EleutherAI's loss-of-control model predicts 25% uncooperative AI labor share

David Johnston, writing on LessWrong and linking to an EleutherAI blog post, explores whether we can build an early warning system for loss of control to AI. He constructs a mathematical model to forecast the outcome of the current AI development trajectory. The model's central predictions, though not robust, place the current path at high risk, with an uncooperative labor share settling around 25%. Key parameters include 'self-propagation' (deliberate production of uncooperative AI by uncooperative AI) and 'leakage' (accidental production). The system is sharply sensitive: a 3–4% increase in self-propagation tips into full takeover, while modest decreases improve safety. The author acknowledges that these parameters are poorly characterized due to weak empirical anchors, making the predictions unreliable as forecasts.

Despite these limitations, Johnston argues that the difficulties in modelling such a system are resolvable with more work. He suggests better empirical trends and research into mechanisms could improve parameter estimation. He envisions such a model informing AI governance by clarifying whether risks are being under- or overestimated, thus aiding coordination. The full report details the model's construction, sensitivity analysis, and comparison to AI 2027 scenarios. Overall, the piece is an early attempt to formalize early warning, with cautious optimism about its feasibility.