Why capability is no longer the bottleneck
For many tasks, today's models already deliver impressive results. What holds back broad deployment in high-stakes areas is something else: uncertainty over whether a given answer is correct, and the lack of traceability into why the model arrived at it (see "Where AI hits its limits").
Interpretability research: looking inside
An active research branch ("mechanistic interpretability") tries to understand which internal patterns and concepts a model actually uses to arrive at an answer - comparable to trying to decode a brain neuron by neuron. Early progress exists in identifying individual, clearly delineated concepts inside large models.
Uncertainty calibration: knowing when you're unsure
A second research approach aims for a model to reliably assess how confident it is in an answer - instead of phrasing every answer with the same confidence. Well-calibrated uncertainty would let uncertain answers get automatically routed to a human, instead of treating every answer the same way.
A realistic take
Both research directions are active but far from a complete solution. For concrete, business-critical applications, the practical answer today isn't "wait until the model explains itself", but building your own review processes (see "How to really check if an AI solution works"). Whether and when research fundamentally solves this isn't reliably predictable - progress is incremental, not a sudden leap.
Why this matters for you as a decision-maker
The trust gap - not raw capability - decides which AI deployments are defensible today. The higher the risk of a task, the more important external controls become (see "Human in the loop: when AI needs approval") - regardless of how impressive a model looks in a demo.