When AI Teaches Differently: Designing Generative AI for Inclusive Learning

Artificial intelligence has entered education with a particularly ambitious promise: personalization.

Unlike traditional digital learning systems, generative AI can dynamically change the way information is presented. It can simplify an explanation, provide another example, break a complex concept into smaller steps, generate practice questions, or respond to a learner's specific question in real time.

For many learners, this already represents a significant shift in how educational technology can support learning.

But for learners with learning difficulties, the question is more complicated.

A system that can personalise learning can also personalise failure. 😶

If a generative AI system misunderstands a learner's needs, produces misleading information, provides inappropriate scaffolding, or encourages the learner to stop thinking independently, its adaptability becomes a potential source of harm rather than inclusion.

This leads to a more important research question:

How can we design generative AI systems that adapt to learners without compromising their learning, autonomy, and long-term development?

From Personalisation to Adaptive Intelligence

Personalised learning is not a new concept.

Educational systems have been attempting to adapt instruction to individual learners for decades. What generative AI changes is the granularity and flexibility of adaptation.

A conventional learning system might determine that a student has answered several questions incorrectly and recommend easier exercises.

A generative tutor can potentially go further.

It might recognise that the learner is struggling with a particular concept and generate a new explanation using simpler language. It could provide an analogy, break a task into intermediate steps, change the amount of scaffolding, or ask the learner to explain their reasoning before revealing an answer.

However, there is an important distinction between generating different content and understanding why a learner needs different content.

A learner who repeatedly makes mistakes does not necessarily need a simpler explanation.

The underlying difficulty could involve working memory, attention, prior knowledge, processing speed, language comprehension, misconceptions, or simply an ineffective learning strategy.

Therefore, genuine personalisation requires more than prompting a large language model with: "Explain this more simply." It requires some representation of the learner.

This suggests an architecture in which the generative model is only one component of a larger adaptive learning system.

The system may need to maintain a learner model containing information such as:

• prior knowledge;
• recurring error patterns;
• preferred forms of representation;
• level of scaffolding required;
• response time and interaction patterns;
• progression across concepts;
• demonstrated misconceptions; and
• changes in performance over time.

The challenge is that such a learner model must itself be treated cautiously. A model's prediction about a learner is not necessarily a fact about that learner. That distinction is particularly important when working with neurodiverse learners and learners with learning difficulties.

The Hallucination Problem Becomes an Educational Problem

Generative AI systems are probabilistic.

They can produce fluent and convincing responses without possessing a guarantee that those responses are correct. In a general productivity application, an incorrect response may be inconvenient. In education, the consequences can be considerably more serious. Consider a learner who asks an AI tutor:

"Why does this algorithm work?"

The system produces an articulate explanation containing a subtle conceptual error. A confident learner might question the response. Another learner might not. This creates an uncomfortable asymmetry: the learners who may need the most support can also be among those who are least equipped to identify when that support is wrong. Consequently, hallucination mitigation should not simply be treated as a technical accuracy problem.

The relevant question is:

Can the system prevent incorrect information from becoming a learned misconception?

This changes how we evaluate generative tutoring systems. Traditional language-model evaluation might measure factual accuracy or benchmark performance. An educational AI system should additionally be evaluated through outcomes such as:

• whether learners detect incorrect information;
• whether misconceptions persist after AI interaction;
• whether learners can correct an AI-generated error;
• whether confidence changes after receiving incorrect feedback; and
• whether the learner can independently solve a similar problem later.

A technically impressive model can therefore still be a poor tutor.

Reliability Is More Than Accuracy

Another problem receives considerably less attention: longitudinal reliability. Many AI education prototypes are evaluated through relatively short experiments. A student interacts with the system for an hour, completes a post-test, and researchers measure learning gain. But education does not happen in an hour.

For learners who require sustained support, the system may be used repeatedly over months or even years. This introduces a different class of problems.

• The learner changes.
• The curriculum changes.
• The learner's knowledge changes.
• The model changes.
• The interaction history becomes larger.
• And the distribution of questions presented to the system may gradually shift.

This is essentially a concept drift problem in an educational environment. A system that performs well during its initial evaluation may behave differently later. Therefore, future generative tutoring systems should be designed with mechanisms for monitoring performance over time. Possible approaches include:

1. Continuous evaluation of generated responses.
2. Drift detection for changes in learner behaviour and model performance.
3. Fallback mechanisms when the system becomes uncertain.
4. Human escalation when a question exceeds the system's reliability threshold.
5. Version monitoring when models or prompts are updated.
6. Longitudinal learner evaluation rather than relying exclusively on short-term tests.

In other words, an educational AI system should not simply ask:

"Did the model work?"

It should ask:

"Does the system continue to work safely for this learner over time?"

Inclusion Is Not the Same as Treating Everyone the Same

One of the most interesting challenges concerns bias. AI systems are often developed around assumptions about an "average" learner. But there is no single standard learner. A student may require additional processing time. Another may benefit from visual representations. Another may need instructions decomposed into smaller steps. Someone else may understand the concept but struggle with the way the question is formulated.

