Tech Disruptions

AI's Job Threat: The New Metric That Changes Everything

March 12, 202616:52Tech Disruptions

This episode challenges conventional wisdom about AI's impact on jobs, revealing that it is currently affecting older, more educated, and higher-earning workers, contrary to popular belief. It introduces a new "observed exposure" metric, which combines theoretical AI capabilities with real-world usage data to provide a more accurate picture than past predictions. Listeners will learn about the significant gap between AI's theoretical potential and its actual, counterintuitive implementation in professional settings.

Key Takeaways

Detailed Report

AI's impact on the labor market has long been a subject of intense speculation, often leading to dire predictions that have yet to fully materialize. However, new research introduces a critical shift in how we measure this impact, moving beyond theoretical potential to focus on "observed exposure"—what AI is actually doing in the real world. This new metric reveals a surprising truth: the jobs currently most exposed to AI are often held by highly educated, higher-earning individuals, challenging much of the conventional wisdom.

The Flaws in Past Predictions

For years, predictions about AI's job threat have been notoriously difficult to get right. Early 2010s studies, for instance, suggested a quarter of US jobs were vulnerable to offshoring, yet many of those sectors later saw healthy employment growth. Similarly, research on robot-induced job losses has yielded conflicting results, and even the massive "China shock" in manufacturing still sparks debate among economists regarding its true scale of job displacement. The core problem has been a reliance on theoretical capability: what AI *could* do, rather than what it *is* doing.

Introducing "Observed Exposure"

The new "observed exposure" metric changes the game by combining AI's theoretical capability with real-world usage data. Instead of asking if a self-driving car *could* navigate a race track, it asks if it's *actually* driving kids to school, and how much of that route is truly automated. This distinction is crucial: if AI merely helps you do your job faster, it's augmentation; if it performs the entire task for you, it's automation, carrying a different displacement risk.

The metric leverages three key data sources: O*NET for job tasks, Anthropic's own Economic Index for usage data, and the Eloundou et al. theoretical capability score (beta), which assesses if a large language model (LLM) alone can double task speed or requires additional tools. The observed exposure metric then filters these theoretical possibilities through the lens of actual, professional-setting usage, weighting automated uses more heavily.

The Gap Between Potential and Reality

A significant finding is the "gap" between what AI *can* do and what it's *actually* doing. For example, "Computer & Math" occupations theoretically show a 94% LLM penetration for tasks. Yet, the observed exposure for this same category is only 33%. This two-thirds gap highlights that AI, despite its incredible theoretical muscle, is only flexing a small portion of it in the real world.

This discrepancy arises from several factors: current model limitations, legal constraints, specific software requirements, the need for human verification in critical tasks, and the sheer inertia and cost of integrating new systems. For instance, an LLM might theoretically be able to "authorize drug refills" (a beta of 1.0), but practical hurdles like legal requirements and patient safety mean this isn't observed as automated usage in the real world. This gap represents the frontier where AI's real-world impact will unfold as capabilities advance and adoption spreads.

Who is Most Exposed?

Based on data from Anthropic's Claude, the jobs seeing the highest "observed exposure" are:

  • Computer Programmers: With 75% observed coverage, coding is a prime area for AI's actual usage.
  • Customer Service Representatives: Their main tasks are increasingly handled by first-party API traffic.
  • Data Entry Keyers: Primary tasks like reading source documents and entering data show significant automation, around 67% coverage.

Conversely, about 30% of workers have zero observed coverage, typically in roles that are physical, unpredictable, or highly social. These include jobs like Cooks, Motorcycle Mechanics, Lifeguards, Bartenders, Dishwashers, and Dressing Room Attendants.

The Demographic Surprise

Perhaps the most counterintuitive finding is the demographic profile of the most exposed workers. Comparing the top quartile of observed exposure to the 30% with zero exposure (using pre-ChatGPT data), the research reveals striking differences:

  • The more exposed group is 16 percentage points more likely to be female, 11 percentage points more likely to be white, and almost twice as likely to be Asian.
  • Crucially, they earn 47% more on average and have significantly higher levels of education. For instance, people with graduate degrees make up 17.4% of the most exposed group, compared to only 4.5% of the unexposed group.

This challenges the narrative that AI primarily threatens low-wage, low-skill workers. Instead, it suggests that highly educated, well-paid knowledge workers are in the crosshairs because their roles often involve tasks—like writing, analysis, synthesis, and code generation—that LLMs are exceptionally good at automating or augmenting.

