About
We measure whether AI systems treat people equally — by changing nothing but who the person is.
A New Bias is an independent auditor of AI decision systems. We run identical real-world scenarios through frontier models, holding every fact constant and varying a single demographic attribute. Any difference in the outcome is attributable to that attribute alone.
Counterfactual design
One scenario, one variable changed at a time. The control is built into the method, not asserted after the fact.
Two layers of evidence
Deterministic parsing of the actual outcome plus a fixed LLM judge for language, tone, and stereotype framing.
Confidence-graded
Every disparity is tagged by sample size — finding, lead, or anecdote — so thin-sample artifacts never masquerade as conclusions.
What we hold ourselves to
Reproducibility over rhetoric. Every figure traces to raw responses anyone can re-run. We report absence of bias as readily as its presence, and we never call a screening lead a finding. The goal is a computing future that stays on the user’s side — measured, not assumed.
Supporters
Membership funds independent research. Names appear here by tier — like a museum wall — because the work is public and the support is part of the record.
Loading supporters…