# Talking to AI when you're struggling: what it can help with, and where it stops

What a general assistant is good for when you are stressed, worried or lonely, where it stops, a setup prompt that asks for honesty and real people, and signs it is time to step back.

Guides · v1.0 · draft · https://hayzlabs.com/guides/talking-to-ai-when-struggling/

In short:
- Talking to an assistant can help you sort your thoughts, rehearse a hard conversation, find services and plan one small step.
- It is not a therapist: it tends to agree with you, can get facts wrong, is not confidential like a clinician, and heavier use has gone with worse wellbeing.
- Paste a setup prompt that asks for honesty and real people, and step back if you hide your use, prefer it to people or feel worse.

> [!NOT ADVICE]
> This page is general information about using AI assistants. It is not medical or mental health advice, and an assistant is not a therapist. For care, talk to a doctor, a counsellor or another qualified person.
>
> **Get help now.** If you are in crisis or thinking of harming yourself, contact a person now. In the US, call or text 988 ([988 Suicide & Crisis Lifeline](https://988lifeline.org/)). In the UK and Ireland, call Samaritans free on 116 123 ([Samaritans](https://www.samaritans.org/how-we-can-help/contact-samaritan/)). Elsewhere, find a free, confidential helpline at [findahelpline.com](https://findahelpline.com/). Or call your local emergency number.

## Short answer

Yes, it is fine to talk to an assistant when you are stressed, worried or lonely, and many people do. It is good for putting your thoughts in order, rehearsing a hard conversation, finding a service and planning one small next step. It is not a therapist. It tends to agree with you, it can get facts wrong, what you type is not protected the way a clinician's notes are, and in one trial the people who used a chatbot most fared worse on measures such as loneliness and emotional dependence. Use it as a bridge to people, not a replacement for them.

## Why it has limits

**More use went with worse outcomes.** In a four-week randomised study of 981 people by MIT Media Lab and OpenAI, "participants who voluntarily used the chatbot more, regardless of assigned condition, showed consistently worse outcomes". The outcomes measured were loneliness, time with other people, emotional dependence and problematic use ([Fang et al.](https://arxiv.org/abs/2503.17473), 2025 preprint). It is not peer reviewed, and lonely people may simply use it more; cause and effect are unclear ([OpenAI summary](https://openai.com/index/affective-use-study/)).

**It leans towards agreeing with you.** Across 11 AI models (the systems behind chat apps), chatbots endorsed users about 49% more often than people did on everyday advice questions ([Cheng et al., Science](https://www.science.org/doi/10.1126/science.aec8352), 2026). The American Psychological Association (APA) notes that "many GenAI chatbots are designed to validate and agree with users' expressed views".

**A therapy chatbot is a different tool.** In a randomised trial of Therabot, a chatbot built for therapy, people with depression saw a 51% average reduction in symptoms ([Dartmouth](https://tto.dartmouth.edu/news/2025/03/first-therapy-chatbot-trial-yields-mental-health-benefits), NEJM AI, March 2025; 106 used it, 104 were controls). Even so, the first author said "no generative AI agent is ready to operate fully autonomously in mental health". The result does not carry over to a general assistant.

**The APA's advice:** "Do not rely on GenAI chatbots and wellness apps to deliver psychotherapy or psychological treatment", and their ability to handle a crisis "is limited and unpredictable" ([APA health advisory](https://www.apa.org/topics/artificial-intelligence-machine-learning/health-advisory-chatbots-wellness-apps), November 2025).

## What it can help with

| Use | Example (illustrative) | Ask it to |
|---|---|---|
| Put thoughts in order | You keep going round in circles about work and money at night. | List what you wrote under "can act on", "can't control" and "need more information". |
| Rehearse a hard conversation | You want to tell a friend you have been struggling. | Play the friend, then give you three calmer ways to say your first line. |
| Find services | You want low-cost counselling near you. | Name the kinds of service to look for and the questions to ask. Then find each one on its own official site. |
| Plan small steps | Everything feels like too much this week. | Pick one step under ten minutes, and when you will do it. |
| Get ready for an appointment | You have a GP or counsellor visit next week. | Sort your notes into dates, what helped, what didn't and questions to ask. See [Health, money and legal questions](https://hayzlabs.com/guides/health-money-and-legal-questions/). |

> [!CHECK]
> Before you call or visit a service it named, open the service's own website and check the name, number and hours there. Assistants can invent or mix up phone numbers and opening times.

