AI Is Being Taught to Do Your Job. Here's What the Data Actually Says, And What It Doesn't

The office is quiet. Fluorescent hum. A 23-year-old named Marcus stares at a screen that drafts emails faster than he can read them. He got this job eight months ago. Entry-level content coordinator. The kind of job that used to be a ladder. Now it feels like a trapdoor.

He isn't wrong to be nervous. Workers age 22 to 25 in the most AI-exposed occupations have experienced a 13 percent decline in employment since 2022. That's not a prediction. That's payroll data. Twenty-five million workers. ADP. Stanford researchers. Real numbers.

But here's where it gets strange. The same data that shows Marcus's problem also shows something else. Unemployment sits at 4.1 percent. The broader labor market hasn't collapsed. And economists, the ones who actually look at this stuff, can't agree on what the hell is happening.

---

The Numbers Nobody Can Agree On

Start with S&P Global. They survey businesses. Thousands of them. Their latest reading shows a net employment impact of -5 percentage points over the past year. That means more firms cut workers because of AI than added them. The forecast for the coming year is also negative. Marginal. But negative.

Now flip to the Atlanta Fed and the National Bureau of Economic Research. They surveyed nearly 750 corporate executives. Their finding: "little evidence of near-term aggregate employment declines due to AI". Smaller firms expect modest gains. Larger firms anticipate reductions. The aggregate picture stays murky.

The CEPR offers the most honest framing. AI substitutes for humans on some tasks and complements humans on others. Both things happen at once. Since 2021, highly substitutable work has grown 3.3 percentage points a year more slowly. Complemented work has grown 3.3 points faster. The net effect depends entirely on which tasks you're measuring, and most studies measure at the occupation level, which flattens the whole thing into a useless average.

The Economist put it bluntly: "the jobs apocalypse is postponed." They counted roughly a million US jobs created against some 200,000 lost. That's a net positive. But it hides the people who got hurt.

Nobody has the clean answer. The machines are here. The jobs are shifting. The story depends on where you stand.

---

What AI Actually Does to a Job

Here's the thing people miss. AI doesn't act on jobs. It acts on tasks. A job is a bundle of tasks. Some of those tasks are replaceable. Some aren't.

Medical secretaries and secondary school teachers score nearly identically on standard AI exposure indexes. On a scale spanning several points, they're 0.027 apart. Practically the same number. But AI can already handle two-thirds of a medical secretary's tasks end to end, scheduling, transcription, drafting summaries. It cannot stand in front of a classroom. Roughly a fifth of a teacher's tasks are substitutable. The teacher's work holds five times the complement share.

One number for two jobs. Two different fates. The index can't express that.

This is the mechanism. The machine eats tasks. It doesn't eat jobs. It eats the parts of a job that can be reduced to a prompt and a response. What remains, the parts that require presence, judgment, physical coordination, emotional reading, the messy human stuff, becomes more valuable. Or it becomes a smaller piece of a smaller pie. Depends on the job.

The Atlanta Fed found a compositional reallocation happening within firms. Routine clerical roles declining. Skilled technical roles gaining. The work isn't disappearing. It's rearranging. The people in the rearranged seats aren't always the same people who were sitting there before.

---

Who Gets Hit First, And Who Doesn't

The young get hit first. That's the clearest signal in the data. The 13 percent decline for workers 22 to 25 isn't about layoffs. It's about hiring. Companies aren't cutting young workers. They're just not bringing them in. The entry-level job, the one that used to teach you how to work, is getting automated away before you can learn from it.

Then there's the gender dimension. A Brookings analysis found that of the 6.1 million workers in highly AI-exposed jobs with low adaptive capacity, limited savings, advanced age, narrow skills, no local alternatives, 86 percent are women. Primarily clerical and administrative roles.

That's not abstract. That's a specific population of people who were told to learn Excel and became the backbone of office operations and now face a transition they may not have the resources to manage.

Meanwhile, the skills gap widens. The World Economic Forum estimates 59 percent of workers globally will need retraining by 2030. A UK government study found that 57 percent of businesses report a technical skills gap. The most significant gap? Understanding AI concepts and algorithms. That figure rose from 55 percent to 60 percent in five years.

Companies need people who understand the machine. Workers need to become those people. The bridge between those two facts is called "retraining," and it's not clear anyone is building it fast enough.

---

What Workers Actually Believe

Forget the executive surveys for a second. Ask the workers.

The Boston Fed surveyed household heads in 2024 and again in 2025. The share of workers concerned about losing their job to AI nearly doubled, from 5 percent to just over 10 percent. Among doctorate holders, zero percent expressed concern in 2024. By 2025, 14 percent did.

But here's the interesting wrinkle. Workers who perceived large productivity gains from AI remained optimistic about their job security. Fourteen percent of them said they were more likely to ask for a raise since AI entered their industry. That group represents only 6 percent of all workers.

The pattern: people who feel like AI is helping them do their jobs better feel safer. People who feel like AI is replacing their tasks without making them more productive feel terrified. The technology itself isn't neutral. Neither is the experience of it.

---

The Skills Nobody's Teaching

There's a gap between what employers think they need and what actually matters.

The UK research identified something called the "iceberg effect." Workers lack confidence in AI skills because the visible part, knowing how to prompt a model, looks manageable. The submerged part, judging the accuracy and reliability of outputs, understanding what the model is doing under the hood, knowing when to trust it and when to override it, is enormous and largely invisible.

Technical skills dominate employee learning wishlists. But soft skills are slipping. Critical thinking. Creative direction. The ability to call bullshit on what the AI produces. These are the skills that determine whether a worker uses AI or gets used by it.

A recent report noted that workers with advanced AI expertise represent only around 1 percent of the workforce. Skills shortages remain a key constraint on adoption. The irony: companies want AI specialists but can't find them. Workers want to become AI specialists but don't know where to start. The gap isn't just technical. It's structural.

---

What Comes Next

The World Economic Forum sets out four scenarios for 2030. More than half of business executives expect AI to displace existing jobs. Twenty-four percent expect it to create new ones. Forty-five percent think it will improve profit margins.

Those numbers don't reconcile neatly. They're not supposed to. The future of work isn't a single outcome. It's a distribution. Some sectors hollow out. Some get transformed. Some create entirely new categories of work that nobody has named yet.

Goldman Sachs notes that where AI augments workers, it can lower the cost of output and increase demand. That can generate a net increase in employment. Software development is already seeing this. Content creation is not.

The honest answer to "will AI replace us" is: it will replace some of us. It will transform the work of many. It will create new work for others. The net effect on employment is modestly negative right now and may shift. The net effect on what work feels like is already enormous.

A livelihood is not a role. The World Economic Forum put it well: it's "a sustained capacity to generate value, dignity and security over time." The machine doesn't care about dignity. It doesn't care about security. It processes tokens and generates outputs and moves on.

The humans have to care about that. The humans have to build the bridges. The humans have to decide what work is for.

Marcus will probably be okay. He'll learn to manage the tools. He'll find a role that the machine can't fill. Or he'll get squeezed and move into something else and remember the fluorescent hum of that office as the place where he learned the machine was coming.

The machine is here. The question is who's steering.