"Real-World Gaps in AI Governance Research Data"

"Drawing on 1,178 AI risk and reliability papers from 9,439 generative AI papers (January 2020through March 2025), we compare research outputs from the most influential research institutions: frontierAI companies (Anthropic, Google DeepMind, Meta, Microsoft, and OpenAI) and leading AI universities(CMU, MIT, NYU, Stanford, UC Berkeley, and University of Washington). We find that Frontier CorporateAI research increasingly concentrates on pre-deployment areas \u2014 model alignment and testing & evaluation\u2014 while attention to deployment-stage issues, such as model bias, has waned as commercial imperativesand existential risk concerns have taken precedence. We identify significant research gaps in high-riskdeployment domains, including healthcare applications, commercial and financial contexts, misinformation,persuasive and addictive features, hallucinations, and copyright usage in training and inference. Withoutconcerted efforts to enhance external observability into AI\u2019s deployment, the growing concentration of AIresearch with frontier corporations could deepen knowledge deficits in these critical deployment areas.We recommend measures to expand external researcher access to deployment data and improve systematicobservability of AI systems\u2019 in-market behaviors."