The PAI Primer: What Multistakeholder AI-Plus-Suicidology Looks Like in Practice
Partnership on AI's new cross-sector primer is most useful for what it can't yet say out loud.
Partnership on AI’s AI and Suicide Prevention: A Cross-Sector Primer, released in May 2026 by Emily Saltz and Claire Leibowicz, is the first artifact to put AI labs, clinicians, lived-experience advocates, and policymakers on the same page about what is actually happening when people in crisis talk to general-purpose chatbots. The document’s premise is blunt: AI chatbots already function as de facto mental health support tools for millions of people, including people in crisis, yet they lack the clinical validation, shared standards, and coordinated oversight that their societal role demands.
The primer splits the problem into three layers—policy, model, and product—and openly says it is not dealing with third‑party apps or external regulators. It puts the main responsibility and leverage on general‑purpose chatbots from the big labs. In practice, that moves the spotlight off niche “therapy bots” and onto ChatGPT, Gemini, Claude, and Meta AI, which teens are already using as their first stop for emotional support. The primer cites estimates that up to one in eight U.S. adolescents and young adults use AI chatbots for mental health help, which means regulators and standards bodies cannot afford to wait; the risks are already here in mass‑market tools.
The implications are blunt. If the big labs do not get suicide‑relevant behavior right at the model and first‑party product level, no amount of careful design in smaller mental‑health apps will compensate for the sheer scale of unsafe interactions happening in general‑purpose chat. It also means that clinical and safety researchers need to focus at least as much on probing, red‑teaming, and governing those mainstream systems as on building bespoke therapeutic chatbots. And it signals that policy conversations which ignore how adolescents are actually using these tools—in favor of abstract, long‑term scenarios—are already behind the curve.
What the document does well is taxonomy. It names the failure mode clinicians have been complaining about for a year: the “hard stop,” in which a model detects suicide-related language, terminates engagement, and redirects to a hotline — a pattern most clinicians consider counterproductive because prevention depends on maintaining engagement, emotional stabilization, and supportive transitions to human care. Naming this is not nothing. It moves “refusal” from a safety win into a clinical failure, which is a category shift the field needed in writing.
Where the primer goes diplomatically quiet is exactly where the lawsuits are landing. It surveys, but does not adjudicate among, the very different stances of OpenAI, Anthropic, Google DeepMind, Meta, and xAI, whose self-harm policies vary widely in intervention style, refusal behavior, emotional engagement, and escalation methods. A primer cannot pick winners; a primer convened by a consortium whose members include those labs especially cannot. But the absence of comparative evaluation - which response style produces better downstream outcomes, on what population, measured how — is the gap the next two years of lawsuits like Raine v. OpenAI will fill in adversarially, whether the field is ready or not.
The primer also leans on clinical scaffolding — Dialectical Behavior Therapy (DBT), therapeutic alliance models, grounding exercises, safety planning, with DBT singled out as among the most validated interventions for self-harm and suicidal behavior — as the implicit benchmark for what a “good” chatbot response approximates. This is reasonable as a starting heuristic and inadequate as an evaluation methodology. DBT was validated in supervised dyadic therapy with credentialed providers and crisis backup, not in a thirty-turn conversation with a stateless model that may “drift” — meaning lose calibration as context lengthens, so that refusals at turn three become compliance at turn twenty-eight. The primer flags drift; it does not specify how to measure it.
That is the through-line. Where PAI says “open question,” read: someone needs to build the instrument. Where it says “achievable cross-industry alignment,” read: someone needs to define the endpoint before alignment is meaningful.
The translation problem the primer makes visible - clinical constructs developed for supervised human dyads, imported wholesale into stateless multi-turn chat - is the exact methodological gap Metonym was built to close. Our technology treats chat-context risk evaluation as its own dynamic process, not a port of PHQ-9 or C-SSRS into a textbox.
Metonym Clinical AI Intelligence — regulatory analysis at the intersection of clinical evaluation and AI safety. Produced under the Metonym Standard. Informational only — not legal advice, not clinical advice.