Should You Let AI Read Your Notes?

Every notes app now offers to read everything you have written and answer questions about it. The pitch is genuinely good: you have thousands of notes and no memory of most of them, and search only finds what you can already name.

It is also the single largest disclosure most people will ever make about themselves, in one click. Notes contain the things you would not put in an email. Worth thinking about for a few minutes before agreeing.

What "AI reads your notes" actually means

The phrase covers at least four different arrangements, with very different consequences.

Arrangement Where your text goes Practical risk
On-device model Nowhere. It runs locally Lowest. Limited by what a phone-sized model can do
Encrypted cloud inference To a server, in a form the provider states it cannot read or retain Low, and depends entirely on the provider's architecture being what they say
Standard cloud API To a third-party model provider, usually under a no-training agreement Moderate. Your text exists on someone else's infrastructure, briefly
Cloud, with training on your data To a provider that may learn from it Highest. Avoid unless you truly do not care

Most good products are the middle two. The important detail is usually buried: retention. Not "do they train on it" but "how long is it stored, and who could be compelled to produce it."

The questions that actually distinguish products

Skip the marketing page and look for answers to these:

  1. Is my whole library sent, or only the relevant part? Retrieval-augmented systems send a handful of matching notes per question, not everything. That is meaningfully less exposure than full indexing on someone else's server, though the index itself may live remotely too.
  2. Is anything retained after the answer? Ask about logs, not just training. Thirty days of request logs is a very different thing from zero.
  3. Is it on by default? A feature that quietly indexed your library the day you updated is a different consent story from one you switched on.
  4. Can I exclude notes? Per-note or per-folder exclusion is the single most useful control, and surprisingly rare.
  5. What happens on cancellation? Does the index get deleted, and can you verify that.

If a company cannot answer these in plain language, that is the answer.

What you get in return

The upside is real, and worth stating honestly rather than dismissing.

The core problem with a large note collection is that you cannot search for what you cannot name. You remember that you thought something about hiring, two years ago, in some context. Keyword search needs the keyword. Semantic search over your own corpus does not, and that is a genuine change in what a note archive is for. It is the difference between external memory as a filing cabinet and as something you can actually ask.

It also lowers the cost of writing notes you will never deliberately reread, which is most of them. See the collector's fallacy for the failure mode that solves.

What you should be careful about

It will find things you forgot were there. So will anyone else with access. Old journal entries, notes about other people, things written in anger. Ask yourself what is in there before deciding it is fine.

Notes about other people are not only yours. Meeting notes, therapy reflections, notes on a colleague's performance. Those people did not choose your app's privacy policy. This is the strongest argument for per-folder exclusion.

Professional obligations may decide this for you. Lawyers, clinicians, journalists, and anyone handling regulated data often cannot make this choice personally, regardless of how good the provider's architecture is. Check before, not after.

Confident wrong answers are worse over your own notes than over the web. When a model summarises your thinking back to you, it is very easy to accept. It can merge two unrelated notes into a plausible claim you never made. Treat answers as pointers to notes, and open the note.

A reasonable position

You do not have to pick all-or-nothing.

  • Split the library. A sensitive folder that is never indexed, and the rest that is. Journal, health, and anything about other people in the first; reading notes, project notes, and reference in the second.
  • Prefer on-device for the private half, cloud for the corpus where retrieval quality actually matters.
  • Turn it on deliberately, after reading the retention answer, rather than by accepting an update prompt.
  • Revisit when the provider changes ownership. Privacy policies survive acquisitions less often than data does.

That split takes ten minutes to set up and removes almost all of the downside while keeping most of the benefit.

The thing that has not changed

Whatever you decide, the underlying rule is the same one that governed notes before any of this existed: the value is in what you wrote, and the risk is in where it lives. Encrypted notes, file over app, and knowing where your notes physically are matter more than any single feature toggle.

More on the practical side in how to take notes with AI, the AI second brain, and the best private notes app. For how one implementation works, see how Ask AI works.

More in Personal Knowledge Management

55 more notes on this branch.