Is AI Actually Displacing Workers Right Now, or Is This Still Theoretical?
This is the first in a series exploring what the research actually says about AI and your career — and what to do about it.
The headlines make this sound simple. "AI will eliminate millions of jobs." Or, from the other side: "AI is just a tool, relax." Neither version is useful, because the real answer is more specific — and more important — than either.
I spent the past year reading the actual research on this, not the takes about the research, but the studies themselves. Three of them, from three very different sources, paint a picture that's more nuanced and more actionable than anything I've seen in the popular conversation. Here's what they found.
1. It's already showing up in hiring data
The strongest evidence comes from a Stanford team led by Erik Brynjolfsson. Their study, "Canaries in the Coal Mine," uses monthly payroll records from ADP — the largest payroll processor in the US — covering millions of workers across tens of thousands of companies. This isn't a survey or a projection. It's what actually happened in the American labor market between late 2022 and September 2025.
But here's what makes this more than a scary statistic. In those exact same roles, employment for workers over 30 didn't fall. It grew — by 6 to 12%. And in roles with low AI exposure, like nursing aides and health assistants, young workers actually grew faster than older ones. The pattern is specific: it shows up where AI exposure is high, and it shows up for young workers, not everyone.
The study also looked at overall employment across the economy. It's still growing. The national unemployment rate is low. The sky is not falling. But employment growth for 22-to-25-year-olds has gone flat since late 2022, and the flatness is driven specifically by declines in AI-exposed jobs.
So who's affected, and why?
The researchers propose a mechanism that I think is the most important idea in the paper: it comes down to what kind of knowledge your work relies on.
Young workers bring mostly codified knowledge — the things you learn in school, from textbooks, from training programs. How to write code in a particular language. How to follow a customer service script. How to run a standard financial model. This is exactly the kind of knowledge that AI systems are now very good at replicating.
Experienced workers bring something different: tacit knowledge. The judgment calls that come from years of doing the work. Knowing which stakeholder will push back on a proposal and why. Understanding that the code compiles fine but the architecture will cause problems six months from now. Sensing that a customer's real issue isn't what they said on the phone. None of this is written down anywhere. None of it exists in a training dataset.
“AI can do the codifiable parts of a job. It struggles with the tacit parts.”
So companies are making a rational (if uncomfortable) calculation: hire fewer people at the entry level where the work is mostly codifiable, and keep (or hire more) experienced workers whose tacit expertise AI can't replicate.
There's one more finding that matters: the researchers controlled for almost every obvious alternative explanation.
Maybe this is just a tech-sector downturn?
No — the pattern holds after excluding technology firms and computer occupations entirely.
Maybe it's an interest rate effect?
No — it holds after controlling for firm-level economic shocks.
Maybe it's a COVID hiring hangover?
No — the pattern didn't exist before late 2022 and appeared specifically when generative AI tools became widely available.
Maybe it's a remote-work thing?
No — it holds for both teleworkable and non-teleworkable jobs.
This is displacement happening right now, in real payroll data, at scale. But it's not the kind most people picture. Nobody got a pink slip that said "replaced by AI." Instead, companies quietly stopped backfilling roles, reduced junior hiring, and let AI absorb the tasks that entry-level workers used to do. The teams got smaller, the tools got better, and the headcount never came back.
2. Whether it affects you depends on how AI is being used in your role
If the Canaries study shows that displacement is happening, the Anthropic Economic Index shows where and how.
Anthropic (the company behind Claude) has been tracking how millions of people actually use their AI system, mapping each conversation to specific work tasks. Their most recent report, published in January 2026, gives us the clearest picture of AI's real-world footprint in the economy.
The scale is significant: by November 2025, roughly half of all jobs had AI being used for at least a quarter of their associated tasks. That's not a projection. That's observed usage.
But the report's most important contribution isn't the breadth of usage — it's the distinction between how AI gets used, which turns out to determine everything.
Automation vs. augmentation: why the mode matters more than the exposure
When someone gives AI a task and takes the output — "write this email," "generate this code," "summarize this document" — that's automation. The AI is doing the work. The human reviews and sends. This is what Anthropic calls "directive" use.
When someone works with AI iteratively — drafts something, gets feedback, revises, asks follow-up questions, uses AI to learn about a topic — that's augmentation. The human is still doing the work, but with an AI assist. Think of it as having a very fast, very knowledgeable colleague looking over your shoulder.
On the consumer side (people using Claude directly), the split is roughly 52% augmentation to 45% automation. Most individual users are working with AI, not handing work to it.
