2026-04-22
Designing hiring signals that stay fair under AI
When embeddings and LLMs enter the loop, the product decision is what *not* to optimize for.
Start from decisions, not models
Before tuning prompts or vector indexes, write down which decisions the model supports and which remain human-only. Examples: ranking for review order (supported) versus auto-rejection (avoid).
Document the boundary
Store a short rationale for why a score exists — one to two sentences — so hiring managers can disagree productively with the model instead of treating it as a black box.
Audit with real panels
Periodically sample applications across demographics, seniority bands, and job families. If scores cluster unfairly, fix the process — not just the temperature parameter.