Why does AI learn to handle common store tasks fast, but still struggles with rare, tricky medical cases?

This explores why AI automates frequent, repetitive tasks first and leaves rare, specialist ones like uncommon medical work for later. The short answer from the corpus is that the reason is mostly economic, not about how hard the task is. Adapting an AI system to a task, whether by fine-tuning, collecting data, or building and testing a pipeline, is a fixed cost. You pay it once, however often the task runs afterward. A retail task that runs millions of times spreads that cost thin. A rare medical procedure that runs a few hundred times a year never earns it back. Does AI's automation advantage depend on how often a task runs? finds that task scale explains about 63% of the variation in whether automating a task is worth it. Two tasks of similar difficulty can end up on opposite sides of the line purely because of how often they happen. So a rare medical task may resist automation simply because nobody can justify paying to set up a system for it, not because it's too hard.

A second, more technical reason points the same way. Rare medical work depends heavily on stored knowledge, and techniques that make models better at step-by-step reasoning don't automatically help with that. Why does reasoning training help math but hurt medical tasks? reports that knowledge recall happens in a model's lower layers and reasoning in its higher layers. As a result, reasoning training that improves math can actually make a model worse at medicine. The general-purpose progress that makes high-volume tasks cheap to automate doesn't carry over cleanly to knowledge-heavy specialties, so those specialties need their own costly adaptation. That brings back the fixed-cost problem.

The cost of getting things wrong is a third factor. Rare cases are where mistakes are most likely and where the person who has to step in is least prepared. A meta-analysis of 18 experiments When does automation harm performance after failures? found that automation improves routine performance but hurts recovery when it fails. The damage is worst when the system moves from analyzing information to choosing actions. Someone who has been watching an automated decision-maker loses the alertness and practice needed to take over. In medicine, rare cases are often exactly those failure moments. Related work on verification What becomes scarce when AI execution grows cheap? argues that as AI execution gets cheaper, the real bottleneck becomes the human time needed to check its work. Checking work on rare, high-stakes tasks is expensive, and the people qualified to do it are scarce.

What you might not expect is that "automated" versus "not automated" is the wrong way to frame it. Once the high-volume parts of a job are automated, the human work that's left often becomes more valuable, not less. Does automating some tasks make other human tasks more valuable? points to US bank teller jobs, which grew alongside ATMs. Does automating some tasks actually increase the value of remaining work? argues that when tasks depend on each other, automating some of them focuses people's time on the bottleneck task, which can push up its value. On this view, the rare medical judgment becomes the part of the job that matters most because the routine work around it has been automated. Field studies also show that automation tends to change manual work rather than remove it. In Does automation eliminate manual work or repurpose it?, workers who used to do classification by hand ended up checking and correcting the algorithm. Are automation and augmentation really a simple trade-off? argues that automating some tasks and assisting people on others feed into each other over time.

A caveat: the corpus has no studies that directly compare retail and medicine, so the explanation above is pieced together from economic and human-factors work rather than tested head-to-head. One practical consequence follows from the fixed-cost argument. If that is what keeps rare tasks unautomated, then anything that lowers setup costs could move rare tasks across the line without models getting any smarter. Examples would be general-purpose models that need little adaptation, or ways to share one adaptation across many rare tasks.