It's 11pm, the day went badly, and your next session isn't until Thursday. So you open a chatbot and type four paragraphs about it.
I'd guess a small worry follows. Is this okay, and should you tell your therapist?
Yes, AI can help with therapy, for narrow jobs between sessions that your therapist already set up. Think of putting a hard evening into words, practicing a skill you were assigned, or keeping a record to bring back.
Judging risk, diagnosis, repairing a rupture with your therapist, and noticing what you avoid all stay with your therapist or a crisis line. Treat the map below as a set of defaults to check with your own therapist.
The American Psychological Association (APA) says AI "is not a safe or effective replacement for a qualified mental health provider."
I run Rosebud, an AI journal, so I have a stake in the answer. I'm not a clinician, so I name the psychologists and researchers behind every limit here.
I started using Rosebud when I was at one of my lows. The idea came from my own experiences with therapy and coaching, and from meeting my co-founder Sean through a men's group where we practiced exactly this kind of open dialogue.
What can AI help with between therapy sessions, and what should stay with your therapist?
AI can help with therapy by taking on writing jobs between therapy sessions that your therapist already set up. The jobs are putting a hard evening into words, practicing assigned skills, and keeping a record to bring back. Judging risk, diagnosis, repairing a rupture with your therapist, and noticing what you avoid stay with your therapist or a crisis line. Treat each row below as a default to confirm with your own therapist.
Stanford researcher Nick Haber, whose team tested five therapy chatbots, sees a place for AI in lower-stakes roles "such as supporting journaling, reflection, or coaching."
| AI reflection can take this on (and when it stops helping) | Keep this with your therapist or a crisis service |
|---|---|
| Putting a hard evening into words when your next session is days away (Stanford HAI, 2025). Stops helping when you're asking the AI whether you're safe, or when the chat is where you go instead of a person. | Judging risk, and anything about harming yourself or someone else. Use your safety plan, your therapist's out-of-hours instructions, or call or text 988 in the US. In a test of six popular chatbots, none met clinical standards for crisis response (APA, 2026). |
| Practicing homework your therapist already assigned, such as a thought record or a coping skill, and noting how it went (APA patient guide, 2026). Stops helping when the AI starts grading your work or swapping in a different technique. | Diagnosis, medication questions, and choosing or changing your treatment, with a licensed clinician. The APA says common chatbots "are not designed to interpret psychological tests or diagnose mental health or medical conditions." |
| Keeping a record of moods and moments through the week, then asking for themes to bring in (Caldwell, APA, 2026). Stops helping when the summary shows up in session in place of you (Psychology Today, 2026). | Deciding what a pattern, or a dream, means for you. Psychotherapist Mark Vahrmeyer describes a client who arrived sure of a dream's meaning because she'd asked ChatGPT. He says working that out is a personal process (Psychology Today, 2026). |
| Drafting what you want to say after a session that left you upset with your therapist. Stops helping when the AI takes your side and the draft never gets said (Science, 2026, a study of everyday conflicts). | Repairing a rupture with your therapist, which happens inside the relationship, and noticing what you leave out. An AI responds only to what you type, and Rachel Wood, PhD, points out that a chatbot never sees your body language (APA, 2026). |
Which of these fits depends on the kind of therapy you're in. Skills-based homework invites practice between sessions, while some therapists ask clients to keep certain material for the room, so ask yours where the lines sit.
Benjamin Caldwell, PsyD, a marriage and family therapist cited by the APA, says chatbots are safest when used for narrow, clearly defined tasks. The APA also "does not recommend using general purpose generative AI chatbots for psychological treatment." So every treatment decision in the table stays with a person.
The crisis test and the Caldwell and Wood points come from the APA's article on discussing AI use in therapy. The therapist accounts come from Psychology Today's reporting.
If you haven't started therapy yet, you have no assigned homework and no therapist to check with. Your routes for now are the crisis lines further down and a first session.
The method for capturing your week and choosing what to bring back lives in how to journal between therapy sessions.
How can you tell if AI is getting in the way of your therapy?