The Unemployment Paradox

Despite this significant observed exposure, the data shows no systematic increase in unemployment for workers in the most exposed occupations compared to the least exposed since late 2022. The trends in unemployment rates have been largely similar, with any slight increase for the exposed group being statistically insignificant. This "unemployment paradox" is a critical finding, as unemployment is considered the most direct measure of economic harm.

This doesn't mean AI is harmless, but rather that its current level of automation and augmentation isn't leading to widespread job losses on the scale of a "Great Recession for white-collar workers"—a level of disruption the analysis *would* have been able to detect.

The "Canary in the Coal Mine": Impact on Young Workers

While general unemployment isn't spiking, a subtle but significant signal is emerging, particularly for younger workers. The research finds suggestive evidence that hiring of *younger workers* (aged 22 to 25) has slowed in these exposed occupations. The job finding rate for this group entering exposed occupations decreased by about half a percentage point in 2024 compared to less exposed jobs, translating to an estimated 14% drop in the job finding rate compared to 2022.

This suggests that companies might be using AI to increase the productivity of their existing, more experienced workforce, rather than hiring new, less experienced talent for those same tasks. For new graduates, this could mean facing a tougher time landing entry-level jobs in fields like programming or customer service. This shift indicates that AI's impact might manifest not through mass layoffs, but through a more gradual, insidious drying up of entry-level opportunities, necessitating a re-evaluation of education and career pathways for the next generation.

Show Notes

Source Materials

References & Resources

  • Anthropic Economic Index: A proprietary index developed by Anthropic to track real-world usage data of AI in professional settings, used as a key component in calculating "observed exposure."
  • Brynjolfsson et al. paper: A related research paper that reported a 6-16% fall in employment for younger workers (aged 22-25) in AI-exposed occupations, primarily attributing it to a slowdown in hiring rather than increased separations.
  • ChatGPT: A popular large language model developed by OpenAI, whose public release in late 2022 marked a significant moment in AI's mainstream adoption and is used as a reference point for the timing of AI's impact.
  • Claude: Anthropic's family of large language models, frequently referenced in the research for its real-world application data and usage in professional settings.
  • Eloundou et al. theoretical capability score (beta): A seminal 2023 research paper that introduced a metric (often called 'beta') to score tasks based on whether a Large Language Model alone could make them twice as fast, representing theoretical AI capability.
  • O*NET: A comprehensive database of occupational information for the United States, used in this research to categorize and analyze job tasks.

Glossary

  • Augmentation: When artificial intelligence assists a human worker, helping them perform tasks faster or more efficiently, rather than replacing the human entirely.
  • Automation: When artificial intelligence performs an entire task independently, potentially displacing the need for human involvement in that specific task.
  • China Shock: A term referring to the economic impact on manufacturing employment in developed countries, particularly the U.S., due to increased competition from China's rapid rise as a global manufacturing power.
  • Current Population Survey (CPS): A monthly survey of households conducted by the U.S. Census Bureau and the U.S. Bureau of Labor Statistics, used to collect data on employment, unemployment, and other labor force characteristics.
  • LLM (Large Language Model): A type of artificial intelligence program trained on vast amounts of text data, capable of understanding, generating, and summarizing human-like text.
  • Observed Exposure: A new metric that measures the actual, real-world usage and automation of AI in professional tasks, combining theoretical AI capability with data on how AI is currently being deployed and used.
  • Offshorability: The ease with which certain job tasks or entire jobs can be moved and performed in other countries, often due to lower labor costs.
  • Theoretical Capability: The potential for artificial intelligence to perform tasks, often measured by how much an AI *could* speed up a task, regardless of whether it's currently being used in practice.
  • Unemployment Paradox: The observation that despite significant AI exposure in certain job categories, there has not been a widespread increase in unemployment rates for those workers, at least not yet.

