow to be white.
But what of it? When I look up a word in a dictionary, does the dictionary know, in a deep metaphysical sense, what the word I’m looking up really means? No. The dictionary is just a book. It is not a knower in the sense that we are knowers. Its knowledge is derivative, depending on the knowledge of the lexicographers that produced it.
Likewise, LLMs don’t know in the sense that humans know. But they are trained on a vast corpus of human writings and are able to sift those writings, drawing connections among them, and providing insightful responses to queries based on that corpus.
Their knowledge is not metaphysical but instrumental. They are able to answer questions which, if posed to humans, would convince us that we are dealing with someone who knows what they are talking about (leaving aside when they hallucinate, or make stuff up, though humans have that failing as well).
Smith’s “inescapable dilemma” is therefore readily escapable. LLMs can provide responses to queries that offer genuine insights. LLMs typically pick up on some signal in their training data that is worth considering and underscoring even if it does not rightly elicit instant credence.
The wrong answers of LLMs have some basis in their training data, and those errors can be interesting on their own terms. And if they’re uninsightful, they can be easily enough discarded. Sure, let’s remain skeptical of LLMs. Let’s treat their output with suspicion. Let’s always require verification. But shouldn’t we take such a skeptical line with anything and anyone that promises to respond to our queries?
When Google, in response to a search query, sends us to a website, do we simply accept that this website is the final word in response to the query? Obviously not. The apostle Paul writes in 1 Thessalonians 5:21 (ESV), “Test everything, hold fast what is good.” This is a conditional, not a blanket dismissal. There’s no reason Paul’s counsel can’t be applied to LLMs.
Querying an LLM is similar to querying an index in a book. With an index, you look up a term, and then go to the pages associated with the term. Some of those pages may give you the information you want about the term, some may not, and perhaps all do not. You, as the inquirer, will need to decide. So too with LLMs.
Now it might be argued that LLMs are more subtle than book indexes and designed to mislead naïve students into accepting their pronouncements as gospel. But since our focus here is on LLMs in education, isn’t that the role of teachers to show students how to handle the educational resources at their disposal, including LLMs?
Instead of tacitly telling students, “We think you are too immature to use this technology,” why not simply show them how to use it properly and responsibly? As it is, students as well as the rest of society will use this technology, so simply on pragmatic grounds it’s better to get familiar with how to use it well than to deny its use, as Smith seems to prefer.
If LLMs are as fraught with such danger, temptation, and misinformation as Smith claims, does it make more sense to shield students from them or to inoculate students to them, helping them to see through “the devices of the enemy”? In all of my apologetics efforts, I’ve always regarded inoculation are far more effective than isolation (see for instance Unapologetic Apologetics, an anthology I edited with Jay Richards).
Misled by Mathematics?
Perhaps the most curious argument Smith makes in his response to my essay on edification and enhancement is that my background as a mathematician makes it difficult for me to see the dangers inherent in LLMs. As he puts it:
I’ve had conversations with several brilliant math PhDs about AI chatbots. All of them have a hard time seeing the dangers of large language models (LLMs). It’s almost unthinkable to them that the #1 use-case of ChatGPT is therapy and companionship. To math geniuses, it seems like only a tiny minority of foolish people should be susceptible to being drawn in.
As a cross-disciplinary scholar who has also worked as a professional philosopher and theologian and who did his undergrad as well as three years as a grad student in experimental psychology, I don’t approach LLMs as a tech nerd who sees them as just big neural networks run by probabilities and linear algebra.
Quite the contrary. I’m sensitive to the abuses of this, and indeed any, technology. Also, I’m well aware that any problem for society with LLMs centers not on their underlying mathematics but on how the resulting technology is deployed based on content ingested into the LLMs.
I therefore reject that too much learning in mathematics somehow renders me incapable of seeing the dangers with LLMs but that Smith is somehow capable of seeing the dangers because he lacks mathematical knowledge that would otherwise confuse him. If selective amnesia suddenly wiped out all my knowledge of mathematics, would I be a better commentator on the proper place of LLMs in society?
