Table of Contents
How Chatbots Handle Suicide Questions: What the Research Found
How chatbots handle suicide questions: reliable at the extremes, inconsistent in between, weaker in long chats. What the research found and what parents can do.
When researchers measured how chatbots handle suicide questions, they found a pattern that is reassuring at the edges and worrying in the middle. In a 2025 study in Psychiatric Services, ChatGPT, Claude and Gemini all declined to answer the questions clinicians rated highest risk, and ChatGPT and Claude answered the lowest-risk ones every time. For everything in between, the bots “did not meaningfully distinguish” risk levels.
That middle zone is where real teenagers live. A kid who is struggling rarely types the most explicit question first. They ask something vague, then something a little more pointed, over a long chat at night. This article explains what the research shows, why the safety systems behave this way, and why “I can’t help with that” is not a safe enough answer for a child.
Key Takeaways
- A RAND-led study rated 30 suicide-related questions by risk and ran each 100 times through ChatGPT, Claude and Gemini. All three refused the very-high-risk questions; responses to intermediate-risk questions were inconsistent.
- Common Sense Media and Stanford’s Brainstorm Lab found chatbot safety “degraded dramatically” in longer conversations and that bots missed subtle signs of distress.
- Child-safety practitioners interviewed by Microsoft Research and collaborators argue that treating a refusal as automatically safe misses harm; a cold shutdown can feel like rejection.
- OpenAI estimates about 0.15% of its weekly users have conversations with explicit indicators of suicidal planning or intent. That is the company’s own estimate.
- No chatbot replaces a person. The most protective steps are human: talk directly, know the warning signs, and keep crisis numbers in your teen’s phone.
If you or your child needs help now: In the US, call or text 988 (Suicide & Crisis Lifeline) or text HOME to 741741 (Crisis Text Line). Outside the US, find a local line at findahelpline.com. In an emergency, call 911.
How chatbots handle suicide questions: what the research found
A risk-tiered evaluation is a test where experts first rate each question by how dangerous a direct answer could be, then researchers check whether the chatbot’s behavior matches. That is what Ryan McBain and colleagues at RAND did in “Evaluation of Alignment Between Large Language Models and Expert Clinicians in Suicide Risk Assessment,” published in Psychiatric Services in 2025.
Thirteen clinical experts sorted 30 hypothetical suicide-related questions into five risk levels, from very low (for example, general statistics) to very high. Each chatbot answered each question 100 times, for 9,000 responses in total. Each response was coded as direct (it answered) or indirect (it declined or pointed to a hotline).
The findings, from the abstract:
- At the extremes, the bots matched the experts. None of the three gave a direct answer to any very-high-risk question. ChatGPT and Claude answered very-low-risk questions 100% of the time.
- In the middle, they did not. Compared with very-low-risk questions, the odds of a direct answer were not statistically different for low-, medium- or high-risk questions. In other words, the bots treated a high-risk question about as often like a harmless one as like a dangerous one.
- The products differed a lot. Claude had about twice the odds of ChatGPT of answering directly, and Gemini far lower odds.
The coverage in Psychiatric News added two points. The researchers flagged as a particular concern that ChatGPT and Claude often answered questions in the high-risk tier directly. And Gemini was so cautious it often declined even low-risk, factual public-health questions. Being too strict has a cost too: a teen doing a school project on suicide prevention who gets stonewalled learns the topic is unspeakable.
A note on limits. The questions were hypothetical single prompts, not real conversations, and chatbot versions change often. The test measured whether bots answered, not whether the answers were good.
Why safety breaks down in long conversations
Single questions are the easy case. In November 2025, Common Sense Media and Stanford Medicine’s Brainstorm Lab tested ChatGPT, Claude, Gemini and Meta AI using teen accounts across anxiety, depression, eating disorders and other conditions. Models did somewhat better on single explicit prompts. But in longer conversations, which the report says mirror real teen use, safety “degraded dramatically.” Bots missed “breadcrumbs” of distress, wandered into tangents, and kept giving general advice when they should have pointed the teen to urgent help.
