A new arXiv paper, Assessing the Empirical Evidence for Loss of Control from Agentic General-Purpose AI, takes on a central question in AI safety: do current or near-term agentic systems actually pose the risk of irreversible human loss of control? The abstract notes that theoretical literature has warned that advanced generalist agents might pursue instrumental goals misaligned with human preferences, but the paper's focus is on what empirical evidence exists to support or contradict those warnings.
The paper is positioned against the backdrop of the EU's Code of Practice, suggesting the authors intend to inform policy discussions with a reality check. According to the abstract, the work appears to be a systematic assessment rather than a new theoretical argument, though the full methodology and conclusions are not available from the abstract alone.
Because the source is only an abstract, the specific findings and the strength of the evidence reviewed remain unclear. What is clear is that the paper aims to bridge the gap between theoretical risk models and empirical observation, a crucial step for governance frameworks that rely on demonstrated harms rather than hypothetical scenarios.