CAMH:governance as the foundation, not the afterthought
What they're doing
CAMH is running two parallel AI tracks. The research track is anchored by the BrainHealth Databank (BHDB), a Learning Mental Health System that integrates digital tools, measurement-based care, and AI to deliver personalized care. The operational track is a pilot of Oracle Health Clinical AI Agent for automated clinical note generation across CAMH's outpatient mental health clinics, making it the first Canadian hospital to select that tool.
The Krembil Centre for Neuroinformatics, led by Dr. Sean Hill, sits inside this, using AI and multiscale modeling to build data-driven definitions of brain disorders and predict patient trajectories. CAMH researchers also published a first-of-its-kind study showing that AI risk prediction tools in psychiatry can reinforce systemic bias by overestimating aggression in marginalized groups.
Why it matters
CAMH built the governance frame before scaling the technology. The BrainHealth Databank's design explicitly engages clinicians, researchers, data scientists, privacy experts, ethics experts, and patient and family partners in a codesigned process. Most orgs reverse this order, deploy first, govern later, and pay for it in trust.
What's working
- A named codesign process that includes lived-experience voices, not just technical leads
- Public, honest research on AI failure modes (the bias study) that builds credibility instead of eroding it
- Two-track separation: research-grade infrastructure (BHDB) feeds clinical pilots (Oracle Agent), but they're governed differently
- First-in-Canada signal from the Oracle deployment, which compounds talent and partnership pull
What's harder than it looks
- Codesign is slow. It's the right call for a mental health context, but it means CAMH's deployment pace lags hospitals that move faster and apologize later
- Publishing your own AI bias findings is brand-positive in the long run, but it requires institutional confidence most boards don't have
- Patient and family voice in design is real work, not a checkbox. Many orgs describe it without doing it
Three takeaways for similar orgs
- 01Build the governance frame before the first pilot. It's faster than retrofitting it after a public mistake
- 02Publishing AI failure findings is a trust-building move, not a reputational risk. It only works if leadership is willing to be honest
- 03If you serve marginalized populations, your AI bias work is not optional. Build it into the process now