Health AI Is Getting Better. Literacy Still Matters.

What changed

On June 18, 2026, OpenAI said GPT-5.5 Instant is improving how ChatGPT handles health questions, including when to seek urgent care, how to ask for relevant context, how to communicate uncertainty, and how to make complex information easier to understand. OpenAI tied the update to physician-led evaluations such as HealthBench and HealthBench Professional.

The same day, OpenAI described an NEJM AI study where researchers from Boston Children's Hospital, Harvard University, and OpenAI used the o3 Deep Research reasoning model to revisit 376 previously unsolved rare-disease cases. After expert review, additional testing, laboratory confirmation, and clinical return of results, physicians established 18 diagnoses. The model generated evidence-linked leads; it did not diagnose patients or make clinical decisions.

On June 23, OpenAI published a research case study about immunologist Derya Unutmaz using GPT-5 Pro to revisit a three-year-old T-cell puzzle. The useful part for readers is not the specific biology alone. It is the workflow: an expert used AI to surface a plausible mechanism, then judged whether that suggestion mattered scientifically.

The useful distinction

Health AI is moving in two directions at once. One direction is everyday health information: people asking tools to explain symptoms, lab results, appointment notes, medication instructions, or what questions to ask a clinician. The other is professional and research support: clinicians and scientists using AI to organize evidence, draft documentation, inspect complex datasets, or generate hypotheses.

Those uses should not be treated as the same thing. A clearer answer to a health question can help someone prepare for care, but it is not the same as a diagnosis. A model-generated research lead can help a specialist look in a promising place, but it is not the same as a clinically confirmed finding.

Health Canada already uses a risk-based lens for software as a medical device, including software represented for medical purposes. That Canadian context is a useful reminder: the more a system influences diagnosis, treatment, or patient management, the more carefully people should ask who is accountable, what evidence supports it, and what oversight applies.

Why this matters for Saskatchewan

Saskatchewan readers will encounter health AI through many doors: public search habits, family caregiving, clinic paperwork, telehealth, student projects, hospital innovation, startup products, research labs, and workplace benefits. The province does not need to wait for a local headline to start building the vocabulary for responsible adoption.

The opportunity is real. Better health explanations can help people prepare for appointments, ask clearer questions, understand medical language, and advocate for themselves. In research and clinical administration, AI may help trained professionals move through literature, notes, and complex data more efficiently.

The risk is also real. Health questions often involve privacy, urgency, incomplete context, and emotional pressure. A confident-sounding answer can still be wrong, outdated, poorly localized, or inappropriate for a person's medical history. That is why AI literacy in health should include both practical use and practical limits.

A simple literacy checklist

For everyday health questions, ask whether the AI response explains uncertainty, points to urgent care when needed, encourages appropriate follow-up with a qualified professional, and avoids pretending it has the full medical record. If the answer sounds absolute, that is a reason to slow down, not speed up.

For health organizations and innovators, ask what the tool is actually for. Is it education, documentation, triage, diagnosis support, patient management, research, or administration? The answer changes the risk profile, the evidence needed, the privacy review, and who should be allowed to rely on it.

For students and community members, the key habit is traceability. Good health AI use should make it easier to ask better questions of reliable people and sources. It should not pressure someone to bypass care, hide sensitive information in an unapproved tool, or treat a model's answer as the end of the conversation.

What AiSK members should watch

Watch for health AI tools that are clear about their role. The best claims will distinguish patient education, clinician support, research assistance, and regulated medical-device use rather than blending them together.

Watch for human-reviewed workflows. The rare-disease and immunology examples are valuable because experts stayed responsible for interpretation, testing, confirmation, and judgment.

Watch for Canadian governance conversations. Privacy, procurement, professional standards, medical-device classification, Indigenous data considerations, and public trust will matter as much as model performance.

AiSK can help by convening health leaders, researchers, students, builders, public institutions, and community members around plain-language questions: where can AI help, where should it slow down, and what safeguards make adoption trustworthy?