10 Things to Avoid When Using AI for Journaling

Key takeaways

• 1. Sharing With Apps That Train On Your Data
• 2. Treating AI Like a Therapist
• 3. Letting the AI Write For You

1. Sharing With Apps That Train On Your Data

The single most important thing to verify before pouring your inner life into an AI journaling app is whether your entries will be used to train the company's models. This is not a hypothetical concern. Independent audits like Mozilla's Privacy Not Included reviews have repeatedly flagged mental-health and journaling apps for vague or permissive data-use language, and the U.S. Federal Trade Commission has taken action against mental-health platforms that shared sensitive user data in ways consumers did not reasonably expect.

The reason this matters is that training data is, by design, hard to fully retract. Once a model has seen your text, even anonymized fragments can influence its weights in ways that are not cleanly reversible. If you later change your mind or the company is acquired by a less careful operator, the data is already absorbed. Paper journals and locally encrypted apps do not have this problem, because there is no model on the other end learning from you.

When you read a privacy policy, look for explicit language. The phrase you want to see is something close to "we do not use your content to train AI models." Soft phrases like "we may use aggregated insights to improve our services" leave the door open to model training under a different name. If the policy is silent on the topic, assume the answer is unfavorable until you can confirm otherwise in writing.

Empath was built with this constraint in the foundation rather than bolted on later. User journal content is never used to train AI models, and the infrastructure is HIPAA-compliant, which imposes stricter handling requirements than a typical consumer app. That is not a marketing flourish; it is a baseline most people should expect from any tool that holds their inner life.

2. Treating AI Like a Therapist

AI journaling tools have become remarkably good at sounding empathetic. Modern language models can validate feelings, ask reasonable follow-up questions, and produce reflections that read as thoughtful. The risk is that this performance of care gets mistaken for the real thing, and people begin to substitute AI conversations for clinical support they actually need.

A therapist does several things an AI cannot. They hold a long-term relationship that itself is therapeutic. They assess risk in real time, including risk you may not be aware of yourself. They adapt their approach based on training, supervision, and accumulated case experience, and they are accountable to a licensing body if something goes wrong. None of that exists in an AI chat window, no matter how kind the responses sound.

The practical danger is most acute around crisis moments. If you are in acute distress, having suicidal thoughts, or working through trauma, an AI journaling app is not the right primary support. Many apps include crisis-resource handoffs, which is good, but the handoff is only useful if the user takes it. People who have built up a sense that the AI "gets them" sometimes resist that handoff, which is the opposite of what should happen in a crisis.

A healthier framing is to treat AI journaling as a between-sessions companion, not a replacement for sessions. Empath was explicitly designed to complement therapy rather than replace it, which is why many users bring summaries and patterns from their entries into their next appointment. The AI helps you notice; the human helps you heal.

3. Letting the AI Write For You

One of the most subtle ways AI journaling goes wrong is when the user stops doing the reflective work and starts outsourcing it. This usually happens gradually. You begin by asking the AI to rephrase an entry, then to summarize a week, then to draft "what you would write if you had more time." At some point, the journal contains more of the model's voice than your own, and the practice loses most of its psychological value.

The expressive writing literature, much of it built on decades of research by James Pennebaker and colleagues, is fairly clear that the benefit of journaling comes from the act of articulating your own experience in your own words. The struggle to find language for something half-formed is where insight tends to happen. If a model finishes the sentence for you, you get a clean paragraph but lose the cognitive work that produces understanding.

There is also an authorship problem that compounds over time. If you reread entries a year later and cannot tell which thoughts were yours and which were the model's, the journal loses its function as a record of your inner life. It becomes a co-authored document with a partner who has no skin in the game. That is fine for a blog post; it is not fine for the only longitudinal account of your own mind.

A reasonable rule of thumb is that the AI should ask questions, surface patterns, and reflect what you wrote back to you, but the prose in the entry itself should be yours. Use AI as a mirror and a prompt machine, not a ghostwriter. The friction of writing badly in your own voice is part of the practice, not a bug to be optimized away.

4. Skipping Encryption and Privacy Policy Fine Print

Most people will not read a ten-page privacy policy, and product teams know this. The result is that a lot of consequential detail lives in language that is technically disclosed but practically invisible. Before committing to an AI journaling app for any length of time, it is worth spending fifteen minutes skimming the policy specifically for a handful of items: how data is encrypted, where it is stored, who can access it internally, how long it is retained after deletion, and what happens if the company is sold.

