From Apprenticeship to APIs: How AI Is Reshaping Knowledge Flows and Transfer
For most of human history, knowledge moved between people the way it always had — through apprenticeship. A junior watched a senior, did the grunt work, absorbed context by osmosis. The knowledge was inseparable from the people who carried it. Geography mattered. Time mattered. Relationships mattered.
Then something shifted. In March 2026, Anthropic's economic research team published a report introducing a metric called Observed Exposure — measuring what AI is actually doing in real workplaces, right now, task by task. The report doesn't predict a future. It documents a transition already underway: knowledge is starting to flow not just from person to person, but from person to API.
Here's why that matters more than any "X% of jobs will be automated" headline.
The Metric: Three Layers of Reality
Observed Exposure is built from three data streams.
Layer one: the theoretical ceiling. Anthropic borrowed a framework from Eloundou et al. (2023) that scores every task in the O*NET database — which covers about 800 occupations in the US — on a simple scale: Can an LLM make this task at least twice as fast? If yes, the task gets a 1. If it needs additional tools built on top of the LLM, it gets a 0.5. If not, it gets a 0. This is the "what's technically possible" upper bound.
Layer two: actual usage. This is where it gets interesting. Anthropic tapped into its own usage data from the Anthropic Economic Index — real conversations with Claude across millions of sessions. What are people actually using AI for, versus what it could do?
Layer three: weighted by purpose. Not all use counts as "exposure." If a programmer asks Claude to generate a complete function and drops it into production code — that's high-weight exposure. The same programmer asking Claude to brainstorm ideas or polish grammar in an email? Much lower weight. Anthropic carefully separates automation from augmentation.
The headline finding: actual exposure is a fraction of theoretical capability. There's a massive gap between what AI can do and what it's being used to do.
That gap — not the ceiling — is what we should be watching.
Who's Actually Exposed?
The report maps exposure across occupations and cross-references it with demographic data.
High-exposure occupations share a profile: highly digital, standardised workflows, quantifiable outputs. Programmers, data entry clerks, translators, customer service reps — jobs that can be broken into clear input-output patterns.
Low-exposure occupations: bartenders, technicians, surgeons. Not because AI can't do parts of these jobs (it can read medical images, analyse recipes), but because these roles require physical presence, real-time physical interaction, and tasks that resist clean modularisation.
What stopped me wasn't the distribution — it's the cross-analysis.
Anthropic found that workers in the most exposed occupations are more likely to be older, female, highly educated, and higher-paid. And here's another finding: the BLS projects these same occupations to grow more slowly through 2034.
Think about what this means. The people most exposed to AI-driven knowledge depreciation are precisely the demographic that looks "safe" — educated, well-paid white-collar professionals. Their jobs aren't disappearing overnight. But the knowledge base those jobs rest on is being encoded, replicated, and devalued faster than ever.
One more finding that stuck with me: since late 2022, there's been no systematic increase in unemployment for highly exposed workers — but there's suggestive evidence that hiring of younger workers has slowed in these occupations. AI isn't laying people off. It might be quietly shrinking the number of entry-level slots available.
What Exposure Tells Us About Knowledge Flow
I read this report through the lens of an old question in economic geography: how does knowledge move between people and across space?
The classic answer is that tacit knowledge — the stuff that can't be written down in a manual — requires physical proximity. Apprenticeship, co-location, face-to-face collaboration. This is why Silicon Valley is one place, not a hundred. Knowledge is sticky, and geography matters because people matter.
Observed Exposure gives us a dynamic view of this process happening in reverse.
When an occupation's exposure score rises, it means tasks that once relied on human memory and manual operation are being converted into programmable logic or API calls. Tacit knowledge is becoming shared code. The cost of moving that knowledge approaches zero.
Economic geographers talk about "global pipelines and local buzz" — the dual mechanism of long-distance knowledge networks and short-range local spillovers. What AI is doing, essentially, is opening a third channel: a pathway for knowledge transfer that doesn't require human contact at all.
It's incredibly efficient at what it does. But it comes with costs that are harder to quantify.
