Will AI Replace Jobs? The Statistics, Decoded (2026)

Let's start with the four numbers you've definitely seen, because every listicle on the internet repeats them, usually wrong.

Number one: Goldman Sachs estimated generative AI could expose the equivalent of 300 million full-time jobs to automation globally (Goldman Sachs, 2023). That's the big scary one. But read the same report and you'll find Goldman says most jobs are complemented by AI, not substituted. They also found that roughly two-thirds of US occupations are exposed to some degree of AI automation, but of those, only about a quarter to half of the workload could actually be replaced (Goldman Sachs, 2023). Oh, and the same report projects AI could raise global GDP by 7%, roughly $7 trillion over ten years (Goldman Sachs, 2023). Funny how that part never makes the headline.

Number two: McKinsey estimates activities accounting for up to 30% of hours worked across the US economy could be automated by 2030 (McKinsey, 2023). Hours, not jobs. And McKinsey's own earlier research found that less than 5% of jobs can be fully automated by existing technology (McKinsey, 2017). Number three: the World Economic Forum projects 92 million jobs displaced and 170 million created by 2030, a net increase of 78 million jobs (WEF, 2025). Read that again. The WEF, the body everyone cites for AI doom, is forecasting a net gain. Number four is the word itself. "Exposed." "Tasks." "Could." That language changes everything, and almost no credible study actually predicts mass job loss.

This is the section nobody wants to write because it's less exciting than the apocalypse. Tough. It's also the truth.

The foundational study here is OpenAI and University of Pennsylvania's "GPTs are GPTs" paper. It found around 80% of the US workforce could have at least 10% of their work tasks affected by large language models (Eloundou et al., 2023). Sounds enormous, until you read the next line: only about 19% of workers may see at least 50% of their tasks impacted (Eloundou et al., 2023). And with LLM-powered software layered on top, the share of tasks that could be significantly accelerated rises to between 47% and 56% of all tasks (Eloundou et al., 2023). Accelerated. Not eliminated. There's a world of difference between "AI makes this task faster" and "this person no longer has a job."

The IMF tells the same story from a different angle. It estimates almost 40% of jobs globally are exposed to AI, broken out as about 60% in advanced economies, 40% in emerging markets, and 26% in low-income countries (IMF, 2024). But here's the part that matters: the IMF estimates roughly half of those AI-exposed jobs may actually benefit from AI integration, becoming more productive, while the other half could see lower labor demand (IMF, 2024). Even Anthropic's own usage data points the same way. Their first Economic Index found AI use leaned toward augmentation, 57%, over automation, 43%, across millions of Claude conversations (Anthropic, 2025). That split shifts report to report, but augmentation has led automation most of the time. When the company building the model tells you it's mostly helping people do their jobs rather than taking them, that's worth a beat.

Here's my favorite part, because this is where the skeptic in me gets to win. Forecasts are models. Models are built on assumptions. And assumptions are where consultants hide the conclusion they got paid to reach. The only honest check on a forecast is the actual data. So what does the actual data say?

The Yale Budget Lab studied the 33 months after ChatGPT's release and found no discernible economy-wide labor-market disruption (Yale Budget Lab, 2025). None. The occupational mix is changing only about one percentage point faster than it did during early internet adoption (Yale Budget Lab, 2025). Read that twice. The thing everyone says is unprecedented is, so far, barely distinguishable from the internet showing up. The OECD backs this up. It found little evidence so far that AI is causing job losses, despite high automation potential, even though occupations at the highest risk of automation account for about 27% of employment on average across OECD countries (OECD, 2023). High potential, quiet reality.

And MIT's Iceberg Index, one of the newest and most-hyped studies, found current AI could technically perform tasks equal to 11.7% of the US workforce, about $1.2 trillion in wages (MIT & Oak Ridge, 2025). Scary number. But MIT itself stresses the index measures exposure, not adoption or actual workforce displacement (MIT & Oak Ridge, 2025). They put the disclaimer right in the report. The gap between the terrifying projections and the flat real-world data isn't a footnote. It's the entire story. Aggregate US unemployment has not moved in any way you can attribute to AI through 2025.

