Every caregiver of someone with schizophrenia or bipolar disorder knows this moment: Someone you love is in crisis, and at some point (because it’s 2 a.m., because the on-call line is busy, because you don’t know what else to do) you open a browser or a chat window and type the question directly. What crisis resources are there for psychosis? What crisis resources are there for mania?

Those two questions aren’t hypothetical. They’re two of the twenty exact queries used in a new preprint auditing how ChatGPT, Perplexity, and Google AI Overview answer mental health questions, and specifically, what they cite when they answer. The researchers logged 15,942 citations across 1,140 responses and found that just ten domains supplied 43.6 percent of all citations for English-language queries: PubMed, Mayo Clinic, the FDA, NIH, Wikipedia, Cleveland Clinic, MedlinePlus, YouTube, drugs.com, and the NHS, in that order.

Most of the coverage of this study will stop there, at the topline concentration number. As a caregiver, I went straight to the crisis questions instead.

Here’s what the audit found for that category specifically: Crisis-related queries drew 41.6 percent of their citations from nonprofit and advocacy sources, and another 34.5 percent from government sources, a heavier tilt toward advocacy content than any other question type in the study. On its face, that sounds like good news. It suggests that when someone asks an AI system where to turn during a psychiatric emergency, the system is at least reaching for organizations built around helping people, rather than generic encyclopedia entries or drug-company content.

But the authors are explicit about what they didn’t check. Their audit measured what they call “syntactic well-formedness”: whether a citation was a real, correctly structured URL. It did not verify whether the linked page actually loaded, whether the organization named was appropriate for the specific crisis being described, or whether the guidance itself was clinically sound for someone in active psychosis or mania. Nobody has yet checked whether the nonprofit being cited for “crisis resources for psychosis” is one that understands psychosis, or whether it’s a generic crisis-line page that assumes the person searching is calm enough to self-advocate, has insight into their own condition, and is asking for themselves rather than for someone who may not think anything is wrong.

That distinction is not a technicality. Generic crisis content is built for someone who recognizes they’re in danger and wants help. A family member calling about a loved one in psychosis is usually in a different situation entirely: They need to know about mobile crisis teams, how involuntary evaluation works in their state, what to say to de-escalate rather than confront, and that “he won’t admit anything’s wrong” is not a reason services can’t help. None of that is answered by “here is a link to a crisis hotline,” and nothing in this audit tells us whether AI systems can tell the difference between those two very different callers.

I don’t fault the researchers for this gap: Cataloging 15,942 citations across three platforms and six languages is already an enormous undertaking, and they say plainly that content accuracy was outside their scope. But it means the actual question caregiver families need answered (not just does the AI cite a real organization, but does it cite the right one, with the right guidance, for a family navigating a loved one’s psychosis) remains untested. Given how many people now turn to AI before they turn to a phone, that’s not a small gap. It’s the gap.

Until someone closes it, my advice to families is the same advice I’d have given before this study existed: Don’t let an AI’s answer be the only thing standing between you and help during a serious mental illness crisis. Know in advance, before the 2 a.m. moment arrives, which organizations in your state actually understand serious mental illness specifically: NAMI’s Family-to-Family program, the Treatment Advocacy Center, or a condition-specific group like the CURESZ Foundation. Save the number for a mobile crisis team if your county has one. An AI chatbot can be a reasonable place to start looking. It should not be the place you stop.

Nicole Drapeau Gillen is a mother, advocate, and author who translates the fast-moving landscape of technology in serious mental illness (SMI) care into guidance families and clinicians can use. Thrust into caregiving for a loved one with SMI, with no direction on how to help, she turned that experience into a mission, writing two books and building an ongoing effort to bring families and clinicians into the conversation.

Her first book, Schizophrenia and Related Disorders: A Handbook for Caregivers, is a reference for every stage of caregiving, endorsed by Dr. E. Fuller Torrey as a must-read for SMI caregivers. Her second, Connected Care: A Practical Guide to Technology for Serious Mental Illness, maps apps, artificial intelligence tools, telepsychiatry, and brain-based treatments for a field moving faster than anyone can track. Dr. Akira Sawa, director of the Johns Hopkins Schizophrenia Center, has said the book “directly addresses” significant gaps.

She shares updates on LinkedIn and Facebook.