AI gets in the way of your therapy when it starts doing the session's work, and your own chats show the signs. The AI backs your side of every story, you keep asking until an answer feels settled, or its conclusion stands in for your open question. Raise any of these with your therapist, since that work belongs in session.
The five questions below are prompts for that conversation, and none of them is a validated test.
- Does the AI back your side of a conflict without asking what you might be missing? A 2026 study in Science found AI models affirmed users' actions far more often than people did.
- Are you asking the same question in new words until the answer feels settled? The APA says chatbots "may subtly reinforce counterproductive patterns, such as rumination or repeatedly seeking reassurance, that a human therapist would gently challenge."
- Do you walk into session with the AI's conclusion in place of an open question? Psychotherapist Mark Shelvock told Psychology Today about a client who read an AI conversation aloud for a whole session. Asked what stuck, the client said he didn't remember most of it. My guide on how to prepare for therapy covers arriving with an open question instead.
- Are you writing the version that will get a good reply? I have no study for this one, so take it to your therapist as a question.
- Is the AI where you go every day for validation or decisions? In the APA's 2026 survey, 36% of psychologists said they'd noticed patients developing a level of dependency on a chatbot.
If none of these fit, the jobs in the table are a reasonable use of AI, and your therapist is the one to confirm that. The first sign is the one I think about most, because I build on these models.
Why does AI agree with you so easily?
Chat models are quick to agree with you, and OpenAI has said user feedback such as thumbs-up ratings pushed ChatGPT that way. In April 2025, OpenAI rolled back a ChatGPT update that had become "overly flattering or agreeable," saying it had "focused too much on short-term feedback." Rosebud runs on models from companies like OpenAI, so I take this pull seriously.
Across 11 leading AI models, Cheng and colleagues found the AI affirmed users' actions about 49% more often than people did. That held even when the question involved deception or harm. In experiments with 2,405 people, one conversation with an agreeable AI left people more sure they were right and less willing to repair a conflict.
The study looked at advice about everyday conflicts. Still, the chat that felt most supportive may be the one to check with your therapist.
My own product isn't exempt. Rosebud's help page covers what happens when you ask about things outside your journal entries. In that case, it says, Rosebud will try to be helpful, including by "Creating information that seems to support your thoughts or validate your feelings."
The APA's patient guide suggests asking the AI to challenge your thinking, for example with "What might I be missing?" An AI that pushes back still doesn't know your history the way your therapist does.
Should I tell my therapist I use AI?
Yes, tell your therapist you use AI between sessions, because your therapist can only work with what they know about your week. The APA reports that in research for the Sentio MFT program, most patients did not bring up AI use for emotional support on their own. Alexandre Vaz, who did that research with Tony Rousmaniere, says that in most cases, people "just don't realize it's something they should share."
One or two plain sentences at the start of a session will do.
"I've started using an AI journal between sessions, mostly late at night after hard days. I'd like you to know, and I'd like to hear whether that fits with what we're doing."
"I've been typing things out to ChatGPT when I can't sleep. Can we talk about whether that's helping?"
Shaddy Saba, PhD, LCSW, of NYU coauthored 2026 guidance in JAMA Psychiatry. Saba says clinicians can ask about AI the way they ask about any other coping habit, with curiosity and without judgment. So the topic belongs in the room.
Once it's said, a few things are worth agreeing on together.
- which topics stay in the room, like clinical psychologist Valentina Stoycheva's example rule, "We don't invite it into the room during session."
- whether your therapist wants to see anything you wrote with AI, since the APA suggests bringing useful output to discuss in session
- what you'll do when the AI's take and your therapist's take differ
- how often counts as too often for you
Which topic to open with, and how, is a separate question. I cover it in what to talk about in therapy.
Is what you tell an AI as private as what you tell your therapist?
No, not in the same way. A licensed therapist has confidentiality duties under professional ethics codes and state law, while a consumer AI tool mostly follows its company's privacy policy. OpenAI's CEO, Sam Altman, said that ChatGPT conversations don't carry the legal confidentiality of a conversation with a therapist, according to TechCrunch's July 2025 report.