Where AI Promises to Improve Education
Toward the end of my edification versus enhancement essay, I offer ten possibilities where AI could significantly improve education, offering new opportunities for learning where they are now largely non-existent. Smith pooh-poohs each of these. To get a flavor of my proposed possibilities as well as Smith’s reaction to them, consider the first three possibilities that I described:
- Accent and Pronunciation Refinement in Language Learning
Speech-recognition AI can analyze intonation, rhythm, and articulation down to the phoneme, offering individualized correction and visualized feedback, especially for non-native speakers who want to eliminate an accent. Students can practice repeatedly with voice models until their accent aligns with native fluency—an outcome rarely attempted, much less achieved, in traditional classrooms.
- Creative Writing with Rhetorical Precision
AI editors can train students to identify and employ rhetorical devices—analogy, parallelism, irony—and revise for tone, cadence, and argument strength. Instead of vague feedback, students get detailed stylistic analysis mapped to classical rhetorical methods. Classical schools would likely be early adopters. - Polyphonic Music Composition and Performance
AI-guided keyboard and ear-training software can help students not just play Bach’s music but think contrapuntally. Thus they will learn voice-leading, harmony, and music theory, not in the abstract but in real time as they compose or perform. Feedback systems can identify harmonic tension, suggest corrections, and let students iterate toward mastery of multi-voice textures.
Here is Smith’s response to each of these:
- “Accent and Pronunciation Refinement in Language Learning”
A focused speech-recognition “AI” (not a chatbot) could help students learn languages more efficiently. That tech has been around for a while. A very narrowly, ethically trained product could be useful here, but it’s no replacement for human conversation.
- “Creative Writing with Rhetorical Precision”
This is a dangerous proposal. Writing is thinking. AI can’t be trusted to generate coherent content, nor to have a voice that is something students should emulate. Do we really want everyone homogenized into writing like a chatbot? No, the sacred work of choosing words must have authoritative human guides. Improve the curriculum, have better books, and invite humans to teach via video. But don’t use chatbots for creative writing.
- “Polyphonic Music Composition and Performance”
I see this as risky. Again, if the AI is not a general purpose chatbot, and perhaps specifically trained on good music, it could potentially be helpful. But today’s AI chatbot-driven song generators are not creating well trained musicians. They’re creating more passive consumers — by design. Chatbot users aren’t interested in learning to “think contrapuntally.” They want quick and easy dopamine hits, which is what chatbots are designed to provide.
Accent and Pronunciation Refinement
Smith’s responses here are emblematic of a general resistance by him against AI. Let’s go through his three criticisms in turn. With regard to accent and pronunciation refinement in language learning, Smith allows that this could be harmless if “narrowly, ethically trained.” But he still sees it as inferior to human conversation.
But Smith’s recommendation here seems misguided. Children who acquire a second language before roughly the age of ten can typically do so with no accent, but after that, the accent endures, and indeed becomes entrenched, unless it is very methodically addressed. Conversation as such is no help here.
A friend of mine from Cuba arrived here in the US at the age of 12. He’s had plenty of English conversation over the years, and yet his accent is as thick as ever. My German mom, now passed on, married my dad in her mid-20s and then moved to the US with him. She never lost her German accent.
What I’m imagining is a neural-net-based program that prompts people with foreign accents to increasingly approximate proper English enunciation. This would not be an LLM. Such training could make a big difference for ESL (English as a Second Language) people wanting to improve their English and fit more readily into the wider American culture.
People who want to keep their accent can keep it. However, people would also have the option of (largely) shedding their accent later in life, which right now is not an option for most people (the exception perhaps being someone with access to a linguist who can specifically train them).
Improving Writing
With regard to the second point, an LLM adapted to assessing and helping improve the writing of users would, of course, need to be tested to determine whether it is indeed offering good writing advice and training. But I’m optimistic here.
When I think of my K-12 education and the often poor to non-existent feedback I got from teachers on my writing, I have a hard time imagining that a suitably trained LLM would do worse than what I experienced.
