Introduction: Who Trains the Experts?

I have been working in data for a really long time and thinking about a lot of really cool things and today I'm going to talk to you about who trains the experts.

I think my slide got squished but so really thinking about how we're actually building expertise in looking towards the future.

The Lens: AI Adoption and the Human-in-the-Loop

So who am I? Leader, strategist, opinion holder, natural blue hair.

This is something that I've been thinking about a lot in terms of kind of AI adoption and the kinds of things that are really important when we start thinking about the human in the loop and using experts to actually assess whether or not the outputs are good

The Apprenticeship Layer We’re Automating Away

and that's great when you have experts but what about when you're not building them and so when we think about how we build sort of that expertise in now or

Software Development: How Juniors Traditionally Learn

historically if we look at it in the software development space so junior developers they debug code they write templates they fix tickets they read read legacy code, and kind of learn from that as they build.

But now we've got tools, you know, Copilot, Claude, GPT, generate much of that instantly. And so seniors can produce output without having to pass things on to the juniors.

Corporate and Policy Work: The Same Pattern

I want to think about it, too, in sort of that corporate and policy space, so not just in IT, but, you know, so junior analysts would summarize reports ports and build slide decks and read things, build memos, prepare briefing papers.

And now Gen AI can produce those first pass documents without any extra input in seconds. And so we're automating that apprenticeship layer out.

Abstraction, “Vibe Coding,” and Shallowing Skill Formation

And so in the vibe coding space, obviously using assisted tools to be able to generate stuff. And it's incredibly powerful.

And it's lowering the barriers to creation which is amazing and accelerates experimentation which is awesome and so we're seeing things come to market quicker and I think that's really great but it changes how skills are actually created and formed so it's

not just the tools that improve productivity it creates abstraction from the complexity of the system it means that we don't have the understanding and deep understanding necessary as we move forward and I kind of think about it

A Familiar Precedent: ORMs and Losing Database Intuition

I've seen this before as somebody who comes from a data background and loves databases deeply. Object relationship models abstracted developers from the database, which meant that we'd have issues with them performance tuning applications and things like that because they don't understand how the data is stored.

For me, this is just accelerating some of that curve in removing some of the ability to understand the complexity and so you know it kind of moves us into into a space where we don't have the skills right from the beginning but at a scale that's just increasing and growing and it matters because you know we're also removing that first stage

and so we don't have we don't have the junior roles that we used to have because you don't need to hire a junior anymore because I've got AI to do it and so there are a few general roles which were already rare and so now you know we're losing we're losing the ability to bring people in we're reducing the debugging

intuition because they're not spending enough time coding we've got weaker understanding of the systems that are being built and used and we've got really shallow expertise and an over reliance on generated code which in the

short term, no problem, because we're producing great stuff and things are moving forward and everything is great.

Long-Term Technical Risk: When Expertise Is Needed Most

But in the longer term, we're not going to have people who are capable of designing architecture, do reliability engineering, do incident response, do security analysis, do performance optimization.

And who's going to fix the AI infrastructure during a cascade failure at 2 a .m. when the AI was the problem in the first place and of course you know

Reliability and SRE: Expertise Built Through Failure

in the infrastructure and SRE space it's the same kind of thing you know the skills are built by running deployments by responding to alerts by operating systems manually by handling failure and if the AI systems are generating the terraform they're managing the configs they're recommending fixes they're automating the remediation we're not developing that failure intuition and so So reliability expertise, which is built traditionally through exposure to failure, how do we generate that without the exposure in the first place?

