The Shifting Career Ladder
AI is changing how work works and quietly removing the pathways through which young people learn to become experts.
For all the noise around AI and jobs, the most important story may not be mass unemployment, but something subtler and more structural. In the midst of all the hype and dread we might actually be missing the shifts that actually matter.
The most important shift for knoweldge workers will not be about the number of jobs, but the type of jobs and how people, especially new entrants to the labor market can access them. AI is reshaping how people build skills, enter professions, and move along the career ladder and through the labour market.
In this conversation, I sit down with Matt Sigelmen founder of LightCast and now the President of Burning Glass Institute. Matt has dedicated his career to understanding the labor market and helping society improve the connections within in it.
Matt and I explore why people and opportunities are often only “a few skills apart,” why entry-level work may be losing its traditional role as the first rung of expertise, and why schools, universities, and employers now need to rethink the pathways that turn potential into mastery.
[The interview has been edited for clarity and brevity]
You’ve built a career at the intersection of data, work, and human opportunity. What moment or experience first made you believe that understanding skills and labor markets could actually change lives? Was there a person or a pivot point that made this feel like a mission, not just a business?”
When I first came to this work, I was very focused on solving the job-matching problem. How do you take a job opening, receive thousands of responses, and stack-rank the candidates best suited for that role? Conversely, how do you take a résumé and figure out which jobs are most relevant to that individual?
We built a very good mousetrap for that. At its core, better matching is a pattern-matching problem: whose career history, skills, and experiences position them for a given role in ways that make them more likely to succeed? I remain very proud of the “little business” we built around that.
But my thinking really began to shift at the start of the Great Recession, around 2008. Until then, we had mostly been helping to solve the job-matching problem one-to-one: how do you support employers in screening candidates, or help individual job seekers find relevant opportunities?
Then we received a contract from New York State to work with the Department of Labor’s job centers. They were suddenly swamped with hundreds of job seekers who needed urgent help, many of whom had very different profiles from the job seekers the system was used to serving. A lot of them were highly educated and highly skilled, coming primarily from the financial sector. The Department simply was not equipped to support them, and that is where we stepped in.
What became clear to me was that the system was still approaching this one person at a time, which is largely how it has always worked and, in many ways, still works.
That is an incredibly inefficient model because it relies on cosmic coincidence: does the perfect job happen to exist for the perfect candidate on the exact day they are looking for it? More often than not, instead of cosmic coincidences, you get misconnections.
You have to think about this as a many-to-many connection problem. You need a bird’s-eye view of the mismatches in the labor market. The labor market’s biggest tragedy is also its biggest opportunity: people and opportunities are often only a few skills apart, but they cannot see each other.
That realization created a major pivot in my thinking. I moved from building matching engines to trying to understand the labor market itself, and what it would take to unlock it.
More practically, that means bringing forward data with enough granularity to show where opportunity exists and where the potential for connection lies.
I have operated for a long time under the premise that, with the right data, we would be capable of all sorts of transformations. We did a lot of that work at Lightcast. The company continues to produce powerful datasets, but we are still waiting for the kind of transformations that those data can inform - and that are truly needed.
You spent 20 years building Lightcast into the world’s leading labor market data company. Then you walked away to found an independent nonprofit institute. What prompted that shift? Did you feel the most important work still hadn’t been done?
I started the Burning Glass Institute with the belief that achieving big transformations means starting not with the data but with the change you are seeking.
Big transformations require different kinds of data and different kinds of connections. At the Institute, we look at the transformations that could create better economic mobility and enable people to build careers that are both upwardly mobile and resilient. We then work backwards and ask: what data is needed to unlock those realities?
That means working with education systems around the world, regional coalitions, public agencies, and employers to help make this possible.
There’s enormous “noise” about AI and jobs (both utopian and apocalyptic). But your data is rare in that it actually tracks what’s happening inside the labor market in real time. What are the two or three findings that have most surprised you?
I think a lot of the discourse around AI and jobs, perhaps most of it, has focused on an often apocalyptic and hyperbolised calculus of body counts: how many people are going to lose their jobs?
I am not going to weigh in directly on what may ultimately come to pass. There is certainly significant potential for real displacement and dislocation. But what we are seeing so far is that the real implication is less about whether you work and much more about how you work.
That also echoes past technological revolutions. General-purpose technologies tend to change how everybody works.
That is important because it is consistent with the evidence we are seeing so far. Much more of the effect we are observing in the labour market today is about significant changes in skill requirements. Augmentation-oriented skills are more likely to grow than other kinds of skills.
What is more surprising is that people have tended to cast augmentation and automation as opposing poles. But what we are seeing is that the jobs experiencing the most automation are also experiencing the most augmentation. The jobs seeing the largest share of tasks automated away are often the same jobs redeploying people into other areas where they can do more. In fact, we see a 0.87 correlation there.
There is some cognitive dissonance in this because our mental model of automation is still shaped by the Industrial Revolution, where work was broken down into discrete and simple tasks. But in the knowledge economy, jobs are more complicated. We end up doing much more than the sum of our tasks.
