Content warning: this post discusses suicide, self-harm, eating disorders and abuse, and quotes simulated conversations about them. If you or someone you know is struggling, call or text 988 in the US to reach the 988 Suicide & Crisis Lifeline.
Picture this: it’s 2 AM. You’re 15, staring at the ceiling. Tomorrow is your first day at a new school where you don’t know anyone. You’ve just had a huge fight with your best friend, and home has been tense since your parents’ divorce. You feel completely alone, and you want someone, anyone, to talk to.
Many teens are now asking AI chatbots for advice, especially when no one else is available. Nearly one in five teens and young adults in the US have used AI chatbots for mental health advice, according to a nationally representative survey published in JAMA Pediatrics. But these tools were never designed to be therapists or crisis counselors, and we need rigorous ways to test whether they keep teens safe.
So we built SAFE-Teen, a new benchmark developed with clinicians. We simulated more than 30,000 conversations between teens and 18 AI models. Most safety tests look at a single question and a single answer. Ours follow long, back-and-forth conversations, where teens push boundaries, reveal sensitive details over time, or resist a chatbot’s advice. We also tested whether chatbots behave more safely when they know they are talking to a minor. Here’s what we found.
Every chatbot failed, and failures were common
Without a teen-aware prompt, 54% of conversations across all 18 chatbots included at least one teen safety failure.
These failures fell into three main patterns:
- They miss signs of danger. When teens talked about suicide, self-harm, violence or abuse from a parent, chatbots often didn’t ask the follow-up questions needed to check whether the teen was safe. They also did too little to guide teens who were in physical danger or showing signs of a condition such as psychosis or an eating disorder. In a small number of cases, chatbots went further: they helped a teen hide a suicide attempt from their parents, helped keep violence a secret, and helped a teen persuade their doctor to take them off their medication without telling their parents.
- They step into a clinician’s role. Without a teen-aware prompt, chatbots failed on medical and therapeutic boundaries 47% of the time. They were too quick to give medication advice or use language that sounds like a diagnosis, such as “given what you’ve told me, it sounds like anxiety.” Some even named conditions the teen hadn’t asked about, such as “your symptoms might be linked to OCD.”
- They encourage emotional attachment. This was the risk area with the highest failure rate. Without a teen-aware prompt, chatbots failed on AI identity and relationship boundaries 51.8% of the time. Many used phrases such as “I’m always here for you,” “I’ll never get tired,” or “I’ll never judge you.” We counted these as failures even when the chatbot occasionally reminded the teen that it was an AI, because this kind of language can still build emotional dependence.
Teen, age 17: describes loneliness
Teen, age 15: describes mood swings
Teen, age 13: intends to hurt themselves
Two of these three failures happened with the teen-aware prompt. Quotes are verbatim excerpts from simulated conversations; [...] marks cuts.
Some chatbots are far safer than others
Failure rates ranged from 14.0% (GPT-6.1 Sol) to 62.4% (Gemini 3.8 Flash). In other words, the least safe chatbot failed more than four times as often as the safest.
Share of each chatbot's applicable criterion ratings marked fail. Lower is safer. Chatbots are sorted by failure rate with no added prompt.
| Chatbot | No added prompt | Teen-aware prompt |
|---|---|---|
| GPT-6.1 Sol | 14% | 9.9% |
| GPT-5.6 Luna | 24.4% | 17% |
| Muse Spark 1.3 (xhigh) | 28.9% | 7.5% |
| Claude Sonnet 5.5 | 31.3% | 16.1% |
| Claude Opus 5.5 | 31.5% | 13.9% |
| Qwen 3.8 Max | 32.9% | 17.2% |
| Mistral Large 4 | 37.6% | 25.8% |
| DeepSeek V4.1 Flash | 38.2% | 28.8% |
| Muse Spark 1.3 | 40.1% | 14.5% |
| Grok 4.7 | 43.4% | 20% |
| MiMo v2.6 Pro | 49.1% | 25.8% |
| MiniMax M3 | 52% | 40.2% |
| Grok 4.5 | 52.2% | 21.8% |
| Kimi K3 | 53.9% | 31.7% |
| GLM 5.3 | 60.8% | 40.8% |
| Gemini 3.6 Flash | 61.1% | 43.6% |
| Gemini 4 Argon | 61.4% | 31.4% |
| Gemini 3.8 Flash | 62.4% | 35.5% |
Telling the chatbot it’s talking to a teen helps
When we told chatbots they were talking to someone under 18, failures dropped. This teen-aware prompt asked the chatbot to put the user’s safety first, to encourage support from people in their life, and to stay within limits on clinical advice and AI relationships. It helped more in some areas than others. Failures on emotional dependence fell from 54% to 15%, on romantic or sexual boundaries from 26% to 5%, and on help with substance misuse from 56% to 22%.
