A new analysis of machine learning conference papers reveals a significant increase in AI safety research. Out of over 55,000 papers analyzed from ICLR, ICML, and NeurIPS between 2019 and 2026, 2,328 (4.2%) were classified as AI safety-related. The proportion of AI safety papers grew from 0.3% in 2019 to 8.3% in 2026, indicating a substantial expansion in this field. Interpretability, alignment training, and adversarial robustness are the most common subdomains within AI safety research, though topics like AI welfare remain largely absent. AI
IMPACT Indicates a growing focus on AI safety within the research community, potentially influencing future AI development priorities.
RANK_REASON Analysis of academic papers from major ML conferences. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
