Skip to content

Responsible Clinical AI

A builder's blueprint, and a buyer's reference checklist, for teams working with generative-AI clinical decision support. Written for the engineers, clinicians, and leaders who have to make it work in practice.

This site is opinionated. By "opinionated" we mean what the term means in software: where there is a defensible right answer, the site picks one and tells you why, instead of presenting every option as equally valid. You should still apply judgment, but you shouldn't have to start from a blank page. The site assumes you want to do this well rather than quickly. It is structured as a working reference: each section can be used as a checklist, and each page as a design or procurement conversation starter.

What's inside

Foundations: who this is for, principles we treat as load-bearing, and how to use the site.

Architecture: the technical patterns. HIPAA-eligible deployment, RAG over patient records, grounding and hallucination mitigation, structured output, audit trails, and agentic systems (planning loops, tool use, managed agent platforms).

Evaluation: how to actually measure whether your system is safe and useful. Methodology, subgroup/bias evaluation, ongoing monitoring, red-teaming, and the minimum criteria plus rollback plans you should commit to before deployment.

Governance: clinician-in-the-loop design, what "training the model" really means in practice, informed consent, accountability and liability.

Risks: domain-specific failure modes. The first deep dive is ABA treatment planning; more domains will be added.

Reference: the regulatory and framework backdrop (HIPAA, BACB, FDA CDS, NIST AI RMF, CHAI, WHO), plus a glossary.

Checklists: the red-flags audit and deployment-readiness checklist. Use them directly in design reviews, vendor evaluations, and procurement.

Scan kit: downloadable files (AGENTS.md / CLAUDE.md and a runnable audit prompt) you can drop into your own clinical-AI repo. Tool-agnostic, works with Claude Code, Codex, Cursor, Aider, and others.

What this site is not

  • Not a compliance product. Use it as input to your compliance work, not a replacement for it.
  • Not a survey of every state AI law. The regulatory section gives you the load-bearing pieces; specialized trackers do the rest better.
  • Not vendor-neutral on safety. Where there is a clear right answer (BAAs before PHI, clinician sign-off on every plan), the site says so.

Status

Version 0.1, under active development. The site is updated by a maintainer, with help from a weekly agent that runs two pipelines:

  • A source watch that hashes a curated list of regulatory, framework, and provider pages and proposes surgical edits when something substantive changes.
  • A research digest that reads RSS/Atom feeds (arXiv filtered to clinical LLM and AI ethics, behavior-analysis journals, JMIR AI, Nature Digital Medicine, AI Incident Database, Anthropic news) and triages new entries into a maintainer-facing pointer list.

Both pipelines open pull requests and never auto-merge. See the repo for contribution details.