Pedagogical Intelligence: What AI Needs to Borrow

A couple of years ago, when we first started wiring large language models into a course builder, the early results were mesmerising. We fed a prototype a dense, sixty-page policy document on biodiversity net gain. In under a minute, it produced a five-module course outline, complete with introductory text, key definitions, and a summary for each section. The English was perfect. The structure was logical. It looked, to all intents and purposes, like a perfectly serviceable e-learning course.

And it was useless.

Factually impeccable, it was useless because it didn't teach. It only presented. It summarised the document, but it had no idea how to take a novice through it, how to build their understanding from the ground up, how to check they'd grasped one concept before introducing the next, or how to test if they could apply the policy in a real-world scenario.

It had linguistic intelligence and subject-matter expertise, derived from the source text. But it had zero pedagogical intelligence. And that, we discovered, was the entire ball game.

Pedagogical intelligence is the gap between an AI that can write about a topic and an AI that can scaffold learning around it. It’s the invisible architecture of good teaching, the craft that separates words on a screen from an experience that genuinely builds competence. It’s the thing that knows a course isn't a document with a ‘next’ button.

This isn't an academic distinction. It’s the core design problem for anyone trying to build tools that help people learn, rather than just tools that help people write.

The Anatomy of an Explanation

At its simplest, pedagogical intelligence is about how you sequence and frame information. An LLM, left to its own devices, will often give you a perfect, Wikipedia-style summary. It’s comprehensive, accurate, and structured for reference, not for learning.

A human expert with pedagogical sense does something different. They don't start with the minutiae. They start with a hook, an anchor point in the learner's existing world. Let's say the topic is the GDPR.

A purely linguistic model might start with Article 5, listing the seven principles of data processing. Correct, but overwhelming. A pedagogically intelligent approach might start with a question: “Have you ever wondered why you suddenly have to tick a dozen boxes before you can read a news article? That's what we're going to unpack.”

It’s the difference between a textbook and a conversation. The textbook lists the facts. The conversation builds the mental model. It uses analogy: “Think of personal data like your house keys. You wouldn't just hand them out to anyone who asks, and if you do, you'd want to know why they need them and what they'll do with them.” It builds from the core concept outwards, from the relatable to the abstract. It checks for understanding along the way, asking small questions before moving on.

This isn't about ‘dumbing down’. It’s about respecting cognition. It’s the craft of knowing that understanding isn’t a single event, but a process of assembly. Providing the raw materials, however well-written, isn't enough. You have to provide the assembly instructions, too. That’s pedagogy.

Beyond Recall: The Assessment Gap

Nowhere is the gap more obvious than in assessment. An AI can generate a factual quiz with its eyes closed. “What year was the Bribery Act passed?” “List three of the six lawful bases for processing data.” This is the lowest level of Bloom's taxonomy: recall. It tests if someone was paying attention, not if they understood.

For any serious professional learning, this is next to useless. You don’t need your people to be able to recite the Bribery Act; you need them to be able to spot a potential facilitation payment and know what to do about it. You need them to apply their knowledge.

And this is where a generic AI falls down. To test application, you need to create context. You need scenarios. Not generic ones, but scenarios with just enough specific detail to feel real and just enough ambiguity to force a decision.

Consider this:

• Recall Question: What is a ‘conflict of interest’?
• Application Question: “You are a procurement manager. Your cousin's company, a firm you know does good work, has submitted a tender for a contract you oversee. It is the second-best bid on price, but you believe their quality is higher than the cheapest bidder. What are your immediate obligations under the company's conflicts of interest policy?”

The first question can be answered by any AI that has read the definition. The second requires an entirely different level of intelligence. It requires the AI to construct a plausible world, a dilemma, and a prompt that forces the learner to synthesise the policy with a real situation. It has to understand the intent of the training, not just its content. It has to know that the goal is not knowledge, but judgement.

This is what pedagogical intelligence builds. It prioritises questions of application, of analysis, of evaluation. It designs for competence, not just for comprehension.

Can You Train an AI in Taste?

So can you teach an AI to do this? Can you give it a library of instructional design theory and tell it to be more pedagogical?

Not really. It’s like trying to teach a person to have good taste by making them read books about it. They might learn the vocabulary, but they won’t acquire the instinct. The problem is that pedagogy isn't just a set of rules; it's a thousand small, context-dependent judgements.

It seems to me that the only way to embed this intelligence is not to try and make the AI itself a master pedagogue, but to build the pedagogy into the architecture of the system that uses the AI. The craft isn't in the LLM; it's in the careful construction of the prompts, the workflows, the content templates, and the user-facing choices that guide the output.

You don’t ask a generic chat model to “write me a course”. You build a system that says, “First, we will define clear, measurable learning objectives. Then, for each objective, we will generate an explanation, an example, and a non-gradable knowledge check. Then we will create a scenario-based question to test application.”

The AI becomes a phenomenally powerful and fluent engine for filling in the blanks, but the structure—the pedagogical intelligence—is provided by the tool itself. The system provides the bones, the skeleton of good practice, and the AI provides the muscle.

This is a harder and less glamorous task than just plugging into an API and being dazzled by the fluency. It means having an opinion. It means codifying a specific method, born of experience, about how learning works.

The real work isn’t just making the AI fluent; it’s making the systems that use it wise.