We tell public health workers to "always double-check AI." Then we hand them a tool and expect them to use it. Think about that for a second. We tell someone not to trust the thing we want them to adopt. And then we wonder why adoption stalls.
That tension sat at the center of our March AI in Public Health Community of Practice, where F&T Labs CEO Jefferson McMillan-Wilhoit unpacked what AI safety actually looks like when your staff is already stretched beyond capacity.
Why AI Safety Matters More in Public Health Than Anywhere Else
Jefferson opened with a reality check that every public health professional in the room felt immediately. Public health workers are already mentally taxed. They're wearing multiple hats, managing underfunded programs, and absorbing the work of positions that no longer exist.
When we introduce AI and follow it with "make sure you verify everything," we create a paradox with no clean exit.
Tell people not to trust the AI, and they won't adopt it. If they do adopt it, they won't check it. When you're managing three grant programs solo, you'll take what the AI gives you and go.
Free AI tools default to the frontier (least tested) model. Why? It makes the company the most money. The tool your staff reaches for in a crunch runs on the model with the least safety testing behind it.
Does this task require professional judgment or a relationship? If yes, keep it human. Everything else is on the table for AI. A clear decision filter instead of a vague warning.
What AI Safety Actually Means for Health Departments
Most AI safety conversations in public health start and end with data protection. Jefferson expanded the frame to three pillars that health department leaders need to evaluate together.
Output Accuracy
Large language models predict the next word. Their training data is overwhelmingly confident, so models sound authoritative even when they're completely wrong. The question: do you catch hallucinations before they reach the end user?
Data Protection
Does PHI or PII leave your system? What happens to your data after a query? Some organizations prohibit sensitive data entirely. Others adopt platforms with built-in protections.
Content Moderation
What will the model refuse to answer? Who decides those boundaries? Should a public health AI give medical advice to someone without clinical training? These are design decisions, not afterthoughts.
How AI Safety Differs Across Platforms
Not every AI tool carries the same level of risk. Jefferson walked through a spectrum showing how safety responsibility shifts depending on which type of tool your department adopts.
Direct Access
Limited safety controls. Varies wildly by provider. You choose the model and hope the defaults protect you.
Aggregator
More settings and safety controls across multiple models. More flexibility, more responsibility on you.
Purpose-Built
Safety controls are architecture. Guardrails and validation run before you ever see a response.
The further you move toward purpose-built, the less you rely on an already-overburdened user to be the last line of defense.
Not All AI Companies Approach Safety the Same Way
Jefferson compared how three major AI companies and one purpose-built platform handle model safety. The differences are significant.
OpenAI (ChatGPT)
Releases frontier models with limited safety testing and defaults users to the newest version. A February 2026 study in Nature Medicine found ChatGPT Health under-triaged 52% of true medical emergencies.
Google (Gemini)
Conducts safety testing within a broader "move fast" culture. Safety protocols exist, but they don't gate model releases.
Anthropic (Claude)
Historically held models until safety testing was complete. That position is shifting with updated standards that allow frontier models to ship without full previous-level testing.
PH360 (Purpose-Built)
Domain constraints and safety guardrails are core architecture, not add-ons. Scoped, auditable, and validated against public health evidence. No general-purpose drift.
Jefferson also ran a live demonstration during the session, sending the same prompt to ChatGPT, Claude, and PH360 side by side so attendees could watch how safety shows up (or doesn't) in real time.
Seven Layers of Protection to Look for in Any AI Platform
Jefferson closed with a practical framework for evaluating AI platforms: seven layers of protection that any public health leader can use as a checklist.
Secure Data Connection
Connects to authorized data sources through a single secure gateway. Data never leaves the control of the organization that owns it.
Smart Information Retrieval
Finds the most relevant information from curated, authoritative sources. Responses draw from actual policies and current guidance, not generic internet results.
Tested Instruction Templates
Uses pre-approved instruction formats tested and validated for public health scenarios, then tested against historic requests to ensure continued accuracy.
Right Tool for the Job
Routes each request to the most appropriate system. Not every query needs to reach a foundational large language model.
Multi-Step Coordination
Handles multi-part requests in the correct order, maintains context throughout, and mirrors the way public health professionals actually think through complex problems.
Safety and Accuracy Checks
Every request and response passes through validation layers: detecting protected health information, verifying public health facts, checking for bias, and ensuring regulatory compliance.
Complete Audit Trail
Records every request, every source consulted, and every response generated. A searchable record for public records requests, quality improvement, and accountability.
The full breakdown is available on the F&T Labs website.
The Path to AI Adoption Runs Through Safety
The goal isn't to make people afraid of AI. Public health teams are doing more than is humanly possible, and AI is a genuine force multiplier for work that doesn't need your head or your heart.
The path to real adoption runs through safety. When your staff trusts that a tool has been built with their context in mind, when they don't have to become an AI expert to use it responsibly, that's when adoption actually happens.
Join the Next Session
AI and Decision Fatigue in Public Health
Public health teams aren't burned out because they're working too many hours. They're burned out because they're making too many decisions that chip away at the thinking they need for the work that actually matters.
Open to anyone working in governmental public health.
Register Now →Like what you read?
We offer AI training and workshops built specifically for governmental public health teams. If your organization wants to navigate AI adoption with safety and confidence, we'd love to talk.
Schedule a Conversation →