To make the exposure-versus-replacement point concrete, here's how the major studies actually phrase their headline numbers. Notice the verbs. "Exposed." "Could be automated." "Displaced and created." Not one of them says "will be eliminated, net."

One of these things is not like the others. The bottom row, Yale, is the only one measuring what actually happened to real paychecks. Everything above it is a model of what could happen. Keep that hierarchy in your head every time you read a stat.

Now I'm going to stop being comforting, because there's one finding that's real, recent, and specific enough to take seriously. If I only told you the reassuring half, I'd be doing exactly what the doom-mongers do, just in the other direction.

Stanford economists Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen ran the numbers in a paper they called "Canaries in the Coal Mine." They found a 13% relative decline in employment for workers aged 22 to 25 in the most AI-exposed occupations since late 2022 (Stanford, 2025). For 22-to-25-year-old software developers specifically, the decline runs about 20% (Stanford, 2025). That's not noise. That's a signal. And here's the kicker: older, experienced workers in the same roles saw stable or growing employment over the same period (Stanford, 2025). The pain is concentrated entirely at the bottom of the ladder.

Why? Brynjolfsson put it in one brutal sentence. What younger workers know overlaps with what LLMs can already do. A fresh grad whose entire value is "I can write a basic CRUD endpoint and look things up" is competing directly with a tool that does exactly that for pennies. The senior engineer who knows the system, the tradeoffs, the business, and how to wield the AI as a force multiplier is not. Stanford's broader finding makes the rule explicit: employment declines concentrate where AI automates rather than augments, while augmentation-heavy occupations saw stable or rising employment (Stanford, 2025). The fix isn't to hide from AI. It's to know what the LLM doesn't, and to be the one operating it. That single distinction is the difference between being the canary and being the miner.

So if it's not everyone, who is it? The data is remarkably consistent here, and it cuts against the gut instinct. This wave hits cognitive desk work, not physical labor.

The OpenAI/Penn study found the most LLM-exposed work involves writing, programming, math, and data analysis, while physical and outdoor work is the least exposed (Eloundou et al., 2023). Let that sink in. The robot apocalypse everyone pictured was supposed to take the factory jobs. Instead it's the spreadsheet jobs that are most exposed. BLS projections put hard numbers on the routine end: bank teller employment is projected to decline about 15% from 2023 to 2033, eliminating roughly 51,400 jobs (BLS, 2024). McKinsey flags office support, customer service, and food service as among the biggest projected decliners (McKinsey, 2023), and it projects an additional 12 million occupational transitions may be needed in the US by 2030, with lower-wage workers up to 14 times more likely to need to change jobs (McKinsey, 2023).

The least exposed jobs are the ones that need a body in a room: physical presence, unpredictable environments, fine motor skills, and human judgment. Skilled trades. Care work. The stuff people assumed would go first is the stuff that's safest. If your job is mostly producing text or shuffling structured data through a predictable process, you're in the exposed bucket. If your job requires showing up somewhere unpredictable and using your hands or your read of a human being, you're not. And if your job is building the AI itself, you're not just safe, you're the one writing the disruption everyone else is afraid of.

I want to tell you about Klarna, because it's the cleanest example of how these stats get inflated and then quietly walked back.

In 2024, the buy-now-pay-later company Klarna bragged that its AI assistant did the work of 700 customer service agents. The press ate it up. "AI replaces 700 workers" is a great headline. Except it wasn't quite true. The "700 agents" was mostly hiring Klarna avoided during a freeze, not 700 actual humans escorted out the door. There's a real difference between "we didn't hire people we might have" and "we fired 700 people," and the headline collapsed it into the scarier version.