Sources / References

Full Transcript

HostOkay, so everyone's been screaming about AI taking our jobs, right? The robots are coming for the call center, for the truck driver, for the factory floor. But what if I told you the jobs most exposed to AI right now, the ones seeing actual real-world automation, are held by people who are *older*, *female*, *more educated*, and earning *47% more* than the average worker?
ExpertAnd not only that, but despite this significant exposure, we're not seeing a massive wave of unemployment. Yet. That's the real kicker here. It defies so much of the conventional wisdom.
HostWait, so the highly paid, highly educated people are actually the *first* in the crosshairs, at least in terms of task automation? That feels counterintuitive to every sci-fi movie I’ve ever seen.
ExpertIt absolutely does. And it’s why this new research introduces a completely different way of looking at AI's impact, moving beyond just theoretical potential to what's actually happening in the trenches.
HostAlright, so let's dig into that, because for years, we've had these grand pronouncements about AI's impact on the labor market. Remember the early 2010s? Everything was about "offshorability" or "robot-proof" jobs. And then... not much happened. Or at least, not in the way everyone predicted. Why has it been so hard to get this right?
ExpertExactly. The paper really highlights this history of humility, right? They point to prominent studies from a decade ago that said roughly a quarter of US jobs were vulnerable to offshoring. Fast forward, and most of those jobs have seen healthy employment growth. Or take robots – studies on their employment effects reach opposing conclusions. Even the "China shock" in manufacturing, which was massive, still has economists debating the scale of job losses.
HostSo, it's not just that the crystal ball was foggy; it was actively misleading in some cases. What makes AI different, or perhaps, what makes *our approach* to measuring AI different, that might finally give us some clarity?
ExpertThe big problem is that most past predictions, and even current ones, tend to focus on theoretical capability. What *could* AI do? The new metric, what they call "observed exposure," changes the game by combining that theoretical capability with *real-world usage data*. It's like saying, "Yes, a self-driving car *could* navigate a race track at 200 miles an hour," but "observed exposure" asks, "Is it actually driving kids to school at 30 miles an hour, and how much of that route is truly automated versus a human taking over?"
HostThat's a huge distinction. So, instead of just the hypothetical, we're looking at what's actually being implemented and used in professional settings, and crucially, how much of that usage is *automated* versus just *augmentative*?
ExpertPrecisely. They're weighting automated uses more heavily, because that's where the real displacement risk lies. If AI is just helping you do your job faster, that’s augmentation. If it’s doing the entire task *for* you, that’s automation, and that’s a different beast. They're also focusing on work-related contexts, which is key. My kids using Claude to write a story for school isn't the same as a financial analyst using it to draft a report.
HostSo, it's about shifting from potential energy to kinetic energy, from "what if" to "what is." And this "observed exposure" metric uses three main data sources, right? O*NET for job tasks, their own Anthropic Economic Index for usage, and then the Eloundou et al. theoretical capability score.
ExpertThat's right. The Eloundou et al. metric, often called beta, scores tasks based on whether an LLM alone can make them twice as fast, or if it needs additional tools. But the observed exposure metric is the crucial layer on top. It asks: of those tasks that an LLM *could* theoretically speed up, which ones are *actually* seeing automated usage in professional settings right now?
HostAnd that leads us directly to this fascinating "gap" you mentioned earlier. The paper shows that AI is still "far from reaching its theoretical capability." Can you unpack that for us? Because I think a lot of people assume if AI *can* do something, it *is* doing it everywhere.
ExpertThat's the common misconception, and this paper provides a really stark visual representation of that gap. Think of it like this: AI has this incredible theoretical muscle, but it's only flexing a small portion of it in the real world right now. The Eloundou et al. metric, that theoretical capability, shows that for categories like "Computer & Math" occupations, LLMs *could* theoretically penetrate 94% of tasks. That's massive.
HostNinety-four percent! That sounds like a sci-fi nightmare for programmers.
ExpertIt does, right? But then you look at the "observed exposure" – what people are *actually* using an LLM like Claude for in professional settings – and for that same "Computer & Math" category, it's only 33%. That's a huge difference. A two-thirds gap between what's possible and what's practical or deployed.
HostSo, AI *could* write almost all the code, but in reality, it's only assisting with a third of the tasks right now. Why such a big discrepancy? Are businesses just slow to adopt?
ExpertIt’s a mix of factors. Some tasks that are theoretically possible might not show up in usage because of current model limitations – they're just not *good enough* yet for certain nuances. Then you have legal constraints, specific software requirements, the need for human verification in critical tasks, or just sheer inertia and the cost of integrating new systems. For example, the theoretical capability might say an LLM can "authorize drug refills." That's a beta of 1.0.
