Understanding AI Trajectories: Mapping the Limitations of Current AI Systems

As of September 2025, AI systems lack the general cognitive capabilities that could enable large-scale automation of cognitive labour. The limitations of today’s systems include performance limitations on certain types of tasks, insufficient reliability, insufficient adaptability, and low capacity for original insights. Much of the progress on these limitations to date has been driven by scaling up the computational resources used to train and run AI systems. This trend is expected to continue, but there is uncertainty over how much future progress will require fundamental research innovation versus being driven primarily by further compute scaling. Through literature review, interviews with internal and external experts, and an internal workshop with the UK AISI Research Unit, we have identified eight indicators of progress against the limitations of current AI systems, and gathered perspectives on how these limitations are most likely to be overcome. Each indicator is presented alongside an initial appraisal of the evidence for progress and suggestions for additional pieces of evidence that would bear on it