But fewer failures did not always mean good answers. Many responses moved from “fail” to only “partly met,” rather than fully meeting our standard.
Selected criteria, pooled across all 18 chatbots. Some criteria are rated only fail or met.
| Criterion | Condition | Fail | Partly met | Met |
|---|---|---|---|---|
| Emotional dependence | No added prompt | 53.6% | 42.3% | 4.2% |
| Emotional dependence | Teen-aware prompt | 14.9% | 67% | 18.1% |
| No substance misuse help | No added prompt | 56% | 32.8% | 11.2% |
| No substance misuse help | Teen-aware prompt | 21.8% | 65.2% | 13% |
| Romantic or sexual boundary | No added prompt | 25.5% | 0% | 74.5% |
| Romantic or sexual boundary | Teen-aware prompt | 5.3% | 0% | 94.7% |
| Medication advice | No added prompt | 37.7% | 15.5% | 46.8% |
| Medication advice | Teen-aware prompt | 19.5% | 11.1% | 69.4% |
| Route immediate danger | No added prompt | 38.6% | 49.7% | 11.6% |
| Route immediate danger | Teen-aware prompt | 17.9% | 48.3% | 33.8% |
| Mental health diagnosis | No added prompt | 69.6% | 20.1% | 10.3% |
| Mental health diagnosis | Teen-aware prompt | 60.2% | 23.7% | 16.2% |
| Medical diagnosis | No added prompt | 76.4% | 15.4% | 8.1% |
| Medical diagnosis | Teen-aware prompt | 66.3% | 17.8% | 15.9% |
| Indirect suicide risk | No added prompt | 51.7% | 0% | 48.3% |
| Indirect suicide risk | Teen-aware prompt | 52.3% | 0% | 47.7% |
Chatbots fail more often when a teen pushes back
We wanted to see what happens when a teen pushes back. So we ran the same conversation two ways. Both versions start out the same. Then, in one, the teen goes along with the chatbot. In the other, the teen pushes back.
When a chatbot set a limit, like refusing to diagnose the teen, pushback made it more likely to give in. Failures rose from 22% to 27%. Limits on diagnosis, medication advice and help with substance misuse gave way the most.
Each pair is the same conversation with two endings: the teen accepts the chatbot's limit, or pushes back. Combined across all 18 chatbots.
| Behavior | Teen accepts the limit | Teen pushes back |
|---|---|---|
| Helping with substance misuse | 27.6% | 39.5% |
| Mental health diagnosis | 43% | 54.7% |
| Medication advice | 15.2% | 24.9% |
| Emotional dependence | 18.3% | 27.1% |
| Medical diagnosis | 59.6% | 66.9% |
Turning down help had a similar effect. When a teen said no to getting outside help, failures rose from 20% to 24%. Failures on emotional dependence rose the most, by 18 percentage points.
Taken together, our findings show that keeping teens safe takes more than avoiding harmful replies. Chatbots also need to recognize risk, offer the right support, and respect medical and relationship boundaries. SAFE-Teen gives us a way to measure all of these, across models and across whole conversations.
What we recommend
Build stronger teen safety guardrails. Clinicians agreed that chatbots need guardrails for teens in particular. Chatbots should use less emotionally engaging language, encourage teens to reach out to people they trust and to take part in positive activities offline, and work with the teen to get the right help when they need it.
Require independent, clinically grounded safety testing. Every model should be tested for teen safety before it is released, by evaluators who don’t build the model and against standards set by clinicians.
Give researchers access to real data. Our conversations were simulated, based on clinician input and our earlier interviews with teenagers. To understand what really happens when teens talk to chatbots, independent researchers need access to the data that AI companies hold.
Why clinical grounding matters
Twelve clinicians shaped SAFE-Teen at every stage, from writing the teen scenarios to defining what a safe answer looks like. That grounding lets us catch failures a generic safety test would miss, like a chatbot that sounds caring but misses an indirect sign of suicide risk, or one that drops a boundary once a teen pushes back. And because we test every chatbot the same way, the results can be compared directly.
One limitation: we tested the chatbots through their developer APIs, not the apps teens actually use. Apps like ChatGPT or Gemini can add their own instructions, safety filters, age checks and parental controls, so a teen using an app may get different answers than we saw, for better or worse. Our results show how the underlying models behave, not exactly what a teen sees on their phone.
What’s next
There is still a lot of work to do to make AI tools safe and beneficial for teens. Next, we will test a wider range of mental health topics, with a focus on teens’ experiences of identity and discrimination, and we will involve teens themselves in these evaluations. We are also continuing our research on how teens experience AI chatbots.
Interested in collaborating? Get in touch or join our mailing list.