The Price of "De-spatialised" Knowledge
The first problem is structural asymmetry.
AI's attack surface is uneven. Knowledge in high-exposure occupations gets rapidly encoded, copied, automated. Knowledge in low-exposure occupations — because it can't be digitised — moves at a glacial pace. The difference in exposure scores between occupations is essentially a measure of who gets to participate in AI-driven knowledge acceleration and who doesn't.
Here's the paradox: the knowledge workers who benefit most from fast knowledge diffusion (programmers, data analysts) are also the ones facing the fastest knowledge depreciation. The "not-digital-enough" occupations (bartenders, technicians) are relatively safe from replacement — but they're also excluded from the productivity gains.
Economic geography has been raising alarms about the Matthew effect in knowledge flows — that innovation resources concentrate in cosmopolitan tech hubs and exacerbate regional inequality. The Anthropic data hints at a counter-current: the Matthew effect might be reversing, because the occupations that received the most digital knowledge investment are now experiencing the fastest knowledge shelf-life decay.
The second problem is more fundamental.
Traditional knowledge flow has always been a person-to-person parasite. Mentors, colleagues, collaborators, job-hoppers — knowledge moves with people. AI decouples knowledge from people entirely. It becomes infrastructure, not expertise.
Economic geography has a concept called "path creation" — how peripheral regions can absorb and recombine heterogeneous knowledge to build new growth engines. In the traditional model, this requires talent flow, institutional collaboration, policy guidance. In the AI era, a developer in a remote town can access roughly the same capability base as someone in San Francisco. Digital "flow space" does, to some extent, decouple knowledge access from geographic privilege.
But here's the question that keeps coming back to me: when knowledge no longer needs people to carry it, what role do people play in the knowledge ecosystem?
Three Questions I Don't Have Answers For
Anthropic's framework is designed to be rerun — to track how exposure changes over time. That's more valuable than any one-shot prediction. But the framework isn't designed to answer the questions I find most interesting.
First: does rising exposure mean symmetric capability transfer? When a programmer uses AI to generate code, knowledge moves from "human understanding" to "system output." The programmer's own understanding might be bypassed in the process. Knowledge reaches the task, but it doesn't reach the person. This isn't the same thing as traditional knowledge flow, where the receiver actually learns something.
Second: does AI-generated "knowledge" have the same reproductive capacity? An apprentice who masters their mentor's craft can innovate, mutate, adapt. A developer who relies on AI-generated code — how much independent problem-solving capacity remains when the AI isn't available? This isn't an anti-AI argument. It's a question about what happens to the human node when the knowledge flow path shifts from "person → person" to "system → task."
Third: if Observed Exposure is itself a measure of knowledge flow — specifically, the speed at which tacit knowledge is converted into system-bound code — then are the low-exposure occupations, the ones that resist digitisation, being left out of this flow in a different way? They're not being replaced. But they're also not being empowered.
But AI Might Also Be a Wall
All the questions above assume AI is enabling some kind of knowledge flow — just maybe in the wrong way, at the wrong speed, to the wrong people.
The more I read about this, the more I wonder about a darker possibility: AI isn't just failing to enable knowledge flow. It might be actively walling it off.
Let me walk through why.
Observation one: the entry-level jobs that carry knowledge between generations are disappearing.
Traditional talent development was apprenticeship in disguise. New hires did the grunt work — data cleaning, first drafts, standard customer queries. The tasks themselves weren't the point. What mattered was the context you absorbed while doing them: how the team actually operates, which unwritten rules matter, who to ask when things get ambiguous. Tacit knowledge moved from senior to junior through shared work.
AI is eating these entry-level tasks. A model can clean data faster than any human, generate first drafts in seconds, handle standard queries at scale. The company saves money. The senior person gets more done.
But here's the catch: people don't absorb knowledge by installing an API. Newcomers need those low-level tasks to build a mental model of the business. When AI takes over, the apprenticeship period doesn't get shortened — it gets eliminated. Junior people lose the "peripheral participation" that let them gradually move toward core contribution.
Knowledge management researchers call this the gate-lock effect. AI locks the gate to the workplace.