If during your high school you are fortunate enough to attend a top prep school, such as Phillips Exeter Academy, where an award-winning writer such as Frederick Buechner could offer comment on your class essays, good for you. In that case you may well do better than any feedback on your writing that an LLM can offer.
But if you’re in an inner-city high school where most students are reading way below grade level and the teachers have minimal writing skills (most K-12 teachers in the US are in the bottom quartile of their high-school classes), an LLM may be just what the doctor ordered (if you are able to take proper advantage of it).
Smith takes a principled stance that humans are to be preferred to LLMs in assessing writing quality and that somehow LLMs are bound to inculcate substandard homogenized writing. But that’s not been my experience. I can ask an LLM to rewrite a passage in the style of a well-known author, and, depending on the author, I get what I think is a reasonably good output (with ChatGPT, Ernest Hemingway seems to work well but James Joyce doesn’t).
While not accomplishing everything I’m imagining for this second point, an AI-based service I use goes at least some of the way toward achieving the goal of improving writing. Specifically, I know of a program that uses an LLM to copyedit articles and book chapters. It is available at Inkwell.net and I’ve used it to copyedit material on a book I’m shortly to be publishing (a biography of Jaime Escalante).
I was hesitant initially to use the tool, thinking that the book in question had been thoroughly copyedited. And yet, I found this copyediting tool as offering solid improvements on text I thought was already clean. As it is, I’m not just an author but an experienced copyeditor. I was impressed with what this program offered, and I don’t think I’m that easy to impress when it comes to writing quality.
Contrapuntal Training and Recovering Lost Arts
On the third point, Smith seems to think that what I have in mind here is to get an LLM to do music training, and specifically contrapuntal training. My initial impulse would not be to use an LLM here. A bespoke neural network, trained on contrapuntal music, would probably be the way to go, though one capable also of exporting to sheet music and mp3s as well as doing ear-training.
Smith’s resistance to such AI-guided contrapuntal training here seems entirely reflexive. We don’t have anything like contrapuntal training for the masses right now, such training being reserved for specialized music students. So the alternative is to use AI to bring about such training or go without it. Anti-AI absolutism thus would seem to encourage not an enriched education but a status-quo that simply resigns itself to doing without contrapuntal training. Anti-AI absolutism thus seems to me a recipe for impoverishing education.
How many activities of the past—how many lost arts and skills—might be recaptured through AI? It boggles the mind to think that ordinary people could visit the theater in Shakespeare’s time, watch and listen to his plays, and be entertained with no special training or advantages.
Yet nowadays reading Shakespeare is often a strain and hardship. For many high-school and even college teachers, getting students to appreciate Shakespeare’s towering genius is like pulling teeth. AI, it seems, could be used to facilitate easy comprehension of Elizabethan English and thus make experiencing a Shakespeare play not a burden but an enjoyment.
More Powerful Than Traditional Search
Every technology entails both promises and perils. Yet, as Aristotle taught, virtue is found at the mean. Courage, for instance, lies between cowardice on the one hand and foolhardiness on the other. So with AI, virtue is not found in its uncritical embrace, as with Sam Altman, or in its reflexive rejection, as with Doug Smith. Discernment is needed, which is always eclectic, choosing the good and rejecting the bad.
I agree wholeheartedly with Smith that forming intimate bonds with chatbots is perverse and should be discouraged. However, I’ve personally found this technology to be quite helpful. For instance, I can perform much more sophisticated searches with ChatGPT or Grok than I can with Google.
Let’s say I’m thinking of someone but can’t recall his name. I know one or two salient facts about the person. Perhaps one of the things that I think is a fact about this person is not a fact but close to it. For example, I tell the LLM that so-and-so wrote a book on hell, but actually the person I’m thinking of wrote about heaven. I find that the LLM can often find what I’m thinking of even with such imperfect descriptions.
It might now be argued by employing this AI capability, I’m atrophying my memory by letting a chatbot conjure up information for me that, without the chatbot, I would have to think harder to recall. But what if thinking harder would not recall it? I find LLMs often being able to bring to light things that are sketchy in my mind.
