Using NotebookLM as a Cornish grammar assistant

One of the things I have found most frustrating about using ordinary AI tools for Cornish is that they are both useful and unreliable.

If I ask a general chatbot to help me with Cornish, it can often give me helpful ideas. It can suggest activities, explain a grammar point, or help me think through how to organise learning. But when it actually produces Cornish, the results are much less dependable. It may mix spelling systems, invent forms, correct something that was not wrong, or give an answer that sounds confident but is not really grounded in reliable Cornish sources.

That is why I have been experimenting with NotebookLM as a different kind of assistant.

NotebookLM is Google’s source-based AI notebook. The important difference is that, instead of simply asking a general chatbot a question, you upload sources into a notebook and then ask questions about those sources. Google describes it as an AI research and thinking tool that can analyse your sources and help turn complex material into clearer explanations. (Google NotebookLM)

For a minority language like Cornish, that distinction matters. I do not want an AI system to pretend that it “knows Cornish”. I want it to work from sources that I have chosen.

Setting up NotebookLM

To try it, go to NotebookLM and sign in with a Google account. Google’s own support page says you can sign up free of charge using a Gmail account, or use it through some qualifying Workspace or Workspace for Education accounts. (Google Help)

Once you are in, the basic process is simple:

1. Create a new notebook.
2. Add sources.
3. Upload PDFs, paste text, or add other supported source material.
4. Ask questions about those sources.
5. Check the answers against the sources, rather than treating them as automatic truth.

For Cornish, I would begin with a dedicated notebook called something like:

Cornish Grammar Sources

or:

Kernewek Grammar Notebook

Then I would upload only a small number of carefully chosen sources at first. It is tempting to throw everything in, but a smaller, cleaner notebook is easier to test.

Why this may be better than ordinary chatbot use

The attraction of NotebookLM is not that it magically solves Cornish. It does not.

The attraction is that it gives the AI a more disciplined job. Instead of asking:

Is this Cornish sentence correct?

I can ask:

Based only on the uploaded grammar sources, what might be wrong with this sentence?

That is a much better question.

It encourages the system to act less like an oracle and more like an assistant sitting next to a pile of grammar books. It can help me find relevant rules, compare examples, and notice possible issues. But I still have to decide.

That is the right relationship, I think. For Cornish, AI should not be the authority. It should be a tool for working more efficiently with reliable sources.

Free sources to start with

I would avoid uploading scans of modern books that are still available to buy. For example, Bora Brav and A Grammar of Modern Cornish are purchasable books from the Cornish Language Board shop, so I would not treat them as free upload material unless I had explicit permission. (cornish-language.org) (cornish-language.org)

For a first experiment, I would use freely available sources instead. Possible starting points include:

1. An Outline of the Standard Written Form of Cornish

This is especially useful because many learners and teachers now work in or around the Standard Written Form. It gives useful information about spelling conventions and variation within SWF.

2. The Akademi Kernewek SWF dictionary PDF

This is not a grammar, but it is still extremely useful. The PDF dictionary explains that it uses the Standard Written Form and gives learners a way to check forms and spellings. (cornishdictionary.org.uk)

5. Wella Brown’s Grammar of Cornish

I would

I would make a clear distinction between free public sources and commercial teaching books. The point is not to avoid modern books; they are often the most useful sources. The point is that a shared NotebookLM workflow should be built from material that people are allowed to upload and use.

What I ask NotebookLM to do

I have found it more useful when I ask quite constrained questions.

For example, instead of asking:

Translate this into Cornish.

I might ask:

Based on the uploaded sources, what grammar points are relevant to this sentence?

Or:

Does this sentence appear to use a mutation correctly? Please refer to the uploaded sources.

Or:

Can you explain the difference between these two possible forms, using only the uploaded materials?

Or:

I am trying to write this in SWF. Which spellings in this sentence may not fit SWF?

Those questions are much safer. They do not ask NotebookLM to become a Cornish speaker. They ask it to help me read and apply sources.

What it is good for

The most useful thing is explanation.

NotebookLM can take a dense grammar point and make it easier to understand. It can also compare passages from different sources and help me find where a rule is discussed. That saves time.

It is also useful for checking my own uncertainty. If I have written a sentence and something feels wrong, I can ask NotebookLM to identify which grammar issues I should look at. It may mention mutations, word order, verbal particles, prepositions, pronouns, or spelling variants. Even when the answer is not perfect, it gives me a route back into the sources.

That is particularly useful for learners who are beyond the first beginner stage. At that point, the problem is often not “I know nothing”, but “I vaguely remember there is a rule here, but I cannot remember where to find it”.

NotebookLM can help with that.

What it is not good for

I would not use it as a final authority.

I would also be cautious about asking it to produce large amounts of Cornish. That is where AI systems often become overconfident. A beautiful answer is not necessarily a correct answer.

There is also the problem of spelling-system drift. Even when I want SWF, a system may pull in spellings from older texts, traditional forms, Kernewek Kemmyn, Unified Cornish, or Late Cornish sources. Sometimes that is interesting. Sometimes it is exactly the problem.

That is why the notebook needs to be built carefully. If I mix Jenner, SWF documents, historical texts, and modern teaching materials, I should not be surprised if the answers reflect that mixture.

A better approach may be to create separate notebooks:

• SWF grammar
• Teaching resources
• Vocabulary and dictionary checking

That way, the source base is clearer.

A useful prompt

This is the kind of instruction I would give at the beginning of a NotebookLM session:

Please answer only from the uploaded sources. Where there is uncertainty or disagreement, say so. Do not invent Cornish forms. If a sentence may involve different spelling systems or revival traditions, identify that rather than choosing one silently.

That instruction does not guarantee perfection, but it sets the right tone.

Another useful prompt is:

I am learning Cornish and want help understanding the grammar. Please explain the relevant rule in plain English, give the source basis for your answer, and avoid producing new Cornish unless the form is directly supported by the sources.

Again, the aim is not to get the AI to do everything. The aim is to make the sources more usable.

Why this matters for Cornish

Cornish learners face a particular problem. There are good resources, but they are scattered across books, PDFs, dictionaries, course materials, websites and older public-domain texts. There are also different spelling systems and revival traditions. For a learner, that can be confusing.

A source-grounded notebook does not remove the complexity, but it can make it easier to manage.

It allows me to ask:

• Where is this grammar point explained?
• Which source supports this form?
• Is this an SWF spelling or something else?
• What should I check before using this in teaching?
• Are there alternative forms I need to be aware of?

That feels like a much better use of AI than simply asking a chatbot to “write Cornish”.

The main lesson so far

My main conclusion is simple: NotebookLM is useful for Cornish when it is treated as a grammar assistant, not as a Cornish authority.

It can help me work with sources. It can help me find explanations. It can help me notice likely problems. It can help me organise my own learning.

But the human learner or teacher still has to listen, check, compare and decide.

For a small language, that seems exactly the right model. The technology should support careful learning, not replace it.