Encryption in transit is now standard and means very little on its own. What matters more is encryption at rest, ideally with strong key management, and in some cases end-to-end encryption where the provider itself cannot read your entries. Each model has trade-offs. End-to-end encryption is the strongest privacy posture but limits what AI features the app can offer, since the server cannot see plaintext. Server-side encryption with strict access controls is a common middle ground that preserves AI features while still protecting against most realistic threats.

Data retention after deletion is another underread section. Some apps continue to hold "backups" or "anonymized" copies for months or years after you delete your account. If your goal is to be able to fully withdraw your information later, you want a clear, time-bound retention policy and a documented deletion path. "We will delete your data upon request, subject to legal and business requirements" is a phrase that quietly preserves the company's right to keep almost anything.

Apps operating under HIPAA, like Empath, are bound to a tighter baseline because health-information regulation imposes specific requirements around access logging, breach notification, business-associate agreements, and minimum-necessary use. That does not make any single HIPAA-compliant app automatically perfect, but it removes a significant amount of guesswork about the floor.

5. Using Apps Without Longitudinal Memory

Many AI tools that look like journaling apps are actually thin chat interfaces over a stateless model. Each session begins from scratch. The AI may seem to remember earlier sessions because you paste in context, but under the hood it has no durable view of who you are, what you have been through, or what patterns are emerging in your life. For one-off venting, that is fine. For journaling, it is a significant limitation.

The reason longitudinal memory matters is that the most valuable journaling insights are almost never visible in a single entry. They show up across weeks and months: the recurring trigger you only notice on the third occurrence, the seasonal mood pattern that takes two years of data to confirm, the slow shift in how you talk about a relationship. A stateless chatbot cannot see any of that, no matter how clever it sounds inside a single conversation.

When evaluating an app, it is worth asking what it actually remembers. Does it index entries so the AI can reference your own prior writing? Can it produce summaries that span a month or a quarter? Does it surface patterns you did not ask about, or only respond to direct prompts? An app that cannot do these things is a useful writing surface but not really a journaling AI in any meaningful sense.

Empath was designed around this longitudinal view. Patterns are detected across entries rather than reset each session, which is what makes voice calls, text messages, and in-app entries combine into a coherent picture over time rather than a series of disconnected snapshots. Without that continuity, the AI is essentially meeting you for the first time every day.

6. Writing for the Algorithm Instead of Yourself

There is a quiet failure mode where users begin to write differently because they know an AI is reading. The entries get more articulate, more structured, sometimes more positive. The rough edges that make journaling therapeutic get sanded off, because the user has internalized an imagined reader who is graded on whether their week looks good. This is a documented version of the broader observer effect in psychology: behavior shifts when people know they are being watched, and the same shift can quietly distort a journaling practice.

The cost is not abstract. Pennebaker's expressive writing research suggests that the therapeutic benefits of journaling come disproportionately from writing about difficult, unresolved, or emotionally messy material. If the practice starts selecting for entries that read well, the entries that would actually help most never get written. Over time, the journal becomes a curated highlight reel rather than an honest record.

One useful corrective is to deliberately write entries you have no intention of sharing or summarizing. Resist the urge to read the AI's response immediately after writing. Some users find it helpful to keep a small portion of their journal completely outside the app, in a paper notebook or a fully local note, specifically for material that feels too tender or too unflattering to entrust to any system.

It is also worth checking your own writing periodically against a simple question: would you write this the same way if no AI were going to read it? If the honest answer is no, that is useful information about the practice, not a failure of the tool. The remedy is usually to write rougher, not to find a better app.

7. Over-Relying on AI Summaries Instead of Reading Your Own Entries

AI-generated summaries are genuinely useful. They can compress a month of entries into a few paragraphs, highlight emotional themes, and surface things you would not have spotted on your own. They are also, by design, a compression, which means information is being lost on purpose. The mistake is treating the summary as a substitute for the underlying entries rather than as a pointer back into them.

There is a specific class of insight that only emerges from rereading your own words. The way you described a relationship two months ago, the small detail you mentioned and then never returned to, the change in vocabulary you use when you describe your work. These artifacts almost never survive summarization, because they are not the kind of pattern the model is trained to extract. They are the kind of pattern that catches you off guard when you reread, and that surprise is often where personal insight lives.

A reasonable practice is to read the AI summary, then go back and skim the actual entries it was built from at least occasionally. Monthly is usually enough. You are not trying to audit the AI; you are giving yourself the chance to encounter your past self directly rather than through a model's interpretation. The two activities are complementary, and neither one replaces the other.