On the enterprise side (businesses building AI into their products and workflows), it's the opposite. Automation dominates. Companies are increasingly building AI into their systems to handle tasks that used to require people — back-office workflows, document processing, email management, scheduling.
The mode of AI usage — automation or augmentation — is what separates displacement from enhancement.
Which tasks is AI actually doing?
Usage is heavily concentrated. The single most common task — modifying software to fix bugs — accounts for 6% of all Claude usage. The top ten tasks account for about a quarter. Computer and math-related tasks make up a third of consumer usage and nearly half of enterprise usage.
This concentration matters for an unexpected reason: the tasks people bring to AI tend to require more education than average tasks in the economy. AI is being used for higher-skill work, not lower-skill work. If AI takes over those higher-skill tasks, the remaining human work becomes less skilled — a deskilling effect. But this plays out differently across roles. A travel agent who loses the complex planning work to AI is left with routine ticket transactions. A property manager who loses bookkeeping to AI is left with contract negotiations and stakeholder management. Same AI, opposite effects on the worker's role.
One more finding: AI's success rate drops as task complexity increases. Simple, well-defined tasks — AI handles them reliably. Complex tasks requiring judgment across multiple domains — AI struggles. This sets a natural boundary on what gets displaced: the codifiable, checkable tasks go first. The judgment-intensive, context-dependent tasks remain human.
3. The economy will create more jobs than it destroys. That doesn't help you individually.
The first two studies look backward at what's already happened. The World Economic Forum's Future of Jobs Report 2025 looks forward — surveying over 1,000 of the world's largest employers about what they're planning through 2030.
The top-line numbers are actually reassuring: 170 million new jobs projected to be created; 92 million displaced. Net gain: 78 million jobs. The economy, in aggregate, will be fine.
But that net-positive headline hides a problem that the report itself identifies as the number-one barrier to business transformation: the skills mismatch.
The 92 million jobs on the displacement side and the 170 million jobs on the creation side don't require the same skills. The fastest-declining roles are clerical and administrative — data entry, secretarial work, accounting clerks. The fastest-growing roles are in AI, big data, cybersecurity, renewable energy, and healthcare. These are fundamentally different skill sets.
“39% of existing workplace skills will be outdated or transformed by 2030.”
When asked what's holding their businesses back from this transformation, 63% of employers pointed to skills gaps — more than any other barrier.
What this means for people, not economies
The macro numbers tell you that the labor market will adjust. History tells you the same thing — every major technology transition has eventually created more jobs than it destroyed. But "eventually" can be a very long time, and "in aggregate" doesn't help if you're one of the 92 million on the wrong side of the transition.
Think of it this way. If you're a financial analyst whose daily work consists mostly of building models in Excel and writing summary reports, the macro fact that the economy will create 78 million net new jobs doesn't change your situation. The question for you isn't "will there be jobs?" — there will. The question is: "will there be jobs that match what I know how to do?" And that depends on which parts of your current role are codifiable (and therefore automatable) versus which parts rely on judgment, relationships, and expertise that AI can't replicate.
The WEF report, to its credit, acknowledges this explicitly. The 85% of employers who say they plan to upskill their workforce is encouraging. But planning and execution are different things, and the 39% of skills expected to change in five years is a pace that most corporate training programs aren't designed to match.
So, is AI displacement real?
Yes. Three independent sources — Stanford's payroll analysis, Anthropic's usage data, and the WEF's employer surveys — all point to the same conclusion, at different levels of the economy:
Three Levels of Evidence
At the individual level, displacement is already showing up in hiring data. Young workers in AI-exposed roles are losing ground, while experienced workers in those same roles are not. The mechanism is clear: AI replaces codifiable tasks and complements tacit expertise.
At the task level, AI is already being used for a significant share of work across half of all roles. Where AI automates complete tasks, employment declines. Where it augments human work, it doesn't.
At the economy level, the net numbers are positive — more jobs will be created than destroyed. But the new jobs require different skills than the old ones, and the gap between displaced skills and demanded skills is the defining challenge of this transition.
The displacement isn't the dramatic, headline version — millions of people marched out of offices on the same day. It's quieter. Companies don't backfill roles. Teams operate lean and discover they can manage. Junior hiring freezes become permanent. Task by task, the work shifts, and by the time you notice, the job has already changed shape.
Understanding this is the first step.
Studies referenced:
- Brynjolfsson, Chandar & Chen, "Canaries in the Coal Mine: Six Facts about the Recent Employment Effects of Artificial Intelligence," Stanford Digital Economy Lab, November 2025
- Anthropic, "The Anthropic Economic Index: Economic Primitives," January 2026
- World Economic Forum, "Future of Jobs Report 2025," January 2025