Last weekend I performed in Mattavilãsaprahasana, a seventh-century Sanskrit play at Christ Church College in Oxford. While preparing, I used large language models (LLMs) to check the pronunciation of some words and phrases. They got parts of it right but missed the finer details. I reckon it has to do with the way these LLMs are trained. They are first built by learning from enormous amounts of written text, and spoken Sanskrit simply isn’t widely used or well represented in those sources. After that comes post-training, where human experts fine-tune the model by correcting its mistakes and showing it better examples. For Sanskrit pronunciation, the amount of expert guidance is small. In fields like business, finance, and technology, though, the picture is different. These models receive far more expert feedback and real-world examples, which gives them noticeably more reliable instincts.
This human-AI synergy has broad economic implications as it puts humans at the forefront of making AI smarter. In fact, the human contribution that now matters most is no longer basic data labelling or annotation, but guiding models through the way real life decisions are made. Even after textual training, a model needs human input to function well in real contexts. It must see how experts weigh evidence, resolve ambiguity, and apply standards within their fields.
Because of this, the talent pool for training AI has expanded far beyond engineers. At micro1, a company I have been partnering with, I see philosophers, linguists, historians, teachers, clinicians and legal scholars now take part in shaping how AI systems perform. The strength of these systems depends on the clarity and depth of the explanations they receive from a range of subject matter experts.
This has led to the creation of a new expert economy. Reinforcement learning environments function as workplaces where people review model outputs, assess alternative answers and demonstrate how a knowledgeable person approaches a problem. Many roles offer flexible hours and remote participation. They provide a path into the technology sector for people whose expertise was often overlooked in previous waves of innovation.
The question of long-term stability often comes up. If models learn from experts, could the experts eventually lose their place? Evidence from current practice points in the opposite direction. Models shift as the world shifts. New facts appear. Norms evolve. Legal and cultural contexts change. A system that performs well at one moment can drift away from acceptable behaviour without steady human adjustment. The more capable these systems become, the more important that correction becomes, because the issues they touch carry greater consequences.
My experience with Sanskrit pronunciation was a small illustration of this dynamic. A model can sound confident, yet still carry errors that only a trained speaker would notice. In many fields the same pattern holds. AI produces fluent output, but only expert review can confirm whether it reflects sound reasoning and current practice. As these systems enter medicine, law, education and finance, human oversight becomes a structural requirement, not a temporary phase.
The future of work will involve fewer routine tasks and more roles that rely on thoughtful analysis. The shift underway is not simply about efficiency. It is about the rising value of human judgment that can be articulated, taught and scaled. That judgment forms the foundation of modern AI and the basis of an emerging economy that rewards expertise across disciplines.
Thus, it is time to pivot from fear-laden conversations around AI taking over jobs and make way for the realignment shaping the new normal of work.