Young people use AI every day, yet almost no one can reliably answer three basic questions: Is it safe? Does it deliver its intended benefit? What does repeated use do to young people’s development?

Too much evaluation is funded as one-off projects, with little incentive or infrastructure to make the evidence add up over time.

This paper is a premortem. Imagining the field in 2036, it works backward from failure, drawing on interviews with the teams behind benchmarks like HealthBench and LegalBench, and on philanthropy’s track record in fields from maternal health to social-emotional learning.

It names six ways an evidence system for AI and young people could fail, and shows how philanthropy can prevent them by funding what no one else owns: shared definitions and records, independent evidence on consequential uses, and pathways that connect evidence to real decisions.

Download the full report to explore where funders should start.

Authors: Allison Salisbury and Andrew Buher