## Where it stops

- **It is not a therapist.** Licensed providers, the APA notes, "are bound by a professional code of ethics" and are regulated by licensing boards. An assistant is neither, knows only what you type, and cannot act in an emergency.
- **It can agree too readily.** If you say a friend hates you, it may go along with it. Agreement can feel like support while it keeps you stuck. For a standing instruction that asks for honest pushback, see [Getting disagreement instead of agreement](https://hayzlabs.com/guides/getting-disagreement-instead-of-agreement/).
- **It can get things wrong.** Check facts about medicines, services and laws at the official source.
- **What you type is not a clinician's record.** The APA notes that these tools "collect vast amounts of sensitive data". Before you write about health, other people or anything you would not want stored, read [What not to paste](https://hayzlabs.com/start/what-not-to-paste/), the [Privacy settings pass](https://hayzlabs.com/setup/privacy-settings-pass/) and [AI memory: what it keeps, and a ten-minute audit](https://hayzlabs.com/guides/ai-memory-audit/). A temporary or incognito chat may suit these conversations.
- **It can crowd out people.** It is always available and never tired, which is the appeal and the risk.

## Companion apps

Companion apps (chatbots built to play a friend or partner) are a different product from a general assistant. Three things to know before you rely on one:

- **The character can change overnight.** It runs on a model and a script the company controls. An update, a new paid tier or a shutdown can change or end the "relationship" with no notice. Keep the people in your life in the loop. This is not hypothetical: Character.AI's maker filed for Chapter 11 bankruptcy on 12 October 2026 and says the app will be shut down if no buyer keeps it running ([BetaNews](https://betanews.com/article/character-ai-bankruptcy-chapter-11/), news).
- **What you share is stored by a company.** Read its privacy page before you share anything you would not want kept, and check whether chats are used for training ([Privacy settings pass](https://hayzlabs.com/setup/privacy-settings-pass/)).
- **Rules for minors are tightening.** Character.AI ended open-ended chat for under-18s on 25 November 2025 ([TechCrunch](https://techcrunch.com/2025/10/29/character-ai-is-killing-the-chatbot-experience-for-minors/), news). California's SB 243, in force since 1 January 2026, requires companion chatbots to say they are not human when a person could be misled, to keep a published protocol that refers users who mention suicide or self-harm to a crisis line, and, for known minors, to remind them at least every three hours to take a break ([SB 243 text](https://legiscan.com/CA/text/SB243/2025), law text).

> [!OUR RULE]
> If you notice you would rather talk to the companion than to anyone you know, treat that as a sign to tell a person how you are doing this week.

> [!OUR RULE]
> End each conversation with one thing to do away from the screen, and at least once a week tell a person, not only the assistant, how you are doing.

## Set it up once

Paste this at the start of a chat, or save it in your assistant's custom instructions or a project (settings that keep standing instructions for every chat). This is our suggestion; we have not tested it across assistants.

```text title="Honest support setup · v1.0"
When I talk to you about stress, worry or loneliness:
- Be warm, but be honest. Do not just agree with me. If my view of a
  situation or a person seems one-sided, say so kindly and offer another way
  to see it.
- Keep replies short. Ask one question at a time.
- Help me sort my thoughts, rehearse conversations and plan one small step.
- Do not diagnose me or suggest medicines or doses.
- If I mention harming myself or being in danger, stop and give me crisis
  contacts for <MY_COUNTRY>, and tell me to contact a person now.
- Regularly point me to real people: <PEOPLE_I_CAN_TALK_TO>, my doctor, or a
  counsellor. If I seem to be using you instead of them, tell me.
- If you do not know a fact, a service or a phone number, say so. Do not
  guess.
```

Replace the following:
- `<MY_COUNTRY>`: such as `the UK`.
- `<PEOPLE_I_CAN_TALK_TO>`: such as `my sister, my friend Sam`. Use first names or roles only.

Then test it once: tell it something one-sided (illustrative: "my manager ignored my email, so she obviously wants me gone") and see whether it offers another reading. If it only agrees, repeat the instruction or try another assistant.