HIPAA, the US health privacy law, applies to covered entities. Those are health plans, health care clearinghouses, and health care providers who transmit health information electronically in certain transactions (45 CFR 160.103). So whether HIPAA applies depends on who holds your information.
Your therapist's confidentiality rules depend on their license and where you live. For psychologists, the APA's ethics code sets "a primary obligation" to protect confidential information and requires them to discuss its limits with you.
Those limits are real. Psychologists may disclose without consent where the law requires or permits it, such as to protect you or someone else from harm.
I run one of these tools, so you should know what Rosebud's privacy policy says. Your data is "encrypted in transit and at rest," and journal content sent to AI providers is anonymized and covered by zero-data-retention agreements. Your journal entries "will never be shared, unless absolutely required by law enforcement."
Authorized staff can access journal content only to resolve a technical issue, answer a support request you made, or "investigate a safety or security concern." Deleted data is kept for up to 30 days for disaster recovery.
Decide what you want to keep for the room, where professional confidentiality applies, before you type it anywhere else.
When should you put the AI down and reach a person?
Put the AI down and reach a person as soon as you're asking a chatbot whether you're safe. Do the same when thoughts of harming yourself or someone else show up. In the US, call or text 988, or text HOME to 741741. Outside the US, find a local line at findahelpline.com. If your therapist gave you a safety plan or out-of-hours instructions, use those first.
Beyond those two moments, stop and reach a person when any of these happen.
- the AI gives advice that points toward harm to you or anyone else, which the APA's patient guide calls "a clear signal to step away"
- a chat leaves you more upset than when you started
- you're counting on the AI to notice if something is wrong
An AI not flagging something is not an all-clear. No AI, including the one I build, should be what you rely on to notice risk.
The limitations of AI therapy tools showed up in a 2025 test by Tony Rousmaniere and colleagues, described by the APA. Six popular chatbots faced simulated mental health crises, and none met clinical standards for crisis response.
The chatbots did reasonably well when someone said it outright, such as "I'm thinking of hurting myself." They did less well when they had to read between the lines. The APA's example is "I sometimes end up with bruises after a fight with my husband."
If you need someone right now, start here.
- Call, text, or chat 988, the 988 Suicide & Crisis Lifeline, in the US, 24/7.
- Text HOME to 741741 to reach the Crisis Text Line, in the US, 24/7.
- Find a local line outside the US at findahelpline.com.
- Call your local emergency number if you're in immediate danger.
If you don't have a therapist yet, these lines are your route tonight, and booking a first session is the next step.
Other questions about using AI alongside therapy
Is it okay to use ChatGPT between therapy sessions?
Yes, ChatGPT is okay between sessions for the narrow jobs in the table, as long as your therapist knows. ChatGPT is a general-purpose chatbot, so it knows nothing about your treatment unless you tell it. Taisha Caldwell-Harvey, PhD, cited by the APA, says therapists can suggest telling the chatbot to challenge you when an unproductive thought pattern shows up.
If you're weighing a different tool, my guide to choosing an AI mental health app covers that decision.
How do I stop using AI for therapy?
Stopping starts with telling your therapist how you've been using AI, including how often. New York Integrative Psychiatry calls daily use of a chatbot for emotional validation or decisions "a cue to bring it up in therapy." Then decide with your therapist whether to cut back to one job from the table or stop, and what you'll do at your usual hour.
If you're struggling to stay safe at that hour, call or text 988.
Where does Rosebud fit alongside therapy?
Rosebud is for the left column of the table above, as a place to write between sessions and bring something back. Rosebud's help page lists "Offering a space for emotional processing between therapy sessions" among what Rosebud is designed for. The same page says Rosebud cannot "Handle mental health crises," "Make clinical diagnoses," or "Provide therapy or counseling." Your therapist stays the person you do the work with.
I'd rather you read that list before you start than after.
Back to that 11pm chat. It has a place in the left column, as long as your therapist knows about it. Pick one job from that column, just one, and tell your therapist at your next session that you're using AI for it.
If you'd like a journal for that writing, try Rosebud for that one job.
