Analysis Reveals Shifting AI Safety Research Priorities at Major Labs—OpenAI Improving, Anthropic Declining
Key Takeaways
- ▸OpenAI's safety research output is higher than reputation suggests and has been improving, despite some work being derivative of Anthropic efforts
- ▸Anthropic shows a significant downward trend in safety-related publications, raising questions about whether its 'safety company' reputation reflects current priorities or 2023-era positioning
- ▸DeepMind's research portfolio remains heavily weighted toward capabilities and applications, with safety researchers reportedly facing difficulties securing resources and publication permissions
Summary
A new analysis examining the research output of three major AI companies—OpenAI, Anthropic, and Google DeepMind—challenges conventional wisdom about their commitment to AI safety research. By analyzing 59 OpenAI blog posts, 86 Anthropic publications, and 233 DeepMind papers through 2025, researchers used machine learning to classify outputs as safety-related or capability-focused, revealing significant trends in how these organizations allocate research attention.
The findings suggest OpenAI's safety research share is higher than commonly credited and has been improving over time, contradicting public perception that it lags peers. DeepMind shows modest safety research growth but remains heavily focused on applications and experimental capabilities. Most strikingly, Anthropic—widely regarded as the industry's "safety-first" company—displays a robust downward trend in safety-related publications, with the analysis suggesting its safety reputation may be largely a artifact of 2023-era output rather than current priorities.
The research acknowledges methodological limitations, including differences in how companies publish (blog posts vs. academic papers vs. indexed publications) and the challenge of comparing organizations with different publication cultures. Still, the data raises important questions about whether AI companies' stated safety commitments match their actual research investment levels.
- Methodological limitations exist due to different publication venues and standards across companies, making direct comparison imperfect
Editorial Opinion
This analysis provides valuable quantitative data on a crucial but often opaque question: how much are AI labs actually investing in safety versus capabilities? While the methodology has acknowledged limitations, the finding that Anthropic's safety reputation may rest on historical momentum rather than current output is particularly important for policymakers and the public to understand. The upward trend at OpenAI and stagnation at DeepMind suggest that safety research investment is neither static nor uniformly prioritized—a reality that should inform both regulatory scrutiny and investment decisions.