Strengths and Weaknesses of LLMs
What about hallucinations? Yes, LLMs make stuff up. It’s in their nature to make stuff up, especially when their knowledge base is deficient and they are being asked leading questions. ChatGPT, for instance, has yet to ingest the content of many books. So when I ask it to provide the chapter titles for such an uningested book, it will tend to make them up.
I asked ChatGPT, for instance, for the chapter titles of Doug Smith’s book [Un]Intentional. It not only got the number of chapters wrong, but also each of the chapter titles wrong. I was unsurprised by this. I’ve worked enough with LLMs to know where they are likely to fail.
I’ve also found ChatGPT to make up references when I ask it for a reference by an author to a work with certain features, even though I know independently the author never published anything like it. Ask an LLM leading questions, and you’ll get some response. Usually the response will be stated with confidence even when the LLM is completely at sea.
Despite all this, I don’t see these weaknesses as a deal-killer against using LLMs. These systems have their limitations, and part of education will be to train students in what these are.
Where I find LLMs to shine is in summarizing work and in rewriting sentences and paragraphs. Say I’m faced with a long op-ed where the author is doing a lot of stage-setting and taking his sweet time getting into the gist of the article. I may feed it into an LLM and ask it for a one- or two-paragraph summary. If not much is riding for me on this article, that is often enough.
Or suppose there’s a problematic sentence in a book or article I’m writing. The sentence sounds clunky to me. I’ve played with rewording it, but still can’t get it to sound decent. Just as I use a thesaurus to find synonyms for words, I can ask an LLM to rewrite the sentence in several ways, essentially finding “synonyms” for the problematic sentence I’ve submitted to it.
By engaging in this exercise, I’ll often find some rewording that seems to me better than what I wrote. Typically, I don’t just take one of the proposed sentences and merely substitute it for the clunky one. Rather, I’ll get some rewording ideas that help me clean up my sentence.
Am I becoming a poorer writer by relying on an LLM in this way? I don’t think so. If anything, I get ideas for new ways to rephrase things that break out of my accustomed writing patterns. My attitude toward LLMs is that they are servants, and I use them very intentionally to help me on specific tasks.
Now Smith might say that I’m simply being seduced by this technology, not realizing what it is doing to me. But in that case, we’re just going to have to differ. I can look better into my own soul than any other human being, and I’m simply not seeing that the quality of my internal life or of my work is degenerating because I’m using LLMs. Sure, I may be deluding myself. But then again critics who see LLMs as wholly negative may also be deluding themselves. Time will tell.
The Need for Accountability
All that being said, I don’t want to minimize the peril of using LLMs. If used uncritically in education, they can do harm. Leaving aside intimate conversations with a chatbot, the peril of LLMs in education is that students will rely on them as a crutch. It’s here that we always need human teachers holding human students accountable so that they can stand on their own two feet without the prop of an LLM.
I’m less concerned about what students do on their own time with LLMs (though the thought of an LLM as someone’s “spiritual director” seems to me perverse and anti-human). But in the classroom, there need to be times when students do writing assignments without access to an LLM; when they are tested on basic facts in geography or history without access to an LLM; when they solve mathematical problems with pencil and paper and nothing else; etc.
The recent revolution in LLMs and AI is indeed a revolution. I sat in on my first AI course in 1981 as a first-year grad student in psychology (taught by Laurent Siklossy, as described earlier). I’ve followed the field over the years. Problems that seemed insurmountable, such as passing Winograd schema tests (finding proper pronoun antecedents based purely on semantics rather than syntactics) or passing the Turing test (a machine being able to produce output over a communication channel largely indistinguishable from a human), have now in the last few years finally been surmounted.
What does all this mean for education? I’m willing to give LLMs, and AI more generally, opportunities to succeed in advancing education. I believe the technology needs to be tightly monitored in the educational arena, and that education always needs to devote significant amounts of time to demonstrating knowledge and skill without the prop of machines.