That same interplay between augmentation and automation also intersects with the importance of skills. There are certain skills where AI makes you both more efficient and more effective. It is not that some skills get automated and others get augmented. Many skills experience both.
Take writing, for example. AI is very good at generating content, and it would be tempting to conclude that AI obviates the need for writing. But it really depends on what you mean by writing. Writing is not just content generation. It is also about evidencing an argument, framing a narrative, and shaping meaning.
That means we now need to think carefully about how higher education embeds these new power skills into the curriculum, so that students leave their education with a higher level of proficiency.
Your research shows that AI is eliminating the entry-level rungs that used to build expertise (the drafting, research, basic analysis that taught junior workers how to think etc.). Can you talk more about that research and specifically touch on how, if those on-ramps disappear, does the next generation actually develop mastery? And who’s responsible for rebuilding that pathway?
There has been a lot of discussion about this expertise upheaval and it’s not theoretical. It is already happening.
We are seeing in the data a very meaningful shift away from hiring people at the entry level and towards hiring people with more experience, which threatens to knock out the early rungs of the career ladder.
This is a real threat, and there are no easy answers. But there are a few promising areas we can address.
First, this is a curricular imperative. Educators need to be deeply aligned with what these changes are, and they need to shift the AI discourse from “how” questions to “what” questions. What do we need to teach? What do we need to keep in the curriculum?
They also need to recognise that one of the reasons employers are shifting away from hiring junior workers is that they are indexing towards greater proficiency. So the curricular challenge is that education systems, which have already struggled to build stronger connections between higher education and work, now need to prepare students for a higher level of proficiency. In effect, students need to be ready to start closer to the middle rather than at the very beginning.
Second, work-based learning is highly effective at ensuring students can apply what they are learning in the real world. There is a lot of evidence for this, whether you are talking about internships, co-ops, or, more broadly, project-based learning and project-based work. The problem is that the same reasons employers are becoming less likely to hire new joiners also create challenges around internships.
We are conducting research on how to make the work interns perform more valuable to employers, because that remains a real challenge. There is also a question of where new models can fit when internships are not available, whether through virtual simulations or other approaches.
More broadly, we need to move beyond degrees alone towards credentials that have real currency in the labour market, and towards the skills that enable people to make an effective transition into work.
There are also responsibilities that lie with employers. As employers come to value expertise more, they need to create pipelines for developing that expertise. Otherwise, they will face a drought. That means finding ways to invest in talent development, even where those investments do not take the traditional form of entry-level hiring.
Finally, we also need to make entry-level jobs more deliberately a pathway to expertise. How do employers reshape work itself to give people greater agency?
All of us are awed by the level of innovation AI allows for, and what we could see here is a golden age of innovation. But for that to happen, we need more workers in jobs where they have the agency to innovate and create. That requires a fundamental reshaping of the workplace, organisational structures, and job definitions.
You’ve said that “70% of K–12 learning objectives need to be taught differently”. That’s a staggering statement. What does a school/university that’s actually getting this right look like? And for educators listening who want to take steps towards that reality, where do they start?”
First, just to put some colour on where the 70% figure comes from: it is based on our research, but any predictive model is exactly that: a probabilistic estimate.
That said, one of the things that impedes effective conversation about curriculum reform, and about creating stronger bridges between education and industry, is the absence of a shared vocabulary. AI is a technology for work. Its impact is therefore felt in work, and in the underlying skills and tasks that make up that work. That is the language employers tend to use.
Education, by contrast, is organised around learning objectives. There is certainly overlap between the two, but the disconnect makes it difficult to translate between learning objectives on the one hand and skills and tasks on the other.
So we built a large-scale knowledge graph to try to connect the two. That is where we saw that roughly 70% of learning objectives may need to be updated through this interplay of automation and augmentation.
The imperative for change is not simply about what we teach. It is about how we teach, how we assess, and what we emphasise.
Take teaching students how to do research. Developing a research plan is something large language models can help you with substantially. LLMs can also help with collecting data. But that means educators will need to place much more emphasis on students’ ability to evaluate and interpret evidence, because they will have access to far more data. Students will also need to think harder about how they translate that evidence into implications for the real world.
All of this has implications for assessment. Today, we often assess what students know. But we need to orient assessment more towards how they know it: what steps they took, what process they used, and how they reached their conclusions.
I hope you enjoyed the last edition of Nafez’s Notes.
I’m constantly refining my personal thesis on innovation in learning and education. Please do reach out if you have any thoughts on learning - especially as it relates to my favorite problems.
If you are building a startup in the learning space and taking a pedagogy-first approach - I’d love to hear from you. I’m especially keen to talk to people building in the assessment space.
Finally, if you are new here you might also enjoy some of my most popular pieces:
The Gameboy instead of the Metaverse of Education - An attempt to emphasize the importance of modifying the learning process itself as opposed to the technology we are using.
Using First Principles to Push Past the Hype in Edtech - A call to ground all attempts at innovating in edtech in first principles and move beyond the hype
We knew it was broken. Now we might just have to fix it - An optimistic view on how generative AI will transform education by creating “lower floors and higher ceilings”.