Monk Skin Tone scale in evaluations

Why PMs are suddenly valuable again | Tulsee Doshi (Head of Product, Gemini)

Topic 9 of 18

Published Nov 8, 2026 · Clip 26:29–29:51 (3:22)

Tulsi recounts her responsible AI work on skin tone representation. The industry-standard Fitzpatrick scale had only six skin tones, was built for healthcare purposes, and skewed extremely light—failing to represent global diversity. Working with Harvard professor Ellis Monk and his 10-tone scale built from years of user research, Google redid product and model evaluations to reflect the Monk Skin Tone scale and used it to improve efforts like the Pixel camera and image generation. She then broadens the lesson to evals generally: they must be thought about holistically—judge quality matters, the production stack and mitigations affect performance, and serving impacts results (especially in audio, where voice naturalness and tenor matter)—so evaluating end-to-end in a clean setup is critical.