It is also worth being skeptical of confident-sounding AI summaries, especially when they categorize your emotional life with clean labels. Real inner lives rarely sort cleanly into "themes," and a summary that sounds too tidy is often smoothing over exactly the contradictions and ambiguities that matter most. Treat AI summaries as drafts of a possible interpretation, not the official transcript.

8. Ignoring Modality Mismatch and Quiet Red Flags

A surprisingly large share of failed journaling habits come down to a simple modality mismatch. The user signed up for an app that requires typing, decided typing felt like a chore on hard days, and stopped journaling on exactly the days when journaling would have helped most. The right modality depends on the moment. Sometimes a five-minute voice ramble in the car is the only realistic option. Sometimes a single sentence sent by text is enough. Sometimes a long, considered written entry is what the day calls for.

This is one of the reasons Empath supports multiple inputs by design: phone call, SMS text, in-app typing, and in-app voice all flow into the same longitudinal record. The point is not novelty for its own sake; it is that a journaling practice survives much longer when it adapts to your day instead of demanding that your day adapt to it. BJ Fogg's habit research has been making variations of this point for years: behaviors that require less friction get repeated, and behaviors that require more get dropped.

The final pitfall worth naming is ignoring quiet red flags about the app itself. Privacy policies that change without clear notice, a sudden acquisition by a company in an unrelated industry, vague answers to direct data-handling questions, or new features that read more like advertising hooks than reflection tools. None of these individually mean an app is unsafe, but together they often mean the product's priorities are shifting in a direction that is no longer aligned with the user's.

A useful habit is to do a brief annual check on whatever app holds your journal. Reread the current privacy policy, confirm export and deletion still work, and make sure the company's public posture still matches what attracted you in the first place. If something has quietly changed, you want to know before you have another year of entries inside it. The whole point of journaling is to know yourself better over time; the tooling around it should make that easier, not quietly trade it away.

Frequently asked questions

Is it safe to use AI for journaling at all?

For most people, yes, provided the app is carefully chosen. The core risks come from how a specific product handles data, not from AI journaling as a category. Look for clear statements that your entries are not used to train models, strong encryption, a defined retention policy, and ideally compliance frameworks like HIPAA. If those pieces are in place, AI journaling can be both useful and reasonably private.

How can I tell if an AI journaling app trains on my data?

Read the privacy policy and look for an explicit statement one way or the other. The clearest sign is direct language such as "we do not use your content to train AI models." If the policy is vague or talks about using "aggregated" or "anonymized" data to "improve services," treat that as ambiguous and ask support for clarification in writing. Silence on the topic generally favors the company, not the user.

Can AI journaling replace seeing a therapist?

No. AI journaling can complement therapy by helping you notice patterns, prepare for sessions, and reflect between appointments, but it cannot provide diagnosis, crisis intervention, or the long-term therapeutic relationship that drives real healing. If you are in acute distress or working through trauma, the AI is best treated as a journaling tool that points you toward professional support, not as a substitute for it.

What is the difference between AI that remembers and AI that does not?

Some AI journaling tools are essentially stateless chatbots that start from zero each session. Others maintain a longitudinal view of your entries and can reference patterns across weeks or months. The longitudinal version is much closer to what people usually mean by "AI journaling," because the most valuable insights tend to require continuity. Empath is built around that continuity rather than around isolated chats.

Should I let the AI write entries for me to save time?

Generally no. The therapeutic value of journaling comes largely from the act of putting your own experience into your own words, even when the words come out clumsy. AI is best used to ask questions, surface patterns, and reflect content back, while the entries themselves stay in your voice. A journal mostly written by a model is a fine artifact but does not produce the same psychological benefits.

What should I do if my journaling app gets acquired or changes its policy?

Treat it as a prompt for a careful re-read. Look at the new privacy policy, confirm whether data-handling practices have changed, and verify that export and deletion still work the way they used to. If the new terms are materially worse, export your data and consider moving to a different tool. A journal is a long-lived asset, and the underlying product's posture toward your data should be something you actively re-consent to over time.

How do I journal honestly when I know an AI is reading?

Awareness of an observer can quietly shift what you write, which is a documented effect in psychology. Useful corrections include writing entries you have no intention of reviewing as summaries, keeping a small portion of your reflection completely offline, and periodically checking whether you would write the same way if no AI were going to see it. The goal is honesty with yourself; the tool is supposed to serve that, not shape it.