Then 2025 happened. Klarna's CEO, Sebastian Siemiatkowski, reversed course and started rehiring human agents. His own words: the company focused too much on efficiency and cost, and the result was lower quality (Entrepreneur / Bloomberg, 2025). Read that as a CEO who tried the all-AI dream, hit the wall every honest operator hits, and admitted the customers could tell. This is the pattern under a huge share of "AI replaced our staff" announcements: a splashy claim, an inflated number, and a quieter correction nobody covers. When you see a company brag about replacing humans with AI, write down the date. Check back in twelve months. More often than you'd think, the story has changed.

Every report that forecasts destruction also forecasts creation, and the creation side points at one role over and over. The role that builds the thing.

The WEF lists AI and Machine Learning Specialists, Big Data Specialists, software developers, and security specialists among the fastest-growing roles to 2030 (WEF, 2025). It found 86% of surveyed employers expect AI and information processing to transform their business by 2030, the top-ranked technology driver (WEF, 2025), and that 85% of employers plan to prioritize upskilling, with 63% citing skills gaps as the single biggest barrier to transformation (WEF, 2025). They expect 39% of workers' core skills to change or become outdated by 2030, down from 44% in 2023 (WEF, 2025), and structural labor-market churn equal to 22% of today's jobs by 2030 (WEF, 2025). Churn, not collapse. The jobs move; they don't vanish.

The single clearest number is from BLS. It projects software developer employment growing 17.9% from 2023 to 2033, much faster than the 4% average for all occupations (BLS, 2024). And BLS explicitly notes that AI may support demand for computer occupations, because developers are needed to build and maintain AI systems (BLS, 2025). Are there real AI-attributed cuts? Yes. Challenger, Gray & Christmas tracked 54,836 layoffs explicitly attributed to AI in 2025, out of 1.2 million total announced job cuts (Challenger, 2026). Real, but a sliver, and they cluster in routine functions, not in the AI-building roles. The durable bet across every report is the same: be the person who builds the AI, not the person whose tasks it absorbs. That's the entire thesis of the AI engineer career path, and it's the one occupation the data agrees on.

One more table, because the gap between public fear and expert assessment is its own data point. Pew Research has been tracking this, and the spread is telling.

Look at the spread. 64% of the public expects fewer jobs, but only 39% of AI experts do (Pew Research, 2025). And only 32% of workers think AI will cost them opportunities personally, even though 52% are worried in the abstract (Pew Research, 2025). That's the fear gap. People are scared of the headline, not the data. The closer someone is to actually understanding how this technology works, the less likely they are to predict a jobs apocalypse. That alone should tell you which group to listen to.

I'll leave you with the framework I use on every one of these stats, so the next scary headline bounces off you instead of ruining your week. Three questions. That's it.

First: is this measuring exposure or replacement? Almost always it's exposure, tasks AI could touch, dressed up to sound like headcount. "300 million jobs exposed" and "300 million jobs lost" are not the same sentence, no matter how the article wants you to read it. Second: is this a forecast model or observed data? Forecasts vary wildly because they bake in assumptions about adoption speed, and those assumptions are where the answer gets decided. Observed data, like Yale's payroll analysis showing no disruption (Yale Budget Lab, 2025), is the adversarial check. When a forecast and the real data disagree, the real data wins. Third, and this one's blunt: who's funding the number, and what do they sell? A consulting firm sells transformation projects. An AI company sells the model. A newsletter sells your fear. None of them are neutral.

Run those three questions and the panic deflates fast. But the framework also points you somewhere, and it's the same place every report points. Across the scary forecasts and the quiet data alike, the one role that grows is the person who builds and runs the AI. BLS says developers grow 17.9% (BLS, 2024). WEF says AI and ML specialists are the fastest-growing roles (WEF, 2025). Stanford says the people getting hit are the ones who only know what the LLM already knows (Stanford, 2025). Add those up and the move is obvious. Don't be the task. Be the operator. That's not optimism, it's just what the numbers say.

