Responsible AI Guidelines

Responsible AI for real public health work.

Backyard-AI helps teams use AI in ways that are narrow, reviewable, source-aware, and accountable to human judgment.

Explore Free Workflows

Book a Readiness Call

Intro

Public health teams do not need vague promises about “AI transformation.” They need practical workflows that make repetitive work faster without weakening trust, clarity, or accountability. Backyard-AI is built around that idea. The goal is not to replace expert judgment. The goal is to support it.

Core principles

The operating philosophy

  • AI should support public health work, not replace expert judgment.
  • Narrow workflows are safer and more useful than broad, ambiguous ones.
  • Important outputs need visible human review.
  • Claims should stay within the limits of the source material.
  • Early pilots should avoid sensitive data whenever possible.
  • Teams should know what the workflow will do and what it will not do.

What Backyard-AI is for

  • evidence synthesis support
  • health communication support
  • workforce training support
  • reporting and workflow support
  • internal knowledge and policy support

What Backyard-AI is not for

  • autonomous public health decision-making
  • clinical diagnosis
  • unchecked legal or policy interpretation
  • hands-off expertise
  • replacing review by qualified professionals

Safeguards

Default safeguards

Human review Important outputs should be reviewed by a person who understands the topic, context, and intended use.

Source boundaries Factual workflows should use a narrow approved-source list whenever possible.

No hallucinated citations If a claim cannot be supported, it should be flagged or removed.

No overclaiming AI-assisted outputs should not sound more certain than the evidence supports.

No sensitive data in public demos Early demos and examples should rely on public, synthetic, de-identified, or explicitly approved material.

Traceability When appropriate, teams should keep a record of the source material, instructions, reviewer, and approved output.

Data handling

What should not be pasted into a public model

  • PHI
  • PII
  • restricted personnel details
  • legally protected case information
  • confidential organizational records unless explicitly approved for that environment

Better early-pilot inputs

  • public guidance
  • synthetic examples
  • de-identified text
  • approved uploaded documents
  • sample data created for demonstration

Risk-tier

Low-risk workflows

Examples include meeting summaries, internal action lists, and training handout drafts.

Moderate-risk workflows

Examples include dashboard narratives, guidance comparisons, public-facing communication drafts, and community resource guides.

Higher-risk workflows

Examples include workflows affecting legal compliance, clinical or regulatory interpretation, sensitive data handling, or high-consequence public communication.

Review checklist

Before using an AI-assisted output, ask:

  • Is the workflow using approved source material?
  • Was any sensitive data included?
  • Does the output overstate certainty?
  • Are dates, names, numbers, and contact details correct?
  • Is the output within the intended scope?
  • Has the right human reviewed and approved it?

Transparency

Disclosure examples

Internal disclosure option This draft was prepared with AI assistance and reviewed by a human before use.

Public-facing disclosure option Drafted with AI assistance using approved source material and reviewed by a human prior to release.