HostWhich sounds incredibly efficient, a doctor's office would love that.
ExpertAbsolutely. But they haven't observed Claude actually doing that task in the real world. Why? Because of all those practical hurdles – legal requirements, patient safety, system integration. So, while the theoretical potential is there, the observed application is nil, or at least not registered as automated usage.
HostThat makes so much sense. It's the equivalent of having a super-fast self-driving car but needing a lawyer to approve every turn, an IT specialist to manage the GPS, and a human to press the "go" button every time. The potential is there, but the real-world deployment is bottlenecked.
ExpertPerfect analogy. And the paper emphasizes that this gap is what their "observed exposure" measure is tracking. As capabilities advance and adoption spreads, that "red area" of actual usage will grow to cover the "blue area" of theoretical possibility. That's the real story unfolding.
HostSo, who *is* actually on the hot seat with this new metric? Who are the jobs seeing the highest "observed exposure" right now? Because I think a lot of our listeners, myself included, have some preconceptions about that.
ExpertWell, based on their data from Anthropic's Claude, the top spot, perhaps not surprisingly given the previous point about Computer & Math, goes to **Computer Programmers**. They have a 75% observed coverage.
HostSeventy-five percent! That's massive. So, coding is definitely where a lot of the rubber is meeting the road for AI in terms of actual usage.
ExpertAbsolutely. The paper mentions that Claude is extensively used for coding. Following them are **Customer Service Representatives**, whose main tasks are increasingly seen in first-party API traffic, and then **Data Entry Keyers**, whose primary task of reading source documents and entering data sees significant automation. They're around 67% coverage.
HostOkay, so programmers, customer service, data entry. That makes intuitive sense for areas where AI excels at repetitive, structured, or information-processing tasks. But what about the other end of the spectrum? Who are the least exposed?
ExpertAt the bottom end, about 30% of workers have zero coverage. These are tasks that appeared too infrequently in their data to meet the minimum threshold, or simply aren't suitable for LLM automation yet. This group includes jobs like **Cooks, Motorcycle Mechanics, Lifeguards, Bartenders, Dishwashers, and Dressing Room Attendants.**
HostOkay, so hands-on, physical, unpredictable, or highly social roles are still largely untouched. That aligns with some predictions, but the top-end jobs are definitely a surprise. It's not just the "low-skill" roles that are seeing this observed automation.
ExpertExactly. And that's where the demographic surprise comes in. The paper takes the top quartile of observed exposure and compares those workers to the 30% with zero exposure, looking at data from before ChatGPT's release. The differences are striking.
HostTell me. This is the part that really made me do a double-take.
ExpertSo, the *more exposed* group is 16 percentage points more likely to be female, 11 percentage points more likely to be white, and almost twice as likely to be Asian. But here's the kicker: they earn **47% more** on average, and have significantly higher levels of education.
HostForty-seven percent more! And higher education. So, the narrative of AI coming for the low-wage, low-skill worker first is just... not playing out in the observed data. It's the opposite.
ExpertIt really is. For example, people with graduate degrees make up only 4.5% of the unexposed group, but they're 17.4% of the *most exposed* group. That's almost a fourfold difference. This challenges the idea that higher education or higher pay inherently provides a shield against automation. In fact, it might make you *more* exposed right now because many of those roles involve tasks that are highly amenable to LLM assistance – things like writing, analysis, synthesis, code generation.
HostSo, if you're a highly educated, well-paid knowledge worker, your tasks are precisely the ones that current LLMs are best at automating or augmenting. It's not about being "replaced" yet, but your day-to-day work is becoming more and more AI-intervened.
ExpertPrecisely. And this makes perfect sense when you consider the nature of LLMs. They excel at language-based tasks, information processing, and problem-solving within defined parameters. These are often the core functions of those higher-paid, more educated roles.
HostOkay, so we've got a new, more accurate way to measure exposure, and we know *who* is exposed – high-earning, educated knowledge workers. But you also said that despite this, we're not seeing a massive increase in unemployment. That's the "unemployment paradox," right? What's going on there?
ExpertYes, the "unemployment paradox" is a critical finding. The paper's core analysis focuses on unemployment because, as they argue, it's the most direct measure of economic harm. A worker who is unemployed wants a job and hasn't found one. Other metrics, like a decline in job postings, don't necessarily signal harm because those workers might shift to new roles.
HostSo, unemployment is the clearest signal of job displacement. And what does the data say since late 2022, after ChatGPT really burst onto the scene?
