Why ChatGPT for Teachers Might Make Things Worse

A critique of a disappointing model and a vision for what AI Platforms should do instead

OpenAI have just released ChatGPT for Teachers which I have just spent the morning looking at it and I have concerns. Here is an initial reaction to it and 3 reasons why I think it’s a huge missed opportunity.

---

There’s a predictable pattern when some edtech companies enter education: they build sophisticated solutions to problems that don’t exist while ignoring the problems real teachers actually face. I am cautiously optimistic about what AI could offer education, particularly around the idea of teacher workload but this is not it, and I’m sure the people who developed ChatGPT for teachers are well-intentioned but what we have here is a step backwards which actively promotes discredited theories and pedagogically bankrupt approaches.

1. “Activity First” Lesson Design

On the explore page the very first ‘tip’ promises the ability to “find activity in seconds.”

The problem with format-first pedagogy and instructional design is that it fundamentally inverts the relationship between knowledge and pedagogy. Instead of asking “what knowledge structure demands this particular pedagogical approach?”, it asks “what activities can I fit this content into?” It’s a completely superficial approach to learning and instruction.

“Activities-first” design treats curriculum as a collection of discrete events rather than a coherent narrative. It sees curriculum design as forward-engineering rather than backward design. When teachers begin with activity templates, ie think-pair-share for this part of the lesson, retrieval practice for that one, a guided practice here etc, they’re assembling a sequence of disconnected episodes rather than constructing a journey through a discipline. As Christine Counsell memorably puts it “You cannot Rosenshine your way into a curriculum, and nor can you Rosenshine your way into mediating content in subject‑sensitive ways.”¹

The deeper problem is that activity-first design reduces knowledge to mere propositions and procedures. It asks: “What do I want students to know or be able to do?” and then selects activities to achieve those outcomes. But much of the time in humanities, arts, and literature, one is not teaching procedures or isolated propositions. McCarthy and Minsky call these “ill-defined domains” (which actually have a long history with AI going back to the 50s) and there, one is teaching knowledge that exists in relationships, structures, and flows that cannot be packaged into discrete “learning objectives” without destroying what makes them educationally valuable.

The efficiency promised by ChatGPT to ”find activity in seconds” is efficiency in exactly the wrong dimension. Yes, you can quickly generate a lesson structure. But you’ve saved time on the professional work that distinguishes effective from ineffective instruction: thinking deeply about the nature of the knowledge you’re teaching, how it connects to what came before and what comes next, what disciplinary features demand particular pedagogical approaches, and how to structure encounters with content so that students experience both the intrinsic value of the moment and the gathering momentum of a larger narrative.

Pedagogy should follow epistemology, not the other way round.

2. Generic Critical Thinking Skills

Another suggestion comes from a real teacher advocating using chatGPT to teach critical thinking. This is a familiar trope in education that is fraught with problems.

The notion that students can be taught “general critical thinking skills” to evaluate AI outputs represents a profound misunderstanding of how critical thinking actually works. Critical thinking is not a portable skill that operates independently of domain knowledge, it is the product of domain expertise.

Link

You cannot “think critically” about whether ChatGPT’s answer about mitosis is credible unless you already possess substantial knowledge of cellular biology. You cannot evaluate the plausibility of an historical claim without understanding the period, the historiography, and the nature of historical evidence. The ability to spot implausibility, recognize internal contradictions, or judge whether a source should be trusted emerges from having rich, well-structured schemas in long-term memory within that specific domain.

What this is describing (advising students to verify information through “other primary resources”) presupposes the very expertise it claims to develop. How does a novice know which sources are primary? How do they judge between conflicting accounts? How do they recognize when ChatGPT has produced plausible-sounding nonsense versus a genuine insight? Without substantial prior knowledge, students are left either credulously accepting whatever sounds convincing or adopting a pose of generic skepticism that rejects valid and invalid claims indiscriminately.

“grammar skills”

The cruel irony is that this approach, marketed as empowering students to navigate AI-saturated information environments, actually leaves them more vulnerable by substituting the appearance of critical evaluation for its substance.

