FactOver the past 18 months, leading large models have reached — and in some cases exceeded — the average accuracy of entry-level professionals in drafting legal documents, pre-screening medical imaging, and reviewing code. This isn't a forecast; it has been confirmed repeatedly in internal evaluations at multiple institutions.
But accuracy isn't the question this story is trying to answer. What we wanted to know is: once "judgment" itself can be outsourced, what's left that genuinely can't be — and shouldn't be — replaced?
I. "Expertise," redefined
Source's view"It took me twelve years to learn to read the hesitation in a single CT scan," says Dr. Chen Jingyi, a radiologist at the National University of Singapore. "AI reads pixel patterns. I read what a patient didn't say out loud across their last three follow-ups." After her department introduced an AI pre-screening system last year, misdiagnosis rates dropped — but her clinic time got sliced into more and more five-minute segments.
"Efficiency went up. Nobody ever asked us what the time we saved should actually be spent on."
InferenceIf this pattern repeats across more professions, the core of expertise may shift from "the ability to complete a task" to "the judgment to know the task's limits" — knowing when to trust the system, and when not to.
II. Three responses from workers
At a data-annotation company in Nairobi, Kenya, we found three starkly different attitudes. Veteran annotator Omari chose to become an "AI auditor," specializing in catching the model's systematic errors. Younger colleague Wanjiku is teaching herself prompt engineering, converting her annotation experience into a new skill. Kiprono, who has since left the job, believes the work itself "was, from the start, training human judgment to become disposable raw material."
FactAccording to the International Labour Organization's 2026 report, over 60% of the global data-annotation workforce is concentrated in Southeast Asia, South Asia, and Africa — while the ultimate beneficiaries of the resulting model capability are concentrated in a handful of countries.
III. What's left may be the capacity to "bear the consequences"
Deep into our interviews, nearly every source, independently, pointed to the same thing: not creativity, not empathy, but "the willingness to bear the consequences of a judgment." A system can generate a recommendation, but no mechanism today lets a model actually be "held responsible."
Source's view"Maybe human uniqueness was never about capability at all," says ethicist Marcus Chen. "Maybe it's that we're willing to stop, apologize, and correct course for each other's mistakes. AI can't do that — and it shouldn't be the one to do it."
This isn't a question one story can fully answer. YouDuJi will keep following these 12 sources' choices over the coming year, as part of our "Futures Not One" hardcore series.