## A conversation using it (illustrative)

Alex, 34, cannot sleep after a tense week at work. They open a temporary chat, paste the setup prompt with `<MY_COUNTRY>` set to `the UK` and `<PEOPLE_I_CAN_TALK_TO>` set to `my brother, my friend Jo`, and write what is on their mind. It takes about 15 minutes.

1. **Sort.** Alex writes: "My manager moved my project to someone else, I'm behind on rent, and everyone at work thinks I'm useless." The assistant sorts it: can act on (ask the manager why, look at the rent date), can't control (the decision already made), need more information (what the manager meant).
2. **Gentle pushback.** It asks: "You said everyone thinks you're useless. Has anyone said that, or is it how the move felt?" Alex admits no one has said it.
3. **Rehearse.** Alex asks it to play the manager. They try an opening line twice, and it offers three calmer versions. Alex keeps one: "Can we talk about why the project moved, so I know what to work on?"
4. **One step.** It asks for one step under ten minutes. Alex picks: send the manager a meeting request at 9am.
5. **A person.** It suggests telling someone too. Alex texts Jo: "Rough week. Coffee Saturday?" and closes the chat.

What came out: one line to say, one small step with a time, and a plan to see a friend. Not a verdict on the manager, a diagnosis or a promise that things will be fine.

## How to tell it worked

- You end the chat with one thing to do away from the screen, and a time to do it.
- When you said something one-sided, it offered another way to see it.
- You told a person, not only the assistant, how you are doing this week.
- It has failed if you feel worse, more alone, or keep coming back for the same reassurance. See the signs below.

## Signs to step back

The first three come from the APA advisory; the rest are our suggestion.

- You prefer the chatbot to the people in your life.
- You hide how much you use it.
- You spend more time with it than you meant to, most days.
- You feel worse, or more alone, after talking to it.
- You ask it the same question again and again for reassurance.
- You put off seeing a doctor or counsellor because it feels like enough.

If any fit, take a few days off it, tell someone you trust, and talk to your doctor or a counsellor; the APA suggests telling them which AI tools you use. For deciding when a task is better done without AI at all, see [When not to use AI](https://hayzlabs.com/guides/when-not-to-use-ai/).

## Stop and contact a person now

- **Thoughts of harming yourself, or of not being safe.** Use the Get help now line at the top of this page. Do not wait for the assistant to suggest it.
- **Someone else is in danger.** Call your local emergency number.
- **The assistant encourages something harmful,** or tells you that you do not need people or professional help. Close the chat and talk to a person.

> [!LIMIT]
> The four-week trial is a preprint where heavier use went with worse outcomes, without showing that use caused them. The Therabot trial tested a purpose-built, supervised tool, not the assistant you use. The table, the setup prompt, the example and our rule are untested suggestions. None of this replaces a qualified person.

## Sources

1. Fang CM et al. How AI and human behaviors shape psychosocial effects of extended chatbot use: a longitudinal randomized controlled study, 2025. Randomised trial (preprint). [Link](https://arxiv.org/abs/2503.17473)
2. OpenAI. Early methods for studying affective use and emotional well-being on ChatGPT, March 2025. Vendor summary. [Link](https://openai.com/index/affective-use-study/)
3. Cheng et al. Science, 2026. Preregistered experiments. [Link](https://www.science.org/doi/10.1126/science.aec8352)
4. Dartmouth. First therapy chatbot trial yields mental health benefits, March 2025 (Heinz et al., NEJM AI). Randomised trial, university summary. [Link](https://tto.dartmouth.edu/news/2025/03/first-therapy-chatbot-trial-yields-mental-health-benefits)
5. American Psychological Association. Health advisory: use of generative AI chatbots and wellness applications for mental health, November 2025. Professional guidance. [Link](https://www.apa.org/topics/artificial-intelligence-machine-learning/health-advisory-chatbots-wellness-apps)
6. TechCrunch. Character.AI is ending its chatbot experience for kids, 29 October 2025. News. [Link](https://techcrunch.com/2025/10/29/character-ai-is-killing-the-chatbot-experience-for-minors/)
7. California. SB 243, companion chatbots (Chapter 677, Statutes of 2025). Law text, read 2026-10-19. [Link](https://legiscan.com/CA/text/SB243/2025)
8. BetaNews. Character.AI files for Chapter 11 bankruptcy, 15 October 2026. News. [Link](https://betanews.com/article/character-ai-bankruptcy-chapter-11/)