Beyond Tech: Law, Policy, and Institutional Judgment

And it happens in the corporate and policy space as well. It's not just a technology problem. So again, when we're not reading, we're not developing papers, we're not summarizing legislation, we're not preparing recommendations, synthesizing conflicting information, building that muscle muscle, then, you know, where does that expertise go? And if the AI is doing this, and again,

you know, we're starting to see a reduction in junior and grad programs, which has issues for, you know, people coming out of uni, and, you know, not having the opportunities to move into those roles. I read an article today that there are law offices that are looking at the fact that they don't need as many juniors anymore with legal specific AIs, so in large language models. So, you know what does that mean for junior law folks to come in and so the challenge with that though

is that then where do the future directors and diplomats and chiefs of staff and regulators and strategists come from if we haven't actually built those skills to begin with and so that you know that judgment that's usually trained through years of learning and doing in sort of low stakes work doesn't happen. So we've got AI compressing output generation in the short term, but we're not actually compressing wisdom formation at the same time. And so what does that mean for us?

Second-Order Effects: Stratification, Fragility, and Pipeline Loss

You know, I think we've got expertise stratification. So, you know, creating a small band of highly skilled, but highly sought after experts who will get older and move on.

we've got a lot of AI -assisted generalists who may or may not be building the right knowledge we're really shrinking that middle layer it'll lead to an increase in inequality potentially in a concentration of institutional power dependency on fewer highly skilled operators which is you know some of that stuff that we're starting to see now and it's only going to get

worse but also it's going to create you know some institutional fragility around not having not having those talent pipelines to continue to build and grow the organization and so you know I think I think there's a bunch of real issues around the resiliency of organizations if we're not actually building that building that full -length capability we're not building tacit

knowledge and understanding we're not accumulating those kinds of things we're not building that senior expertise because we haven't had it.

Junior level jobs are going to require more experience to get because you're going to need to be more experienced to understand the outputs of the AI.

A Small-Country View: Capacity and Critical Infrastructure

And then from a tiny country perspective, you know, when I think about New Zealand as a whole, if we're not creating those senior engineers, the cybersecurity experts, the infrastructure operators, the experienced analysts, this could become a real concern for us in

pretty short order and so it matters for critical infrastructure it matters for defense it matters for resilience matters for governance and so you know this is the it can be a real capacity issue not just an economic one but you

What AI Improves: Productivity, Learning, and New Opportunities

know it's not all doom you know cuz I think you know the great thing about AI is it does democratize creation it does move it does move creation to a broader a range of people.

It does accelerate learning. It can increase productivity. It does help small teams compete with larger organizations because it does accelerate that delivery.

And it enables, you know, rapid prototyping and all of the great things that can bring products to market and actually grow things. And it can create new jobs.

And so the transition from the way that we have

Open Questions: Redesigning How We Build Expertise

been to the way that we want to be is incredibly important and so my concern is less that you know AI is bad because it's not my concern is whether as a society we're preserving those pathways to actually create expertise and so I've got some questions I don't have the answers because you know this is a hard

problem and I am just one tiny blue haired lady so you know some of the the things that I'm thinking about for my own organization is I'm writing AI strategy and thinking about, too, for New Zealand as to how we kind of manage this stuff. So, you know,

how do we actually preserve junior hiring, should we? And what does that look like? Should we be deliberate to make sure that we are still creating those junior positions?

Should we make sure that we are augmenting, not replacing with AI and making sure that that's part of our strategic delivery, and thinking about how we actually train judgment in that AI -insisted environment.

Organizational and Government Choices

And then from a government perspective, you know, the apprenticeship model that we've known and loved is obviously changing, so how do we redesign that? As public institutions,

so coming from the public sector for the last 20 -ish years, you know, this matters to me, And so, you know, should we be actively building and maintaining that sort of human capacity to make sure that we can straddle that bridge? And, you know, what becomes strategically important?

What are the things that we need to build in people to be able to maintain that human in the loop review of outputs? outputs.

Conclusion: Preserving Pathways to Create New Experts

Every profession depends on a generation that actually learned from doing imperfect work. If AI removes the imperfect work, then we need a way to create new experts.

And that's kind of me so you know i don't have the answers yet and you know i'm keen for us to have conversations following this because i do think you know i'm not the only one thinking about this

um it's kind of nice to see it starting to pop up in my linkedin alongside all of the um enthusiasm and excitement for replacing humans with computers um but yeah i'm really keen to kind of continue those discussions and see where we get to and if