ChatGPT as a therapist: what the research actually tests
ChatGPT can help you put a worry into words, examine an interpretation or prepare for a conversation with a therapist. Those are useful tasks. A fluent, sympathetic reply on its own doesn't establish that a chatbot can provide psychotherapy.
This article looks at what the research actually tests and what has changed in ChatGPT. For a practical reflection prompt, see a way to structure a self-help conversation. If you're choosing a service, the separate Nearby and ChatGPT comparison covers features, access and data terms.
A useful conversation needs a specific task
At three in the morning, typing can feel easier than trying to explain everything out loud. You can start with an unfinished thought. A language model can help organize it, ask a follow-up question or describe a familiar exercise such as a thought record.
The distinction is in the task you give it. "Help me separate what happened from what I assumed" gives you something to inspect. "Tell me what is wrong with me" asks the system to make an assessment from a few messages. The same conversational interface can make these requests look much more alike than they are.
A reply is useful when you can point to what it helped you do: name the feeling, write down an alternative explanation, or prepare a question for your next appointment. Feeling understood is valuable too. It isn't the same outcome as a sustained change in symptoms.
What the studies measured
HAILEY offers a good example of a defined task. It suggested wording changes to human peer supporters on TalkLife. In a non-clinical randomized study with 300 supporters, AI-assisted responses had a 19.6% higher expressed-empathy score than responses written without that feedback. This measured the wording of supportive replies, rather than patient recovery, and the people responding were peer supporters, not a sample of treating clinicians. Sharma and colleagues, 2023.
That finding supports a particular role for AI: helping a person write a more empathic response. It doesn't turn every chatbot with a warm tone into an effective therapist.
Other research tests where models get a conversation wrong. Moore and colleagues, 2025, evaluated models including GPT-4o with mental-health scenarios. They found stigmatizing and inappropriate responses, including agreement with delusional beliefs. These were tests of particular systems and versions; they shouldn't be presented as measured failure rates for every current ChatGPT conversation.
The Stade roadmap, 2024, proposes staged development and evaluation of clinical LLMs. It's a research proposal, rather than a trial showing that a general assistant delivers treatment. If an app cites a study, check which system was tested and what the researchers measured. Our MIND study breakdown shows how much an apparently simple percentage can leave out.
Memory and responses to distress have changed
The claim that ChatGPT starts from zero in every session is outdated. OpenAI documents memory that can use relevant context from past chats and other available sources when enabled. Controls vary with the account and workspace. Memory can personalize the next reply, but a remembered detail isn't a checked treatment record. OpenAI memory documentation.
Data controls are a separate choice. Turning off "Improve the model for everyone" stops new conversations from being used to train OpenAI models, without deleting those chats. Temporary chats don't create or update memories while they remain temporary; the personalization options available at the start determine whether they can use existing context. Check the controls in your account when deciding what to share. OpenAI data controls.
It's also inaccurate to say that ChatGPT has no measures for conversations about distress. In October 2025, OpenAI described updated model behavior, mental-health evaluations and guidance toward real-world support. Those are the provider's reports about responses and evaluations, rather than clinical treatment outcomes. OpenAI's account of the changes.
Compare the actual products
A specialized label tells you what an app is intended for. It doesn't guarantee a tested clinical protocol, complete memory or contact with a human specialist. Those features need to be checked for each product.
| What to check | General assistant such as ChatGPT | App focused on emotional support |
|---|---|---|
| Conversation structure | You can set the task and format with a prompt | May offer a guided flow; inspect what it actually does |
| Continuity | Memory and history depend on available settings | Storage and personalization depend on the app |
| Evidence | Research on a model applies to the tested version and task | A CBT-inspired design still needs its own outcome evidence |
| Human support | Review the provider's documented options | Check whether there is an actual human service or only information about finding one |
Nearby offers support chats and structured self-reflection using CBT-informed questions. The published product facts explain its features and the evidence available. Nearby wasn't evaluated in the studies above, and its terms don't promise automatic crisis detection or contact with a specialist.
FAQ
Can I discuss personal worries with ChatGPT?
Yes. A specific reflection task gives you a useful starting point. Ask it to help describe an event or consider another interpretation, then check whether its account fits what happened.
Does a chatbot remembering me make it a therapist?
Memory can save you from repeating context. Psychotherapy also involves assessment, an agreed approach and responsibility for care; recalling past messages doesn't establish those things.
Should I choose a specialized app?
Choose around the conversation you want and the features you can verify. Start with the features and data comparison, rather than assuming that a specialist label proves better outcomes.
If there's immediate danger of self-harm, contact local emergency services or a crisis service directly.
Sources
- Sharma and colleagues (2023): Human–AI collaboration in peer support.
- Moore and colleagues (2025): Stigma and inappropriate responses in tested LLMs.
- Stade and colleagues (2024): Development and evaluation of clinical LLMs.
- OpenAI: Memory in ChatGPT.
- OpenAI: Data controls.
- OpenAI: Responses in sensitive conversations, October 2025.
Product documentation checked on October 9, 2026.