Insufficient task-specific training data limits AI automation of tasks requiring physical or tacit knowledge.
Assessed first pass
Budget was limited (one turn left after reading the claim), so this was a light first pass with no decomposition. Assessed as supported, credence 0.75, matching the parent's seed. Subclaims on data scaling and on whether simulation or video can substitute for real task data are left for a funded deeper pass.
Assessed Supported
verdict confidence 0.65 · credence 0.75
The claim holds that a shortage of task-specific training data is one of the things holding back AI automation of work that depends on physical manipulation or tacit, unwritten know-how. The evidence favors it. Large language and vision models improved by training on web-scale text and images, but no comparable store exists of robot actions, force and touch signals, or expert craft demonstrations. Robot-learning researchers commonly name this gap as a main bottleneck. Efforts such as the pooled Open X-Embodiment dataset and large teleoperation data-collection programs exist to close it, and their results generally improve as data grows. The disagreement is about how much of the limit is data, not whether data matters. Other constraints also hold back physical automation: hardware cost and reliability, dexterity and sensing, safety requirements, and the economics of deploying robots. Simulation, video pretraining and transfer across tasks may partly replace task-specific data. So data scarcity is well supported as one real limiting factor, but not as the only or main one in every case. Tacit knowledge that never appears in recorded form fits the claim's reasoning, though direct measurements of that effect are scarcer.
Claim entered the graph