Note-taking is still king, but can AI help kids understand context?

Debates about the role of AI in education often drift toward extremes.

For some, generative AI promises a long‑imagined future of a personalised tutor for every learner and for others, it raises concerns about rampant AI-enabled cheating and cognitive atrophy.

But perhaps what lies ahead sits somewhere in between these two ends of the spectrum – a future that thoughtfully blends the best of what we know from decades of research and practice in education with new tools and technologies that can enrich learning rather than replace it.

To move the discussion forward, together with colleagues at Cambridge University Press & Assessment and Microsoft Research, we conducted one of the first large‑scale classroom experiments on using large language models (LLMs) as reading aids.

Over 400 students across seven secondary schools in England studied two short history passages – one with the help of an LLM and one using either note‑taking alone or combined with the LLM. Three days later they completed a short test assessing what they retained and understood from the texts.

Note-taking remains highly effective

Our primary finding was clear. Students performed better when they took notes – whether or not they also used an LLM – than when they relied on the LLM alone.

Note‑taking remains a highly effective, well‑understood method for consolidating information, and a strong baseline for any new tool to surpass.

But in further exploring the students’ interactions with LLMs, we gained some rather striking insights into unexpected ways that the students used the technology.

In particular, most students (over 90 per cent) engaged with the LLM not to cheat or avoid reading, but instead to enhance their understanding of the text.

Many asked for information that went beyond the text, asking about historical context, unfamiliar references or the significance of key events.

For instance, one student reading about apartheid asked, “What was Mandela’s life story?” Another, studying the Cuban Missile Crisis, wanted to know, “Why was America afraid of communism?”

Why didn’t it raise scores?

From an instructional perspective, this is precisely the kind of curiosity we hope to spark. Students weren’t primarily summarising or shortcutting. They were situating the material within a broader landscape.

So why didn’t this translate into higher scores?

First, note‑taking is a proven strategy that requires cognitive engagement – selecting, paraphrasing, and organising information – known to support learning.

Second, in the combined note‑taking‑plus‑LLM condition, we saw a reasonable number of students (about 25 per cent) copy LLM output directly into their notes.

Copying, of course, doesn’t yield the same benefits as generating one’s own summaries, and likely diminished benefits of the combined approach.

But third – and perhaps most thought provoking – is that our assessments were all confined to assessing recall and comprehension of what was contained within the text itself.

While the LLM helped students explore surrounding context, our test was not designed to assess this aspect of learning.

Preserving strengths while incorporating benefits

This may also help explain a gap between the students’ perceptions and their performance.

Many reported feeling that the LLM was more “helpful” than note‑taking, even though their performance suggested the opposite.

On one hand, LLM‑assisted explanations can make material feel more accessible, but this can remove desirable difficulties that lead to learning.

On the other hand, it’s plausible that having a reading aid that can answer questions beyond the text is in fact quite helpful in addition to having a space to take notes.

These findings raise a practical question: how might we design learning activities that preserve the strengths of traditional strategies while incorporating the benefits students clearly value in AI tools?

One option is to sequence activities more deliberately, encouraging students first to read and explore with an LLM and then to take independent notes without the ability to paste text.

Another is to provide explicit guidance on using LLMs as tools for clarification and inquiry rather than as substitutes for effortful processing.

What could it mean for the future?

It is also important to consider the context of our study in interpreting these results.

Students were in supervised classroom settings, focused on using the entire period to read these passages. They had no access to other tools, nothing else competing for their attention, and were interacting with a then novel AI system.

Students working independently at home, trying to complete as much homework as possible in as little time as possible, might behave differently.

Yet our experimental environment might also point toward towards a productive future learning environment: periodic, in-class LLM‑supported reading sessions, paired with in‑class, LLM‑free assessments that gauge both core understanding of material and the broader insights that students develop, linking material to the rest of their studies.

If designed and used carefully, perhaps AI need not replace or diminish established learning practices, but can instead complement them, broadening exploration without compromising the desirable difficulties that lead to effective learning.