FishyGrits

Documented history

What the Face-Recognition Test Found

Measured error can expose unequal technical consequences.

A 2019 NIST evaluation found demographic differentials in the majority of face-recognition algorithms it studied. Results varied by algorithm, application, and data, so the finding is not that every system fails identically. It is evidence against treating facial recognition as neutral merely because its decisions are automated.

Error rates become social harms through deployment: a mismatch in a photo organizer differs from a mismatch used in policing, travel, or access control. Rigorous evaluation must therefore connect performance numbers to context, appeal, oversight, and the people made to bear mistakes.

Research notes

Source trail

These sources inform the archive’s account; citation does not imply an institution’s endorsement of FishyGrits.