If a generative AI system assumes that faster responses, shorter explanations, or conventional interaction patterns indicate better learning, it can unintentionally disadvantage learners whose cognitive processes differ from those assumptions.

This is where inclusive AI needs to move beyond the idea of simply providing accessibility features. Inclusion should be embedded into the intelligence of the system itself. For example, rather than deciding that a learner is "weak" because they require more scaffolding, the system should treat scaffolding as an adaptive variable. 

The same learner might require substantial support when encountering a new concept but progressively less support after demonstrating mastery. This creates a more useful principle:

Adapt the support, not the expectations.

The objective is not to create a permanently simplified learning environment. It is to create an environment that provides the appropriate level of support while allowing the learner to progress toward greater independence.

The Danger of Making Learning Too Easy

This leads to another major concern: AI overreliance. Generative AI is extremely good at reducing cognitive effort. That sounds beneficial. But learning often requires cognitive effort. If an AI tutor immediately provides the solution every time a learner encounters difficulty, the learner may become increasingly dependent on the system. The system has successfully completed the task. The learner has not necessarily learned how to complete it. This is why the goal of an educational AI system should not be:

"How quickly can I give the learner the answer?"

A better objective is: 

"How can I provide enough support for the learner to solve the problem themselves?"

This suggests interaction strategies based on graduated assistance. For example:

• Level 1: Ask the learner to attempt the problem.
• Level 2: Provide a hint.
• Level 3: Identify the conceptual area that may require reconsideration.
• Level 4: Provide a partial solution.
• Level 5: Provide a complete explanation only when necessary.

Such a system transforms AI from an answer generator into a learning scaffold. The distinction is subtle but fundamental. A tutoring system should optimize not merely for task completion, but for learning independence.

Can AI Actually Understand a Learner?

This may be the most difficult research question. Large language models can infer patterns from interactions, but inference is not equivalent to cognitive assessment. If a student takes a long time to answer a question, what does that mean?

• Perhaps the student has difficulty processing the concept.
• Perhaps they are distracted.
• Perhaps they are thinking deeply.
• Perhaps the interface is confusing.
• Perhaps they are simply not interested in the question.

A generative AI system should therefore be extremely cautious about making strong claims about a learner's cognitive profile based solely on behavioural traces. Instead, learner modelling should ideally combine multiple sources of evidence. For example:

Learner model = behavioural evidence + performance data + explicit learner feedback + educational assessment

Even then, uncertainty should remain part of the model. Rather than representing:

"This learner has poor working memory."

a responsible system might represent:

"Current interaction evidence suggests that the learner may benefit from shorter instructional segments."

That difference is not merely linguistic. It represents a fundamentally different philosophy of AI-supported education. The system is making an actionable prediction, not assigning an identity to the learner.

Human Agency Must Remain at the Centre

Ultimately, inclusive educational AI should not be designed around replacing teachers. Nor should it aim to replace learners' cognitive effort. The strongest systems may instead operate as adaptive intermediaries between learners, teachers, and educational content.

The AI can generate explanations. It can identify patterns. It can provide scaffolding. It can adapt presentation. But teachers remain essential for interpreting context, understanding learners as people, and making pedagogical decisions that cannot be reduced to interaction logs. Likewise, learners must remain active participants.

A successful AI tutor should ideally make the learner more capable without the AI, rather than making the learner more dependent on it. This provides a useful criterion for evaluating educational AI:

If removing the AI causes the learner's ability to collapse, have we created an effective tutor, or merely an effective dependency?

Toward Safer Generative Tutors

The future of generative AI in inclusive education is therefore unlikely to be determined by the size of the language model alone. It will depend on the architecture surrounding it. A robust generative tutoring system may require several interacting layers:

Generative Model → Verification → Learner Model → Adaptation → Interaction → Evaluation

The generative model produces content. The verification layer assesses reliability. The learner model represents what the system currently knows—and does not know—about the learner. The adaptation layer determines how much support should be provided.

The interaction layer determines how that support is delivered. Finally, the evaluation layer continuously examines whether the intervention is actually improving learning. This creates a shift in perspective.

We should stop thinking about educational AI simply as:

"an LLM that teaches."

Instead, we should think about it as:

"an adaptive socio-technical learning system in which a generative model is one component."

That distinction matters.

The Research Opportunity

Generative AI has given educational technology something previous systems struggled to provide flexible, natural-language interaction at scale. But flexibility introduces uncertainty. The more freedom a system has to generate responses, the more difficult it becomes to guarantee that those responses are appropriate for a particular learner. That tension between adaptability and reliability is at the heart of the research problem. For learners with learning difficulties, the stakes are even higher. The objective should not be to build an AI that simply talks differently to every learner. The objective is to build an AI that can adapt responsibly.

It should know when to explain.

• When to scaffold.
• When to challenge.
• When to reduce assistance.
• When to verify.
• When to admit uncertainty.

And, perhaps most importantly, when not to answer. The future of inclusive learning will therefore not be determined by whether generative AI can become a better teacher than humans. A more interesting question is whether we can design AI systems that extend human teaching capacity while preserving learner agency, safety, and independence. That is where generative AI becomes more than a technological novelty. It becomes a research problem.