3. The Myth of Learning Styles Now Powered by AI

If there exists low-hanging fruit in the orchard of educational neuromyths, then learning styles is surely it; the claim that students learn better when instruction matches their preferred modality (visual, auditory, kinesthetic) has been so thoroughly debunked that its persistence serves as a litmus test for whether someone knows what they’re talking about when they talk about learning.

On another page, we are told that ChatGPT can personalise content for “different levels and learning styles.” This is the most disappointing aspect of ChatGPT for teachers.

The fact that OpenAI, a company positioning itself as an authority on the future of education, appears entirely unaware of decades of research showing the theory of learning styles has no empirical support is deeply concerning. More troubling still is that teachers using this product will now receive algorithmically-generated affirmation of a discredited theory, entrenching precisely the kind of pseudoscientific thinking that evidence-informed practice seeks to displace.

Time spent diagnosing spurious learning styles, differentiating instruction according to imaginary preferences, and reassuring students they are “visual learners” is time stolen from practices that actually matter.

When teachers are encouraged to match instructional modality to supposed learner preferences rather than to the inherent demands of the content itself (some concepts simply require visual representation; others demand narrative; still others need procedural modeling), they are being led away from attending to knowledge. The cruel irony is that this is sold as personalisation when it is, in fact, the abdication of the professional judgment that genuine responsive teaching requires; the kind that emerges from understanding both the learner and the disciplinary object, not from sorting students into pseudoscientific categories.

For a long-read on why Learning Styles theory should be avoided, read this.

A Wasted Opportunity

Ultimately I am cautiously optimistic about what AI can do for instructional design but this product is very disappointing and has raised my AI-educational p-doom. What emerges from this is an uncanny valley of educational materials: content that looks professional, sounds educational, and checks boxes for “relevance” while being ultimately unmoored from the science of how learning happens.

The other broader worry is the way in which I sense that this kind of thing will lead to a kind of flattening out of subject specific meaning, disinterred from local knowledge and expertise. LLMs work by predicting the most likely next word. They are engines of regression to the mean. They cannot produce culturally responsive or locally relevant materials because they are trained on the “global average.” A lot more to be said on this point.

Implementing Adaptive Learning With AI

Any learning platform looking to implement AI should be doing almost the precise opposite of ChatGPT for teachers.

Here’s where we must be precise about the nature of knowledge and where AI could genuinely help: some knowledge structures are genuinely componential and benefit from systematic decomposition, while others are irreducibly holistic and lose meaning when fragmented. The challenge is that ChatGPT for Teachers treats all knowledge as if it belongs to the first category when distinguishing between these types is the essence of instructional design and expertise.

Consider phonics instruction, where decomposition is not just appropriate but essential. The alphabetic code genuinely can be broken into grapheme-phoneme correspondences that students must master systematically. The knowledge structure itself is componential; mastery of /sh/, /ch/, and /th/ sounds contributes independently to reading capability.

The fatal error is applying this decomposition logic indiscriminately. ChatGPT for Teachers seems to make no distinction between knowledge that is genuinely atomic (phonics, multiplication facts, chemical symbols) and knowledge that derives meaning from its disciplinary texture and flow (historical narrative, literary interpretation, philosophical argument).

When you decompose a historical period into disconnected “facts” about dates and events, you lose the causal relationships, the narrative coherence. When you break a poem into atomic elements like “metaphor,” “imagery,” and “tone,” you end up feature-spotting and destroy the aesthetic encounter that gives literature its power.

We know from decades of research that retrieval practice, spaced repetition, and interleaving produce robust learning gains. We know that there’s an optimal difficulty zone. We know that spacing intervals should dynamically adjust based on individual mastery. We know that students need practice distributed across weeks and months, not crammed into single units.

The problem isn’t knowing these principles work; it’s implementing them with precision for thirty students simultaneously across multiple topics over an entire academic year. This is exactly the kind of coordination puzzle at which AI excels and humans struggle. An adaptive learning system *could* co-ordinate elements of this and provide teachers with actionable data about which students need intervention on which content.

The issue isn’t just that ChatGPT for Teachers solves the wrong problems; it’s that in doing so, it may make the right problems (such as implementing what we actually know about how learning happens), even harder to solve.

1. From a forthcoming Civitas essay