Will even such a constrained use of AI technology end up being counterproductive? Will students unexposed to AI technology during their K-12 years, for instance, end up being better learners and well-rounded human beings than those exposed to it in a controlled and disciplined way? I suspect Doug Smith would opt for the former. I opt for the latter.
That said, I don’t think we need to wait a decade to determine which option is superior. These technologies are being tried in real time, and we should be able to see in real time how students are faring with them. But note, I’m not advocating a complete laissez-faire approach to using AI in education. It needs to be used deliberately, and at times it needs to be firmly excluded so that we don’t become its dependents.
AI in Monitoring Attention
On one point in my edification and enhancement essay, I want to make a concession to Smith. I mention there that technology could be used to scan brain activity of students to determine their degree of engagement in the classroom. I raised this as a speculative possibility (though for something unobtrusive, like a ring around a finger, rather than a headband). Smith helpfully pointed out in his response to me that this has been tried in China, as noted by the Wall Street Journal in the following short video:
To this Smith adds:
[Dembski’s suggestion for attention monitoring in my earlier article] mirrors the ultra-authoritative Chinese state’s use of AI monitoring for students. In China, students are already wearing AI-powered headbands to constantly monitor student’s [sic] brain activity. That data will almost certainly be part of their “social credit score,” among other nefarious things. The video in that Wall Street Journal article is not a dystopian sci-fi story, it’s today’s reality. Those are real kids wearing those devices and being shaped, controlled, and dehumanized.
It may well be that the very intrusiveness of this technology will undermine learning. At this point, we can’t know for sure. In any case, China’s example doesn’t answer whether this technology may prove useful for learning. It seems that EEG headbands and related engagement-tracking tools depicted in the above video may not improve learning outcomes. EEG readings in uncontrolled environments—especially noisy, movement-filled classrooms—are apt to be unreliable and prone to false signals.
What in the Chinese school system these devices seem to be measuring is not concentration but behavioral stillness—the surveillance itself pressures students to suppress movement and to appear attentive (this video even describes a student whose parents beat him when his attention scores dip). This produces compliance, not understanding. Genuine learning requires curiosity, free play, autonomy, and psychological safety—conditions undermined by constant monitoring and fear of parental, teacher, or state punishment tied to attention scores.
Teachers may see short-term improvements in orderliness because of this technology. But such technologies seem unlikely to increase comprehension, retention, or intrinsic motivation. Instead, they will likely introduce stress, reduce working memory, and foster extrinsic pressure that weakens deep learning. The function of the Chinese system seems not pedagogy but data collection and behavioral conditioning, feeding government-backed AI research projects and normalizing surveillance for children. Simply put, the technology shapes behavior but not understanding—and likely harms the very cognitive processes it claims to enhance.
But what if students and teachers alone were given access to the data from such a device? What if students didn’t have to wear the device all the time, but only at certain times when their level of engagement with class activity was clearly flagging? What if certain patterns could be spotted, where certain types of instruction or certain teachers proved more or less effective at being correlated with brain activity indicative of learning? What if, to distance this technology from the daily classroom setting, such a device were solely used in research settings to determine the effectiveness of various types of pedagogy?
To the degree that this technology is diagnostic and not for “shaping, controlling, and dehumanizing,” not for surveillance, and not for going on a child’s permanent record but only for temporary use, I could see it being helpful. Or perhaps not. Perhaps whatever these devices are measuring will never be sufficiently diagnostic of true learning. Perhaps all that’s needed is an attentive teacher who can track which students are engaged and which are zoned out.
Whether my proposal for tracking student engagement via physiological markers is ultimately for the good will depend on whether such markers truly measure engagement (rather than mere compliance). It will also depend on how the associated technology is implemented. Most importantly, it will depend on whether students do indeed become better learners through this technology because it correctly tracks when they are engaged in class and identifies what facilitates and hinders their learning. My approach here is principled only in respect of this technology not being used to enforce compliance. It is practical in the sense that it seems worth trying and worth implementing only if found to be helpful in helping students become better learners.
