Framework
The Bias Audit Triad
Three practices that prevent AI models from encoding discrimination: clean and representative data, demographic testing before launch, and continuous monitoring for drift.
How it works
- 1Step 1 — Clean data: make sure the training set is not obviously wrong. A model trained only on one demographic will encode that demographic's patterns as universal truth.
- 2Step 2 — Model cards: for every model you ship, build a table showing false positive and false negative rates across demographic groups and intersectional groups. Not just 'Black' and 'women' — Black women is a separate intersectional group.
- 3Step 3 — Monitoring: models drift as populations shift. A model that is fair on launch can become biased as the user mix changes. Set up ongoing drift detection.
- 4In hiring, discrimination is legally regulated. In healthcare, it is just unethical — and often invisible. That absence of regulation means the responsibility falls entirely on the builder.
- 5The goal is not equal treatment across groups. It is precise treatment. Men and women need different things at different points in their healthcare journey. Precision requires knowing which group you are serving.
Hear it on the podcast
The Behavioral Health Crisis Is a Data ProblemListen to the episodeMore from this episode
The Continuity GapThe most valuable data in healthcare — and in most SaaS products — exists between appointments, not during them. Building for the gap is the actual product opportunity.The Paying Problem FilterCool technology that no one will pay for is not a product. Before you fall in love with your insight, find the customer whose budget line it solves.Build for Signal, Not AssumptionWhen training models for human behavior, go in with as few assumptions as possible. Let the data surface the features. Your intuitions about what matters are often wrong.
Get the next framework first
One email a week with the frameworks, tools, and founder stories from the podcast. Free.
More frameworks for AI & Automation
The Centralized Data LayerOne AI data analyst serving the whole company beats fifteen teams building their own agents from different sources.Build for Agents, Not Just HumansAgents have fundamentally different interface preferences than humans. Designing for both requires treating them as separate audiences.The Three Stages of Data IntelligenceMost companies are stuck at Stage 1. The real value - and the real competitive advantage - lives in Stages 2 and 3.Agent Access Control as a Safety FoundationYou cannot let agents go wild if they have read-only access to two tables. Granular permissions are the foundation of safe agentic deployment.Context Over ModelThe AI model you use matters far less than the quality and completeness of the context you feed it. Managing context is the real competitive advantage.Personal AI LibrarianAn AI that knows everything in your library — and only your library. Trust through constraint.