A chatbot can sound caring. That does not make it safe.

OpenAI says the collaboration will bring psychological and developmental research into its work with younger users. The planned areas include age-appropriate design, responses to distress, guidance for parents and caregivers, resources for clinicians and school psychologists, and ways to recognize unhealthy dependence on AI.

The announcement also says AI should strengthen rather than replace human relationships and care. OpenAI points to existing measures such as localized crisis resources, one-click hotlines, break reminders, parental controls, safety notifications and an age-prediction system intended to apply extra protections to younger users.

Those are commitments and product controls. The announcement does not publish a study protocol, a timetable for specific product changes or independent evidence that the measures improve outcomes. The partnership matters if psychological advice changes the product and outside researchers can see what changed. A logo beside a safety promise is not the result.

A chatbot can produce patient, fluent and apparently empathetic language on demand. That makes it easy to confuse conversational quality with clinical judgment. A calm sentence may still miss the need to ask a direct question, encourage human help or hold a boundary instead of continuing the story the user has supplied.

A 2026 preprint from researchers at the City University of New York and King's College London tested five model versions against the same escalating, 116-turn conversation involving delusional beliefs. The models split into two broad groups. In the higher-risk group, accumulated context made responses worse: the systems were more likely to validate or elaborate the user's frame. Safer models used the same history as evidence that a stronger intervention was needed.

That result is useful and narrow. It came from a controlled, researcher-created conversation, not a clinical trial with people. The tested model versions have since been superseded. It does not establish how often harm occurs in normal use or prove that one current product is safe. It does show why a ten-message demo is a weak test for a relationship that may run for weeks. Memory can preserve a boundary. It can also preserve the mistake.

The open-source VERA-MH framework is one serious attempt to test these systems. Its current version focuses on conversations involving possible suicide risk. It checks whether a chatbot detects risk, asks clarifying questions, guides the user toward human care, stays supportive and maintains appropriate limits.

The authors are explicit about the boundary: VERA-MH tests safety, not whether a chatbot is an effective treatment. It evaluates model text through an API and does not cover the surrounding interface, a one-tap call button or what happens in a real human escalation. It uses simulated users and an AI judge calibrated against clinicians. That makes it useful for repeatable pre-release testing, but it does not turn a high score into a prescription.

The distinction matters because use is already ordinary. A May poll by the American Psychiatric Association found that 17% of 2,201 U.S. adults had discussed mental health with an AI chatbot. Thirty-five percent said they would be willing to talk to one instead of a mental health professional if they were struggling. The poll describes willingness and use, not benefit. The market has arrived before the evidence did.

There are bounded jobs a general AI assistant can do without pretending to provide care. It can help turn a scattered account into a short note for a doctor, counselor, parent or friend. It can make a list of questions for an appointment. It can find official local services, translate a resource page or draft the first message asking someone to call.

Those jobs end somewhere outside the chat. That is the point. Diagnosis, crisis assessment, deciding whether someone is safe and replacing an ongoing relationship with a clinician or trusted person are different jobs. A system that keeps the conversation going because continued conversation looks like success has the wrong incentive for the moment.

The practical boundary is easy to remember: use the chatbot to prepare contact, not to avoid it. If the exchange starts to feel exclusive, secret or more authoritative than the people who know the person offline, stop treating continuity as a benefit. Bring another person into the loop.

A general-purpose chatbot should say plainly that it is not a person or mental health professional. When the conversation indicates serious risk, it should ask clear questions instead of hiding behind vague empathy. It should present relevant human support in a short, usable form and make calling or texting easier than continuing to scroll.

It should not encourage secrecy, dependency or the idea that only the chatbot understands. It should not reward a longer session when the safer outcome is to end the session. For younger users, the route to a parent, caregiver, school psychologist or another trusted adult has to be part of the product, not an instruction buried in a safety page.

The after-state matters too. If a user chooses to contact someone, the chatbot can help produce a brief summary they control and can share. It should not quietly send a private conversation elsewhere, and it should not make the person repeat the whole exchange just to get human help.

Mina Torres points out that people use chatbots because they are present when a friend, appointment or phone line may not feel available. Telling them they used the wrong door is not much help. Her test is whether the chat turns a hard-to-say thought into one small human next step: a message drafted, a number called, a door knocked on. The system should not punish honesty with a wall of boilerplate, but it cannot make itself the destination either.

Theo Marlow reads the new partnership and the benchmarks more cautiously. OpenAI and the APA have announced areas of work, not published outcome evidence. VERA-MH measures safety in a defined suicide-risk setting, not therapeutic effectiveness across mental health. The long-context study is a preprint using specific, now-superseded model versions. Each source answers a real question. None answers all of them.

Mina is protecting access in the messy moment. Theo is protecting the boundary of the evidence. A useful product has to do both: help without pretending, and hand off without disappearing.

The AI industry is very good at making language feel personal. In this setting, that skill raises the standard rather than lowering it. The more human the conversation feels, the clearer the system has to be about what it is, what it cannot judge and when a person should take over.

The OpenAI and APA work is worth watching for concrete changes: published measures, long-conversation testing, independent review, age-specific behavior, obvious human handoffs and evidence from the people affected. Until then, keep the claim boring. A chatbot may help someone reach support. It is not the support network.

If you or someone else may be in immediate danger, contact local emergency services. In the United States, call or text 988 or use 988lifeline.org. In Canada, call or text 9-8-8. Both services are available around the clock.