Can AI Build New Knowledge?

Modern AI systems are remarkably good at using knowledge acquired during large-scale training. But it is still unclear whether they can genuinely build new knowledge from experience: identify what they do not know, acquire the right evidence, form new abstractions, revise their understanding of the world, validate what they have learned, and retain it for future use. One way to view current AI progress is as a human-machine continual-improvement loop. A model is trained; humans inspect its failures; researchers design new datasets, benchmarks, architectures, losses, memory mechanisms, or post-training methods; the model is retrained and evaluated; and the cycle repeats. In this sense, AI systems are improving continually, but much of the knowledge-building still happens outside the model, through human diagnosis and design. The question for next-generation AI is whether more of this loop can be internalized. Can future systems identify their own knowledge gaps, decide what evidence they need, learn from video or interaction, store useful memories and exceptions, validate new knowledge, and update themselves without losing previous capabilities?