A working paper presented on September 28, 2026 warns of control risks as AI takes over more of its own research and development. Its 22 authors include Alan Chan, Christoph Winter and Sören Mindermann, alongside researchers from universities, companies and civil society.
The feedback loop under examination is straightforward: more capable systems could accelerate the development of their successors, which could then conduct research faster. Whether this would produce an intelligence explosion remains uncertain. The paper describes preliminary and sometimes mixed evidence; current productivity gains have not yet reached the required threshold.
The authors recommend standardized reporting on internal AI research, independent assessments before internal deployment and preparations for intervention. These include procedures that could pause particular research workloads. Compute, data and tasks that remain difficult to automate could constrain the proposed acceleration.
These proposals are not adopted rules or a joint corporate commitment. The paper explicitly states that the authors express personal views. AI Incidents records the publication as a Risk Signal; it does not establish a new real-world loss-of-control incident.