Primacy of the Curriculum
In closing, I want to make one final point, and it is this: Any use of LLMs and AI in education must be in the service of a curriculum that clearly specifies what is to be learned and tracks whether students are indeed learning what the curriculum so specifies.
I would regard it as educational malpractice to simply stick a student in front of ChatGPT and say, “Now have at it and learn something.” I imagine an AI-assisted educational curriculum will take the form of programmed learning (in the sense developed by psychologist Edward Thorndike over a century ago) where rather than passively absorbing content, students are required to actively engage the content and correctly answer queries about it.
Specifically, Thorndike offered the following vision of programmed learning:
If, by a miracle of mechanical ingenuity, a book could be so arranged that only to him who had done what was directed on page one would page two become visible, and so on, much that now requires personal instruction could be managed by print.
LLMs and AI more generally promise to make Thorndike’s vision of programmed learning, or programmed instruction, not just a reality but also to expand it far more powerfully than he could have imagined. (By the way, I happen to own the domain names programmedlearning.com and programmedinstruction.com, which I’m happy to make available for pursuing such a vision of education.)
In any case, to return to the importance of curriculum in any AI-assisted educational effort, much of the material in such a curriculum will need to be clearly delineated in advance. In particular, it must not be made up seat-of-the-pants as one goes along. To the degree that this material is generated by an LLM, an educator will need to verify and edit it. That said, I can see an LLM in real time generating problems whose correct answer will test the degree of mastery in a given area of study.
For a simple example, with which I close, suppose you want to use ChatGPT to test and build vocabulary. Consider, therefore, the following prompt to ChatGPT: “Provide a multiple choice vocabulary test of ten questions with four possible answers to each question, marking each correct answer, and where the test lays out words of general interest at the beginning graduate student level (17th grade).”
Below is ChatGPT’s response to this prompt. Note that in a real vocabulary test, the correct answer would not be marked, as it is here, but would be confirmed after the student offered his or her own answer. ChatGPT is well able to handle bookkeeping of this sort. Obviously, this example can be vastly generalized to math, geography, history, and any other subjects.
People may question, in the case at hand, whether the vocabulary words below are truly graduate-level words. I’m less concerned about the particulars here as about a general proof of concept. In my view, students learn best when they are active participants (hence programmed learning, as above), needing to perform actions that demonstrate knowledge, skill, or competence in real time. AI, suitably primed and predisposed, seems to me ideally suited for eliciting and evaluating such actions and thus for determining whether effective learning is indeed taking place.
Graduate-Level Vocabulary Assessment (10 Questions)
1. Abstruse most nearly means:
A. Easily understood
B. Difficult to comprehend ✔️
C. Pleasantly surprising
D. Excessively ornate
2. Inchoate refers to something that is:
A. Fully developed
B. Not yet fully formed; just beginning ✔️
C. Excessively complex
D. Devoid of structure
3. Pernicious describes something:
A. Gradually improving
B. Highly destructive or harmful ✔️
C. Temporarily confusing
D. Intensely bright
4. Equanimity denotes:
A. Creative insight
B. Moral virtue
C. Calmness under stress; composure ✔️
D. A tendency toward self-doubt
5. To excoriate someone is to:
A. Praise them warmly
B. Criticize them severely ✔️
C. Remove them from office
D. Warn them indirectly
6. Recalcitrant most nearly means:
A. Fragile or easily broken
B. Stubbornly resistant to authority or control ✔️
C. Pleasantly compliant
D. Awkwardly positioned
7. Sedulous effort is:
A. Half-hearted
B. Frenetic and impulsive
C. Diligent, persistent, and careful ✔️
D. Unfocused and meandering
8. Vicissitude refers to:
A. A moral failing
B. A sudden revelation
C. A change or variation, often unexpected ✔️
D. A cyclical ritual
9. Assiduous attention involves:
A. Sporadic interest
B. Constant, careful, and persistent effort ✔️
C. Distracted focus
D. Passive observation
10. Perspicacious most nearly describes someone who is:
A. Easily discouraged
B. Insightful and keen in mental perception ✔️
C. Verbose and rambling
D. Emotionally impulsive
