Write first, ask second, and keep a record of what you accepted. That is the whole method. Ten minutes of raw writing with the AI panel closed, then one bounded question to the model about your own entries, then two lines noting which of its phrasings you kept and which you rejected. In that order, the tool does what it is good at, reading across weeks of entries faster than you can, and never gets the chance to tell your story in its own words.
In The Art of Journaling, chapter 15, I name five tensions that come with inviting AI into a journal: authenticity, privacy, dependence, bias and ownership. The practical answer I give there is to treat the model as a conversation partner, alternate assisted and unassisted entries, and write your own boundaries down before you start. This page turns that into a session you can run tonight.
Quick Answer: Open the journal, close the AI panel, and write for ten minutes. Then ask the model one bounded question that quotes your own words back to you, such as which three topics you returned to most in the last 30 entries. Finish with two lines: the phrase it got right, the phrase it got wrong. Ask nothing open-ended, and never ask it what an entry means.
Definition: AI journaling is ordinary journaling in an app that can read your past entries with a language model and answer questions about them, generate prompts, or summarise themes.
Key Evidence: Baikie and Wilhelm (2005), reviewing the expressive-writing research, report that writing about stressful or emotional events for 15 to 20 minutes on three to five occasions produces better physical and psychological outcomes than writing about neutral topics. That benefit belongs to the writing, and the model adds nothing to it.
Context: Science journalist Angela Chen, testing three AI journaling apps on her own decade-old entries for Psyche, found that the same entry scored 50 percent anxiety on one run and 40 percent nostalgic on another, and that the suggestions were the advice anyone would give. Her conclusion was to keep the privilege of interpretation for yourself.
Key Takeaways
- Ten minutes of writing before the model sees anything. Set a timer and keep the AI panel closed until it rings.
- One question per session, and it must quote your own words back. "What does this mean?" is banned.
- Two lines of log at the end of every assisted session: one phrase kept, one phrase rejected, and why.
- If two of your last five entries were written with an eye to what the model would say, close the panel for a week.
- Before the first entry, check five privacy points: plain-text export, opt-in processing, per-entry exclusion, an on-device option, and what happens if the company is sold.
The write-first session
Copy this template into the front of the journal. Fifteen minutes.
- Write raw (10 minutes). AI panel closed, prompts off. Write whatever is there, in the order it arrives, without fixing sentences. If the app cannot hide its prompts, write in a plain file and paste in afterwards.
- Ask one bounded question (3 minutes). Pick from the five below. Each one forces the model to quote you and stops it interpreting.
- List the three topics I returned to most in the last 30 entries, quoting my own words for each.
- Which people are named in the last 30 entries, and how many times each?
- Quote every sentence from the last 30 days in which I wrote "I should".
- Which weekday do my shortest entries fall on?
- Quote the last entry in which I wrote about the same problem as tonight's, with its date.
- Log the exchange (2 minutes). Two lines under the entry: "Kept: ..." and "Rejected: ... because ...". After a month the log shows whether the model's phrasing is creeping into your own.
What the tool can and cannot do
| It does this well | It cannot do this |
|---|---|
| Recall across entries: find every mention of a name, a phrase or a topic in months of writing, in seconds | Sit with a feeling that has no words yet; the model will offer a summary before you are ready for one |
| Generate a prompt when the page is blank | Know your history beyond what you typed into it |
| Summarise themes, in its own phrasing, which belongs in the log only | Replace a clinician, a friend or a night's sleep |
| Combine entries with sleep, location and phone-use data to shape prompts, as the MindScape prototype at Dartmouth does | Give consistent answers: Chen's re-analysed entry came back with different numbers each time |
The research is younger than the products. The MindScape paper (Nepal and colleagues, 2024) describes a design and a preliminary study of contextual prompts for college students; it reports no outcome that separates the AI layer from the writing underneath it. Any claim that AI journaling improves health is borrowed from the expressive-writing literature, which studied pen and paper. For the wider evidence, see what the research shows on therapeutic journaling.
The same entry, two ways
This example is invented for illustration. Suppose the raw entry reads: "Told Sam about the job. He said congratulations twice, the second time looking at his phone. I do not know if I am angry at him or at myself for wanting him to be more excited than I am. I am not excited. I said yes because saying no needed a reason I did not have."
A model's tidy summary of that entry might read: "You shared news of your new job with Sam and felt his response was lukewarm. You are processing mixed emotions about the decision and about wanting validation."
The summary is accurate and useless. It dropped the phone, which is the detail that carries the hurt. It dropped "I am not excited," which is the sentence the whole entry was written to reach. And it replaced a sentence about having no reason to say no with "processing mixed emotions," which could sit under any entry ever written. Logged as "Rejected: lukewarm, mixed emotions, because it lost the phone and the yes," the summary has done its work. Pasted in as the entry, it makes the journal record the model's account of your evening.
The switch-off rule
Rule: If two of the last five entries were written with an eye to what the model would say back, close the panel for a week and write on paper or in a plain file. Reopen it only when the entries have gone back to being ugly.
Chen describes the general version of this trap, citing the digital-culture researcher Jill Walker Rettberg: people shown data about themselves tend to retell their lives to fit the data. A journal that is being written for its reader has stopped being a journal, whatever the reader is.
Five privacy checks before the first entry
- Export: can you get every entry out as plain text or markdown, without paying, today?
- Opt-in: is AI processing off until you switch it on, per entry or per journal?
- Exclusion: can you mark a single entry as never sent to the model?
- On-device: is there a mode where the model runs on your phone or computer and nothing leaves it?
- Acquisition: does the policy say what happens to your entries if the company is sold or closes?
If the answer to the first is no, stop there. The digital journal privacy guide covers the rest, including how to read a policy. Chen's own line is worth copying: one month of decade-old entries went in, the rest stayed off the service.
When this does not work
Three cases call for a paper notebook. In acute grief or a crisis, the model's prompting can feel like being hurried. In entries about other people, the third party never agreed to have their words processed; use the exclusion setting every time. And if you already over-analyse, the tool feeds the habit; Chen quotes James Pennebaker, who started the expressive-writing research, saying that too much introspection turns into rumination. In that case the ten minutes of raw writing are the whole practice and the panel stays closed.
Why AI journaling matters
Re-reading is the step most people skip, and in The Art of Journaling, chapter 5, pattern analysis (regular review, marking recurring themes, tracking them over time) is the method I give for getting past a journal that is only ever written and never read. A model can do the finding in seconds, which makes re-reading possible for someone with two years of entries and no free evening. The condition is that the model only finds and quotes. Once it is allowed to interpret, you have handed over the one part of the practice that was doing you any good.
Related reading
- Top Digital Journaling Apps for Mental Health
- Digital Journal Privacy: How to Protect Your Entries
- How to Start Digital Journaling: Setup, Tags, Photos and Backups
- Why Your Personal Journaling App Isn't Helping You Heal
- Digital Journaling
Sources
- Baikie and Wilhelm (2005), Advances in Psychiatric Treatment – Review of the expressive-writing paradigm: 15 to 20 minutes of writing on three to five occasions, with better physical and psychological outcomes than neutral-topic writing
- Nepal et al. (2024), Contextual AI Journaling: the MindScape App – Design paper and preliminary study on combining a language model with sleep, location and phone-use data to generate journaling prompts for college students
- Angela Chen, Psyche – First-person test of Rosebud, Insight Journal and Mindsera on a decade of entries: inconsistent sentiment scores, generic advice, the Rettberg point about retelling a life to fit the data, and Pennebaker on introspection becoming rumination