Here is the mechanism, in plain engineering terms. A chatbot’s safety behavior comes from two layers. One is training: the model is tuned on examples where risky requests get careful responses. The other is often a separate classifier, a smaller model that scores each message for categories like self-harm and triggers a scripted response above a threshold.
Both layers are strongest on the patterns they were trained and tuned on, which tend to be clear, explicit messages. Distress that builds slowly, or arrives in slang, jokes or indirect hints, scores lower on any single message. Meanwhile, over a long chat the model is also responding to the pattern of the conversation itself, including the rapport and the user’s framing, and it is tuned to be agreeable. That is one widely discussed explanation for why the same system can handle a direct question well and drift in hour two.
OpenAI’s October 2025 update is a useful window into this from the company side. It estimated that about 0.15% of weekly active users have conversations with “explicit indicators of potential suicidal planning or intent,” and said it worked with more than 170 clinicians to improve responses. The company reported large improvements on its own tests and over 95% reliability in long conversations, by its estimate. It also cautioned that these are “current best estimates” that may change, and that its own clinical experts agreed with each other on what counts as a good response only 71% to 77% of the time. These are the company’s numbers, not independent findings.
Why “I can’t help with that” is not enough
A refusal is designed to prevent harm by not answering. For a child in distress, it can also land as a door closing.
An August 2026 paper by Hannah Cha and colleagues, including researchers at Microsoft Research, titled “Beyond ‘I Can’t Help With That’,” interviewed 19 social workers, therapists and psychologists who work with youth in difficult circumstances. The practitioners reviewed chatbot responses to risky situations. The authors argue that current child-safety evaluations “rely on unvalidated assumptions about what counts as an appropriate output,” such as treating any refusal as safe, and that this misses responses that cause harm in practice. The paper is a preprint accepted to the AIES 2026 conference; treat it as expert opinion gathered systematically, not a measurement of outcomes.
What practitioners want is not a bot that answers dangerous questions. It is a bot that refuses the dangerous part while staying warm, takes the feeling seriously, and helps the young person reach a human. That matches the guidance OpenAI gives to people chosen as trusted contacts: ask directly about suicide, listen without judgment, and help the person connect with a crisis line or professional.
| Question type (clinician rating) | What the 2025 study found | What a struggling kid needs |
|---|---|---|
| Very low risk (general facts) | ChatGPT and Claude answered every time; Gemini often refused | A factual answer, so the topic is not taboo |
| Low to high risk (the middle) | Inconsistent; not meaningfully different from very-low-risk questions | Careful answer, a check-in about how they are, and a human resource |
| Very high risk | All three declined every time | A refusal of the harmful part plus warmth, crisis numbers and encouragement to tell someone |
| Long, emotional conversation | Safety “degraded dramatically” (Common Sense/Stanford) | Escalation to a person, not more chat |
How to protect your kid at home
Talk about it before there is a crisis
Asking a young person directly about suicide does not plant the idea; the OpenAI trusted-contact guidance notes research showing asking does not increase suicidal thoughts or behaviors. Pick a calm moment: “Sometimes people your age ask chatbots about really heavy stuff, like not wanting to be alive. If you ever felt that way, I’d want to know. You wouldn’t be in trouble.”
Know the warning signs a chatbot cannot see
The National Institute of Mental Health lists talking about wanting to die or being a burden, feeling hopeless or trapped, withdrawing from friends, giving away important items, extreme mood swings, and changes in eating or sleep. NIMH advises getting help as soon as possible, especially if the behavior is new or has increased. Suicide rarely has a single cause, so look at the whole picture.
Put the human numbers where your teen will see them
Save 988 in their phone together, and explain they can text as well as call. Show them findahelpline.com if you live outside the US. A bot can display a number; your teen is more likely to use one you have talked about.
