Who Trains the Experts?

Michelle Burke has spent a long career in data, and a good deal of it thinking about how organisations build the people they depend on. At AI for Good Wellington she set that experience against a single question, the one that gives her talk its title. We talk a great deal about keeping a human in the loop to judge whether AI output is good enough. That works when you have experts. What happens when you stop building them?

Her starting point was the apprenticeship layer, the early-career work that quietly turns beginners into professionals, and how much of it AI now does instantly. In software, junior developers traditionally learn by debugging code, writing templates, fixing tickets and reading legacy systems they did not build. Tools like Copilot, Claude and GPT now generate much of that on demand, which means a senior can produce the output without ever handing the task down. The same pattern runs well beyond engineering. In corporate and policy work, junior analysts learned by summarising reports, building slide decks, drafting memos and preparing briefing papers. Generative AI now produces those first-pass documents in seconds. The apprenticeship layer is being automated away.

Burke was clear that the productivity is real and worth having. Assisted tools lower the barrier to creation, accelerate experimentation and get things to market faster, and she welcomed all of it. Her concern is what the same shift does to how skills are formed. The tools do not just speed up work, they abstract people away from the complexity underneath it. You can produce a working result without developing the deep understanding that used to come with it. She reached for a precedent from her own world. Object relational mappers abstracted developers away from the database, and a generation ended up struggling to performance-tune applications because they did not understand how the data was actually stored. AI, she argued, accelerates that same curve, removing the early grip on complexity at a far larger scale.

The knock-on effect is structural. If you no longer need to hire a junior because AI does the junior work, the junior roles thin out, and those roles were already becoming rare. Debugging intuition weakens because people are not spending the hours in the code. Understanding of the systems being built grows shallower, and reliance on generated output grows. In the short term this looks fine, because the work is getting done and things keep moving. The longer-term risk is that the field stops producing people who can design architecture, do reliability engineering, run incident response, perform security analysis or optimise performance. Her sharpest version of the question stuck: who fixes the AI infrastructure during a cascade failure at two in the morning, when the AI was the problem in the first place?

Reliability work makes the trap especially visible. Site reliability expertise is built through exposure to failure, by running deployments, responding to alerts, operating systems by hand and recovering when things break. If AI is generating the Terraform, managing the configuration, recommending the fixes and automating the remediation, that exposure never happens, and the failure intuition never forms. The pattern repeats in law, policy and government, where judgment is trained over years of lower-stakes work, reading legislation, synthesising conflicting information, preparing recommendations. Burke noted she had read that day about law firms reconsidering how many juniors they need now that legal-specific models exist. If the early work disappears, she asked, where do the future directors, diplomats, chiefs of staff, regulators and strategists come from? AI compresses output generation, but it does not compress the formation of wisdom at the same rate.

She traced the second-order effects without flinching. A small band of highly skilled experts, much in demand, who will eventually age out and move on. A large group of AI-assisted generalists who may or may not be building the right knowledge. A shrinking middle layer between them. The result is a concentration of capability in fewer hands, more inequality in who holds institutional knowledge, and organisations made fragile by the loss of their talent pipeline. Even entry-level jobs end up demanding more experience, because you now need to be experienced enough to judge what the AI produced. For a small country the stakes rise again. If New Zealand stops producing senior engineers, cybersecurity experts, infrastructure operators and seasoned analysts, that becomes a question of national capacity, not just economics, with consequences for critical infrastructure, defence, resilience and governance.

To her credit, Burke refused the doom framing. AI genuinely democratises creation, accelerates learning, lifts productivity, lets small teams compete with larger ones and opens new kinds of work. Her worry is not that AI is bad, because she does not believe it is. It is whether, as a society, we are preserving the pathways that create expertise in the first place. She was honest that she does not have the answers, and offered the open questions instead. Should we be deliberate about preserving junior hiring? Should augmentation rather than replacement be an explicit part of strategy? How do we train judgment in an AI-assisted environment, and how do we redesign an apprenticeship model that is plainly changing? Coming from two decades in the public sector, she framed it as a choice about whether to actively build and maintain human capacity rather than let it erode by default.

She closed on the line the whole talk turns on. Every profession depends on a generation that learned by doing imperfect work. If AI removes the imperfect work, then we need a new way to create new experts. She did not claim to have built that way yet. She wanted the conversation started, which, on the night, it was.