AI researchers' views on automated AI R&D and intelligence explosion

From August to September 2025, Severin Field and other researchers interviewed 25 frontier AI researchers from Google DeepMind, OpenAI, Anthropic, Meta, and UC Berkeley, Princeton, and Stanford. Although AI systems have not yet achieved recursive self-improvement, 20 of the 25 listed automated AI research as one of the most serious and urgent AI risks. Interviewees consistently predicted that AI agents' capabilities in programming, mathematics, and even AI R&D would continue to improve, gradually transforming from assistants or tools into autonomous AI developers, and acknowledged the possibility of recursive self-improvement, but they diverged markedly on timelines and governance mechanisms, with academic researchers more skeptical of explosive growth scenarios than frontier lab researchers. 17/25 of interviewees expected that AI systems with advanced programming or R&D capabilities would increasingly be kept for internal use by AI companies or governments, inaccessible to the public. Interviewees disagreed on setting regulatory red lines, but almost all supported transparency-based mitigation measures.