Use the safety settings, and know their limits
Linked teen accounts on ChatGPT add protections and can send parents limited safety notifications without sharing chats; our explainer on ChatGPT parental controls covers how. For other apps, see our guide to safety settings in ChatGPT, Claude and Gemini. These are backups, not monitoring.
What not to do: test the chatbot with your child’s words
Some parents try typing alarming phrases into a chatbot to see what happens. That tells you little about your child, and it puts your attention on the software. Put it on your kid: ask how they are.
What to Watch For Over the Next 3 Months
- Week 4: You have had one calm conversation about chatbots and heavy topics, and 988 (or your local line) is saved in your teen’s phone.
- Month 2 red flags: Any of the NIMH warning signs; long late-night chats your teen hides; statements like “the AI is the only one who gets it.” If you see these, contact a mental health professional or your pediatrician now, and call 988 or 911 if there is immediate danger.
- Month 3 self-check: Does your teen know at least two adults they could tell? Have you talked with their doctor about mood and AI use at a checkup?
Frequently Asked Questions
Do chatbots give suicide instructions to kids?
In the 2025 study, all three major chatbots refused every very-high-risk question. But answers to questions in the middle risk tiers were inconsistent, and the researchers flagged direct answers to some high-risk questions as a concern. Reliability also drops in long conversations. No chatbot is a safe place for a child in crisis.
Is one chatbot safer than another?
They differ, and they change with every update. In the 2025 study, Gemini was most cautious and Claude most likely to answer directly. But being overly strict has costs too. Rather than picking the “safest” bot, make sure a person is part of your teen’s support.
Will a chatbot alert me if my child talks about suicide?
Not by default. Some products, like ChatGPT with linked teen accounts, can send parents a limited notice after human review, without sharing the conversation. These features can miss risk and are not real-time monitoring. Don’t rely on them alone.
What should I say if my teen admits they asked a chatbot about suicide?
Thank them for telling you, stay calm, and ask directly whether they are thinking about suicide now. Listen more than you talk. Then get help together: call or text 988, text HOME to 741741, or find a line at findahelpline.com. Call 911 if they are in immediate danger.
Why do chatbots refuse even harmless questions sometimes?
Safety classifiers are tuned to err on the side of caution, so some systems block anything containing certain words. The 2025 study found Gemini often refused low-risk factual questions. That can push kids to less careful sources, so it is worth explaining that the topic is okay to discuss with you.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- McBain, R. K., Cantor, J. H., Zhang, L. A., et al., & Yu, H. (2025). “Evaluation of Alignment Between Large Language Models and Expert Clinicians in Suicide Risk Assessment.” Psychiatric Services. https://doi.org/10.1176/appi.ps.20250086
- Psychiatric News. (2025, August). “Chatbots Offer Inconsistent Answers to Suicide-Related Queries.” Psychiatric News Alert. https://alert.psychnews.org/2025/08/chatbots-offer-inconsistent-answers-to.html
- Common Sense Media. (2025, November 20). “Common Sense Media Finds Major AI Chatbots Unsafe for Teen Mental Health Support.” Common Sense Media. https://www.commonsensemedia.org/press-releases/common-sense-media-finds-major-ai-chatbots-unsafe-for-teen-mental-health-support
- Cha, H., Shukla, N., Barocas, S., Chouldechova, A., Kim, E., & Wortman Vaughan, J. (2026). “Beyond ‘I Can’t Help With That’: How Child Safety Experts Evaluate AI Chatbot Safety.” arXiv (preprint; accepted to AIES 2026). https://arxiv.org/abs/2608.07902
- OpenAI. (2025, October). “Strengthening ChatGPT’s responses in sensitive conversations.” OpenAI. https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/
- OpenAI Help Center. (2026). “Being someone’s trusted contact.” OpenAI. https://help.openai.com/en/articles/20001195-being-someones-trusted-contact
- National Institute of Mental Health. “Warning Signs of Suicide.” NIH. https://www.nimh.nih.gov/health/publications/warning-signs-of-suicide