And what about the seniors? Without juniors to mentor, their tacit knowledge doesn't flow anywhere. The things that can't be written in a manual — when to break the rules, how to make judgement calls in grey areas — simply stop being transmitted. Senior workers become isolated islands of expertise that vanish when they leave.
Observation two: knowledge is being "assetised" — and that's killing person-to-person transfer.
When an occupation has high observed exposure, organisations tend to do the obvious thing: extract the core operational knowledge from their best workers, distil it into an AI model or automation module, and store it on the company server. The tacit knowledge gets disembodied and immortalised. From an efficiency standpoint, this makes perfect sense.
From a knowledge flow standpoint, something fundamental shifted: knowledge stopped moving between people. It now moves between people and systems.
Person-to-person flow has a property that person-to-system flow doesn't: transmission is also transformation. When knowledge passes from mentor to apprentice, it mutates. The apprentice doesn't just copy — they interpret, adapt, recombine, sometimes improve. Knowledge is a living thing.
When knowledge becomes a module on a server, it's an asset. It's perfectly preserved and perfectly frozen. You can call it via API. You can't argue with it, improve it, or make it your own.
There's an even subtler layer. I've come across research suggesting that workers in high-exposure roles are starting to engage in defensive knowledge hoarding. They know their expertise is easily captured and replicated by AI. So they stop sharing the unique tricks that make them valuable. They execute critical tasks in unscripted, untraceable ways. It's a natural human response to being "distilled" — but it builds new walls inside the organisation.
Observation three: a paradox is forming.
Here's the strange situation we're in:
Digital knowledge accessibility has never been higher. You can ask an AI to explain almost anything, have a conversation about it, get a custom tutorial. If you want to learn something, the barriers have never been lower.
But if you want to enter a profession — become a competent practitioner in a field — the barriers are higher than ever. Companies prefer to hire experts who can "hit the ground running and direct AI effectively." Training juniors is a luxury fewer firms can afford.
The result: knowledge has never been more "free" at the digital layer, and never more "stagnant" at the social layer.
Economic geography worries about regional inequality — resources concentrating in metropolitan hubs while peripheral areas fall behind. This framework gives us a sharper version of that concern. It's no longer just about geography. It's about who can command AI effectively. Resources — knowledge, opportunity, growth — concentrate around that capability. Everyone else — including the young people who need apprenticeship time to develop — gets locked out.
If this trend holds, organisations will face a mid-career talent断层 (断层 is the Chinese word for "broken layer" — a gap where a generation of expertise should be). The digital shortcut comes at the cost of long-term knowledge reproduction.
This is a harder problem than "how many jobs will AI replace?" Replacement is at least measurable. Generational knowledge fractures — you don't see them until they've already happened.
So
Anthropic's report isn't the kind of article that makes you go "wow." No apocalyptic numbers, no utopian promises. It's a scaffold — a reusable framework for measuring AI's real penetration into the labour market.
But its sharp edge is this: it moves the conversation from "can AI do this?" to "is AI actually doing this?" And what it's doing, quietly and unevenly, is rerouting the channels through which knowledge flows.
For centuries, the pipeline was apprenticeship. Person to person, generation to generation — knowledge transmitted through shared work, imperfect copying, and slow accumulation of tacit understanding. The new pipeline is different. Person to API, system to task. Faster, cheaper, massively scalable — but it doesn't reproduce the carrier. It doesn't build the next generation of practitioners.
The quietest finding in the report might be the most consequential: since late 2022, hiring of younger workers has slowed in high-exposure occupations. No mass layoffs. No dramatic headlines. Just a slow contraction of the entry-level paths that used to turn newcomers into experts.
That's the real transition the title points to. From apprenticeship to APIs. And we're only beginning to understand what gets lost along the way.
What do you think?
Based on Anthropic's economic research report "Labor market impacts of AI: A new measure and early evidence" (March 5, 2026). Primary data sources: Anthropic Economic Index, ONET database, Eloundou et al. (2023) task-level exposure estimates. Personal judgements are clearly marked — they are not predictions or investment advice.*