ExpertTheir analysis of the Current Population Survey data shows no systematic increase in unemployment for workers in the most exposed occupations compared to the least exposed. The trends in unemployment rates have been largely similar between the two groups. If anything, there's a slight, statistically insignificant increase for the exposed group, but it's indistinguishable from zero.
HostThat's a huge relief, honestly. All the doomsday predictions haven't materialized into mass layoffs for programmers or customer service reps. But does that mean we're completely in the clear? Or is it too early to tell?
ExpertIt's definitely too early to pop the champagne. The paper actually gives a good sense of scale. They say that if AI caused a "Great Recession for white-collar workers," meaning unemployment rates in the top quartile of exposure doubled from, say, 3% to 6%, their analysis *would* be able to detect that. So, the fact that they *haven't* detected anything significant means we haven't hit that level of disruption yet.
HostSo, the current level of automation and augmentation isn't leading to widespread job losses. That's good news for now. But there's a subtle signal, a "canary in the coal mine," that *is* starting to show up, particularly for younger workers.
ExpertAbsolutely. This is where the nuanced evidence emerges. While general unemployment isn't spiking, the paper *does* find suggestive evidence that hiring of *younger workers* has slowed in these exposed occupations. Specifically, they looked at workers aged 22 to 25.
HostAh, the new entrants to the workforce. This is a big one. So, it's not about people getting fired, but about new people not getting *hired*?
ExpertPrecisely. Brynjolfsson et al., in a related paper, reported a 6-16% fall in employment for this age group in exposed occupations, attributing it primarily to a slowdown in hiring rather than increased separations. This new paper corroborates that. They found that the job finding rate for young workers entering exposed occupations decreased by about half a percentage point in 2024, compared to the less exposed jobs.
HostHalf a percentage point might not sound like a lot, but what does that translate to?
ExpertIt translates to an averaged estimate of a 14% drop in the job finding rate compared to 2022 for these young workers in exposed occupations. That's pretty significant for new graduates trying to break in. And crucially, there's no such decrease for workers older than 25. This suggests companies might be using AI to increase the productivity of their existing, more experienced workforce, rather than hiring new, less experienced talent to do those same tasks.
HostSo, if you're a young programmer, for example, who's just graduated, you might be facing a tougher time landing that first job because the company can now use AI tools to make their current senior programmers even more efficient, reducing the need for entry-level hires.
ExpertThat's one of the key interpretations. The paper acknowledges a few caveats, like young workers might simply be staying in school longer, taking different jobs, or that job transitions can be mismeasured in surveys. But the signal is there. It's a subtle shift, not a dramatic explosion, but it's definitely something to watch. It suggests that AI's impact might manifest not through mass layoffs, but through a more gradual, insidious drying up of entry-level opportunities in certain fields.
HostThat's a critical distinction. It's a different kind of disruption. Less of a sudden earthquake, more of a slow, shifting tide that makes it harder for new boats to get to shore.
ExpertExactly. And it means we need to think differently about how we prepare the next generation for the workforce. If these jobs are becoming harder to enter, what does that mean for education and career pathways?
HostAlright, this has been an absolutely fascinating deep dive into how AI is *actually* impacting the labor market, not just theoretically. So, let's distill this down. What are the key takeaways our listeners should really internalize from this research?
ExpertFirst, we need to be skeptical of past predictions and buzzwords. The new "observed exposure" metric, which focuses on real-world, automated usage rather than just theoretical capability, is a much more grounded way to understand AI's impact.
HostSecond, there's a massive gap between what AI *can* do and what it's *actually* being used for in professional settings. This gap offers a crucial buffer, but it's also the frontier we need to watch as AI capabilities and adoption continue to grow.
ExpertThird, the jobs and demographics most exposed to AI right now are not what many might expect. We're talking about higher-paid, more educated, often older and female knowledge workers, like computer programmers and customer service reps, rather than solely low-wage, physical laborers.
HostFourth, despite this significant observed exposure, we are *not* currently seeing widespread unemployment increases. So, the sky isn't falling for the general workforce, at least not yet, and not in terms of direct job displacement.
ExpertAnd finally, the most telling signal, the "canary in the coal mine," is a subtle but significant slowdown in hiring for young workers, specifically 22-25 year olds, in these highly exposed occupations. This suggests a potential shift in how companies staff roles, favoring augmentation of existing talent over new entry-level hires.
HostThat's a lot to chew on. So, for our listeners, I'd leave you with this: How does understanding "observed exposure" change your perspective on your own career path or hiring strategies? And if this trend of slowed hiring for young workers in exposed fields continues, what ripple effects might that have on education and economic mobility in the coming years?