When the Chatbot Summarises: AI Reading Support and the Limits of Helpful Struggle

A Year 9 class is reading a two-page explanation of resistance in electrical circuits. One pupil highlights the whole paragraph, pastes it into a generative AI tool and receives a five-line summary. The summary is accurate, clear and rather bland. The pupil copies it into their book. The lesson moves on. Nothing obviously goes wrong — and that is precisely the problem.

The appeal of the reading assistant

AI reading assistants now arrive in several forms: text-to-speech tools that read aloud, chatbots that explain or simplify, summarisers that condense, and question generators that claim to check understanding. Their appeal to busy teachers is obvious. Science texts are dense, and classrooms contain pupils with widely different reading abilities. A tool that can produce a glossary, simplify a sentence or read a passage aloud might plausibly help. The question is not whether such tools can be useful. It is what cognitive work they do, and what cognitive work they remove.

What comprehension requires

Cognitive science describes comprehension as an active construction. Readers connect new information to prior knowledge, monitor whether they understand, make inferences and reorganise ideas in memory. Summarising is not a neutral act of copying; it requires selection, deletion and generalisation. Explaining a concept in one's own words is a form of generation that tends to strengthen memory. Retrieval practice, spacing and elaborative questioning all depend on the learner doing something with the material. These processes are effortful, and that effort is often the point. The phrase desirable difficulty captures the idea: some struggle improves learning, provided it is productive rather than overwhelming.

Where the scaffold is genuine

There are real cases where AI support is a scaffold rather than a shortcut. For a pupil with dyslexia or a visual impairment, text-to-speech can provide access to the same science content as peers. For an EAL learner, a just-in-time explanation of an academic word may reduce extraneous load and allow attention to the science. A teacher might use a generative tool to produce a simplified version of a text for a pupil who is not yet fluent, while keeping the core concept intact. These uses share a feature: they support access to the text and free capacity for thinking about the science. They do not replace the pupil's own comprehension work.

Where the shortcut bites

The risk appears when the tool performs the inference, selects the key idea and presents the conclusion. A pupil can then produce a neat paragraph without having constructed meaning. The result may look like comprehension, but it may be closer to transcription. Cognitive load is reduced, but so is the productive struggle that builds durable knowledge. There is also a metacognitive danger: a fluent summary can create an illusion of understanding. Pupils may recognise the words and assume they could reproduce the reasoning. Without retrieval or explanation, that assumption may go untested until an assessment.

A practical settlement

Teachers do not need to choose between prohibition and uncritical adoption. A more defensible approach is to make the cognitive work visible and decide where the tool sits. AI can help prepare a lesson: generating a glossary, drafting a simplified text for teacher checking, or producing possible questions. It can support access: reading aloud, defining a word, translating an instruction. It should not normally replace the pupil's own summarising, explaining or inferencing when those are the learning goals. A simple test is to ask who is doing the thinking. If the answer is the machine, the scaffold has become a substitute.

Designing tasks that keep the pupil thinking

Science teachers can use AI without diluting the subject. One approach is to give pupils an AI summary alongside the original text and ask them to identify what has been lost. That turns the tool into an object of critique rather than an authority. Another is to use a chatbot as a questioning partner: pupils ask it to explain a confusing sentence, then check the explanation against the textbook and the teacher's model. After any AI-supported reading, pupils should close the tool and write from memory, answer a retrieval quiz or explain the concept to a partner. The tool may have helped them into the text, but the assessment should test what remains.

The verification problem

Science has an additional safeguard and an additional risk. The safeguard is that claims should be checkable against evidence, diagrams and data. The risk is that AI-generated explanations can be confidently wrong. Pupils need to be taught to verify, not merely to consume. That means comparing the tool's explanation with the source text, checking units and definitions, and asking whether a claim is consistent with what they already know. These are disciplinary reading skills, not generic digital skills, and they are worth teaching explicitly.

The test of a good scaffold

AI reading assistants are neither a revolution nor a menace in themselves. They are tools with particular affordances and particular costs. The educational question is whether they help pupils do more of the cognitive work that matters, or whether they quietly do it for them. A good scaffold is temporary, contingent and aimed at independence. It gives a reader enough support to attempt the next step and then removes itself. If an AI tool cannot pass that test, it may be efficient — but it is not teaching reading.