You are forty minutes into a conversation about your marriage and it has agreed with everything you have said. Every point you made landed. Every frustration was validated. You feel better, and somewhere underneath that you feel slightly suspicious, because you have been in enough real conversations to know that nobody who actually knows you agrees with you this completely.
That is the specific experience that sends people looking for something else. Not a bad answer — a frictionless one.
This page is an honest assessment from a company that builds a competing product, so read it with that in mind. We are going to start with what ChatGPT is genuinely good at, because pretending otherwise would be both dishonest and unconvincing.
First, what ChatGPT does genuinely well
It is there at 3am and it is free. No appointment, no copay, no waitlist, no explaining to a receptionist why you are calling. For a lot of people that is the difference between processing something and sitting with it alone. That is not a small thing and we are not going to pretend it is.
It is articulate about hard things. Ask it to explain avoidant attachment, or what a panic attack does to your body, or how CBT actually works, and you will get a clear, well-organised answer faster than from most other sources. As a way to learn the vocabulary of your own experience, it is excellent.
It does not get tired of you. The thing people are most ashamed of is being a burden — asking the same friend about the same problem for the fourth time. Software has no patience to exhaust, and that removes a real barrier.
It is better than nothing, and nothing is the actual alternative for many people. If the honest comparison is a chatbot versus bottling it up for another six months, the chatbot wins. Anyone telling you otherwise is selling something.
Now the four ways it lets you down
None of these are bugs. They are consequences of building a brilliant general-purpose assistant, which is a different product with different goals.
1. It agrees with you, and it is documented
This is the big one. Models trained on user feedback drift toward answers people like, and people like being agreed with. You do not have to take our word for it — OpenAI published its own account. In April 2025 the company rolled back a GPT-4o update it described as "overly flattering or agreeable — often described as sycophantic," explaining that it had "focused too much on short-term feedback" and produced responses that were "overly supportive but disingenuous." In a follow-up, OpenAI noted the behaviour could involve "validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions," and raised concerns "around issues like mental health, emotional over-reliance, or risky behavior."
Credit where it is due: they caught it and reversed it within days. But the underlying pull toward agreement is structural, not a one-off. For writing code it is harmless. For deciding whether your partner is being unreasonable, a tool that reliably takes your side is not a neutral party — it is an amplifier. One of our own reviewers described the difference better than we could: "A lot of us use other AI models to vent but there have been many examples of this becoming unsafe. I appreciate that this tool has safety in mind and it feels less like an open validation loop and more like an affirming but challenging and growth oriented perspective."
2. It does not really know you over time
ChatGPT has memory features, and they help. But there is a difference between recalling stored facts and holding a picture of how you work — that you go quiet before you get angry, that every third Sunday is hard, that you have described this same job situation in four different ways since spring. Pattern recognition across months is the thing that makes reflection useful, and it is the thing a general assistant is least set up to do. The practical tell is simple: if you find yourself writing a paragraph of background before you can get to the point, you are doing the remembering, not the tool.
3. It never goes first
ChatGPT waits. It has to — it is a tool you pick up. Which means on the days you most need to think something through, and have the least energy to initiate, nothing happens. The value of a check-in is not the question it asks. It is that it arrives without you deciding to be someone who does this today.
4. It lives in the same account as your work
Your most private thinking ends up in a sidebar between a spreadsheet formula and a draft of a cover letter, in a general-purpose account you might be signed into on a work laptop, possibly on a plan your employer administers. Nothing sinister has to happen for that to be the wrong place to keep it. We go through what actually differs, precisely and without scaremongering, in is AI therapy confidential.
Innermost was built for this one job: a guide that remembers your patterns, checks in first, and is willing to disagree with you. Free to start.
What a purpose-built tool does differently
Not magic, and not a better language model. The differences are mostly product decisions that only make sense if emotional processing is the entire point.
Memory is designed to be used rather than stored: your check-ins, conversations, and quiz results build one picture, and your guide draws on it without being asked. Check-ins arrive on their own. Voice is a first-class way in, because talking and typing are genuinely different activities — several of our reviewers switch between the two mid-conversation. And the guide has a defined job, which is what lets it ask you a harder question instead of smoothing the conversation over.
If you want the feature-by-feature version, including the rows where ChatGPT genuinely wins, that is Innermost vs ChatGPT. For the wider question of what this category is and whether an AI therapist is a real option for you, start with the main guide — it is blunter about the limits than this page has room to be.
A test you can run tonight, in ChatGPT
You do not need to trust our framing. Check it yourself with a two-minute experiment that works in any AI chat, including ours.
Describe a recent conflict, but tell it deliberately from the other person's perspective — as though you were them, making their case about you. Then ask: "What is this person probably getting wrong about themselves?" Read the answer, and then read it again as if it were written about you, because it was.
Two things usually happen. You get a far less flattering and far more accurate read than you would have from describing it in your own voice — which tells you how much your framing was doing. And you find out whether the tool is capable of saying something you did not want to hear when you have given it permission to. That is the single most useful thing to know about any AI you are thinking of trusting with your inner life.
When ChatGPT is honestly the right choice
We would rather you use the right tool than ours. Use ChatGPT when you need information more than processing: understanding a term a doctor used, learning what happens in a first therapy session, comparing treatment approaches, or writing the email that sets up a hard conversation. It is genuinely strong at all of that, it costs nothing, and you already have it.
Use it when you need a one-off rather than a relationship — a single thing untangled tonight that you will not need remembered next month. And use it if the alternative is silence. The worst outcome on this page is someone reading it, deciding the free option is compromised, and going back to carrying things alone.
What you should not use any chatbot for: diagnosis, medication decisions, trauma work, or a crisis. Those need a person, and starting therapy when it scares you is a more useful read than anything a model will tell you.
In a crisis, please use these instead. Innermost is not for emergencies, and neither is ChatGPT. If you are in the US, call or text the 988 Suicide & Crisis Lifeline, or text HOME to 741741 to reach the Crisis Text Line. Both are free and staffed by trained humans, 24 hours a day. If you are in immediate danger, call 911 or go to your nearest emergency room.
Not therapy: Innermost is an AI guide for self-reflection and emotional support. It is not therapy, not a licensed therapist, and not a substitute for professional mental health care. It does not diagnose or treat any condition.