Every statistic on this page traces to a named, primary source. Where a primary site blocks automated access, the figure was confirmed against the publisher's own quotation and cross-checked against multiple independent reports. A few widely-repeated stats that couldn't be verified to a primary source were left off on purpose. Here are the sources, with the specific findings used.

Goldman Sachs Research (Briggs & Kodnani, 2023), "The Potentially Large Effects of Artificial Intelligence on Economic Growth" and "How Will AI Affect the US Labor Market": origin of the 300 million exposed jobs figure, the two-thirds-of-occupations-exposed finding, the ~25% of work hours estimate, and the 7% / $7 trillion GDP projection.

Eloundou, Manning, Mishkin & Rock, "GPTs are GPTs" (arXiv, 2023), later peer-reviewed in Science: the 80% / 10%-of-tasks and 19% / 50%-of-tasks figures, the 47-56% of tasks accelerated with software, and the most-exposed (writing, programming, math, data analysis) versus least-exposed (physical, outdoor) split.

IMF Staff Discussion Note (2024), "Gen-AI: Artificial Intelligence and the Future of Work": the ~40% global exposure figure, the 60% / 40% / 26% advanced / emerging / low-income breakdown, and the roughly-half-complemented finding.

World Economic Forum, Future of Jobs Report 2025 (1,000+ employers, 14M+ workers, 55 economies): 170 million created / 92 million displaced / +78 million net, 22% structural churn, 86% expecting AI transformation, 39% of skills changing (down from 44%), 85% prioritizing upskilling, 63% citing skills gaps, and AI/ML specialists among the fastest-growing roles.

McKinsey Global Institute, "Generative AI and the Future of Work in America" (2023) and "Jobs Lost, Jobs Gained" (2017): up to 30% of US work hours automatable by 2030, 12 million additional occupational transitions, lower-wage workers up to 14x more likely to switch, office support / customer service / food service as biggest decliners, and under 5% of jobs fully automatable.

The Budget Lab at Yale (2025), "Evaluating the Impact of AI on the Labor Market": no discernible economy-wide disruption 33 months after ChatGPT, occupational mix shifting only ~1 point faster than early internet adoption.

Stanford Digital Economy Lab (Brynjolfsson, Chandar & Chen, 2025), "Canaries in the Coal Mine?": 13% relative employment decline for workers aged 22-25 in the most AI-exposed jobs, roughly 20% for young software developers specifically, stable-or-growing employment for older workers in the same roles, and declines concentrated where AI automates rather than augments.

MIT & Oak Ridge National Laboratory, The Iceberg Index (2025): current AI could technically perform tasks equal to 11.7% of the US workforce, about $1.2 trillion in wages, with an explicit note that it measures exposure, not adoption or displacement.

OECD Employment Outlook 2023: occupations at highest automation risk account for about 27% of employment across OECD countries, with little evidence so far of AI causing job losses.

Anthropic Economic Index (2025): AI use leaning 57% augmentation over 43% automation across millions of conversations, with the note that the split shifts across reports.

US Bureau of Labor Statistics (2024-2025), Occupational Outlook Handbook and The Economics Daily: software developer employment growing 17.9% (2023-2033) versus 4% average, bank teller employment declining ~15% (roughly 51,400 jobs), and the note that AI may support demand for computer occupations because developers build and maintain AI systems.

Pew Research Center (2025): 52% of US workers worried about AI's workplace impact, 32% expecting fewer personal opportunities, 64% of the public versus 39% of AI experts expecting fewer jobs over 20 years.

Challenger, Gray & Christmas (2026): 54,836 layoffs explicitly attributed to AI in 2025, out of 1.2 million total announced cuts.

Entrepreneur / Bloomberg reporting (2025): Klarna's reversal of AI-driven workforce cuts under CEO Sebastian Siemiatkowski, who admitted the company focused too much on efficiency and cost at the expense of quality.