How much of ML research is about AI safety, what is it about, and who's doing it?

A comprehensive analysis from LessWrong examines how much machine learning research focuses on AI safety. Researchers classified every paper accepted at ICLR, ICML, and NeurIPS from 2019 through 2026—roughly fifty-six thousand papers total—and found that two thousand three hundred twenty-eight, or just four point two percent, address AI safety. What's striking is the growth trajectory: safety's share of papers jumped from zero point three percent in 2019 to eight point three percent in 2026—a roughly twenty-five-fold increase, even as the conferences' overall acceptance rates stayed flat. Interpretability research leads the field, followed by alignment training and adversarial robustness. The analysis reveals an interesting tension: while the dedicated AI safety ecosystem—organizations like MATS, Mila, and Google DeepMind—continues publishing more safety work each year, their share of the total is actually declining, suggesting that mainstream ML research is increasingly addressing safety questions.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton