The forecasters changed their minds. The data didn’t get the memo.

A field note on what a year of AI-and-jobs predictions actually told us — and why the predictions turned out to be the least reliable thing in the room.

For about a year, the people building AI told us it was coming for our desks. Not in the hedged, someday-maybe register that futurists use to stay safe, but with numbers and deadlines attached. Half of entry-level white-collar jobs. A few years. A "white-collar bloodbath."¹ These were not pundits reaching for clicks — these were the chief executives of the companies shipping the models, the rare insiders willing to say out loud what their peers only muttered.

And then, sometime around the spring of 2026, several of them quietly took it back.²

That reversal is the actual story here, and it's a stranger one than either the original panic or the walk-back, taken alone, would suggest. Because while the forecasts swung from doom toward reassurance, the ground-level data did something neither camp predicted: it didn't line up behind either of them. I want to lay the whole thing out — what was said, what was unsaid, and what was happening underneath — because the gap between those three is where the real lesson lives, and it isn't the lesson anyone was selling.

Layer one: the forecasts, and who made them

Start with the record, because it's easy to misremember a year later.

In May 2025, Anthropic's Dario Amodei said AI could eliminate up to half of entry-level white-collar jobs, possibly pushing unemployment to 10–20% within one to five years. He used the word "bloodbath" and accused the industry of sugar-coating.¹ In June 2025, OpenAI's Sam Altman warned that a lot of jobs would go away and that entry-level roles were squarely at risk.² In July 2025, Microsoft Research published a paper scoring occupations by how much they overlap with what people actually use AI for — drawn from roughly 200,000 real Bing Copilot conversations — with information work (writing, gathering information, communication) scoring highest.³ By February 2026, Microsoft's AI chief Mustafa Suleyman was putting a clock on it, telling the Financial Times that most white-collar tasks — naming law, accounting, marketing, project management — would be "fully automated by an AI within the next 12 to 18 months."⁴

Bill Gates played a different note, the optimist's hedge — only three fields, he said repeatedly through 2025 and into 2026, stay reliably human for now: coding, energy, biology.⁵ The phrase "for now" did a lot of quiet work in that sentence, and we'll come back to it.

Put the predictions on one axis and a pattern is already visible: confident, specific, dated, and made by people who are, at minimum, not disinterested parties to the question of whether AI is powerful. Hold that thought.

Layer two: the walk-back

Here's where it gets uncomfortable for anyone who took the forecasts at face value.

By May 2026, Amodei had changed his frame. Sitting onstage next to JPMorgan's Jamie Dimon, he reached for the Jevons Paradox — the nineteenth-century observation that making something more efficient tends to increase, not decrease, the demand for it. Automate 90% of a job, he argued, and the remaining 10% expands to fill the day at roughly ten times the productivity.⁶ Jobs multiplying, not vanishing.

But the reframe was not the clean 180-degree reversal it's often described as — and that detail matters. In the same exchange Amodei also invoked Amdahl's Law (a system is only as fast as its slowest component) and explicitly cautioned that AI is moving faster than past technologies, which could produce "weird behaviors and this big disruption." His own caveat conceded the weak point: the Jevons rebalancing depends on time, and he wasn't sure there'd be enough of it for the workers caught in the transition.⁶ So the shift was real, but it was a shift in emphasis under the same uncertainty — not the discovery of a fact that overturned the earlier one.

Altman was blunter. He called himself "pretty wrong" about AI's near-term economic impact and said he was "delighted to be wrong" — that he'd expected more entry-level white-collar elimination by now than had actually happened.²Goldman's David Solomon, who never bought the apocalypse, kept pointing to a century of American history — electrification, then computing, now AI — each wave creating new work after it destroyed the old.²

And the most telling reversal came from Microsoft Research itself. The paper that launched a thousand scary headlines was quietly re-titled — from Working with AI: Measuring the Occupational Implications of Generative AI to Working with AI: Measuring the Applicability of Generative AI to Occupations — and the authors published a follow-up that amounted to a correction of how their work had been read.⁷ Their study, they wrote, "does not draw any conclusions about jobs being eliminated"; in the paper they had explicitly cautioned against that reading. The line they drew was simple and devastating to the headline framing: "A job is far more than the collection of tasks that make it up."⁷ Their score measured what a chatbot is used for, not what a human will be fired for. The list everyone read as a death-row roster was never that.

I'll resist the urge to narrate anyone's private motives here — I can't see inside these reversals, and pretending to would be its own kind of dishonesty. The structural point stands without mind-reading: each of these forecasts was made under deep uncertainty by someone with a material stake in how the technology is perceived, and within a year the same people told a calmer story — to a market grown nervous about an AI bubble, and as both OpenAI and Anthropic moved toward enormous IPOs for which an employment-apocalypse risk factor is poison.⁸ You don't need to assign a motive to notice that a statement which softens 180 degrees in twelve months — without the slow structural data having moved to justify it — was never functioning as a measurement in the first place.

Except the measurements were coming in the whole time — and they didn't move the way the talk did.

Layer three: what the data actually did

This is the part that breaks the tidy narrative in both directions — but not because the data is contradictory. It isn't. It's that different datasets measure different things, and people kept stacking them as if they were one number.

On one layer, the layoff trackers. In 2025, Challenger, Gray & Christmas attributed roughly 55,000 US job cuts directly to AI, amid the highest total layoffs since 2020.⁹ By May 2026, tech layoffs had passed 115,000 for the year, already nearing the ~124,000 logged in all of 2025, with Meta, Amazon, and Snap among the companies naming AI as a driver.² On that layer, the original doom forecast looks more right, not less — exactly as its authors were backing away from it. (A caveat the trackers themselves invite: "AI" is a more palatable thing to cite in a press release than "we over-hired in 2022," and analysts at Forrester expect a sizable share of these AI-attributed cuts to be quietly reversed.)

On a different layer, the hiring data. Big-tech new-grad hiring fell something like 30–50% from pre-pandemic levels; payroll data showed employment for 22–25-year-olds in highly AI-exposed jobs falling about 13% relative to less-exposed roles, and the entry-level software role that defined Silicon Valley for two decades got quietly deprecated.¹⁰ The on-ramp is narrowing even where the occupation itself is intact. Some of that is AI; some of it is the end of cheap money and the hangover from 2022 over-hiring. The honest reading is that the entry level is being squeezed by several forces at once, and AI is one of them — not that AI is the sole cause.

And on a third layer, the structural measures. The Yale Budget Lab, tracking the composition of employment through March 2026, found no significant change in the occupational mix or in unemployment duration for high-AI-exposure jobs since ChatGPT launched in late 2022 — and noted that the shifts that are visible were already underway in 2021, before generative AI was widely available.¹¹ Separately, MIT's "Iceberg Index" (built with Oak Ridge National Laboratory and released in November 2025) estimated AI could already perform tasks overlapping about 11.7% of the US labor market — roughly $1.2 trillion in wages — while noting that only about 2.2% of that is visible in actual adoption so far.¹²Capability that exists but hasn't, on the structural layer, translated into a measurable reshaping of who works where.

Read across the layers and the apparent paradox dissolves into a more precise statement: declared layoffs blamed on AI are climbing, the entry-level on-ramp is closing, exposure is high, and the macro structure of employment looks almost untouched — all at once, because each of those is a different instrument pointed at a different part of the system. The doom-callers were reading the layoff layer. The optimists were reading the structural layer. Both were right about their own layer and wrong to treat it as the whole.

A note on what each number can and can't see

It's worth being explicit about the instruments, because most of the public argument came from quietly swapping one for another.

Declared layoffs — the Challenger figures, the company statements — measure attribution, not causation: cuts where AI was named as a reason. That makes them fast and concrete, but they fold in restructuring, over-hiring hangover, and the simple fact that "AI" is a more palatable cause to cite than a hiring mistake. They tell you what employers say they're doing.

Employment structure — the Yale Budget Lab work — measures the composition of the labor market: which occupations exist, in what proportion, with what unemployment duration. It's the slowest instrument and the hardest to move, which is exactly why it's the best check on whether something fundamental has shifted versus merely been announced. It tells you what has actually been rebuilt.

Exposure / capability — the MIT 11.7% — measures overlap: how much of the work, in principle and at competitive cost, AI could perform. It's a ceiling, not an event. High exposure means the pressure exists; it says nothing about whether, when, or how that pressure becomes a headcount decision. (MIT's own "iceberg" framing — a small visible tip over a large submerged mass — is the same layering point this essay is making, in different clothes.)

Hiring / on-ramp — the new-grad data — measures flow at the entry point: who is being let in. This is the leading indicator the other three miss, because a firm can freeze junior intake long before any of it shows up in layoff counts or in the occupational mix. It tells you about the future shape of the workforce before that shape exists — though it's also the layer most easily moved by interest rates and over-hiring corrections, so it must be read with care.

Stack those four and you can see why a year of argument went in circles. Each side held up the instrument that flattered its forecast. Almost nobody said which layer they were standing on.

What the gap is actually telling us

Three things fall out of laying it side by side like this, and they build on each other.

The forecasts were the least reliable instrument in the room. Not because the forecasters were foolish — Amodei in particular took professional heat for saying an unpopular thing he appeared to believe. But a prediction made under deep uncertainty, by someone whose standing is tied to the technology seeming inevitable, is a weak instrument regardless of the person's sincerity. When such a statement softens within a year without the structural data having moved to justify it, the honest conclusion isn't "now he's right" or "he was right before." It's that these statements were never measurements. The arrow on a forecast tells you about the conditions under which it was made at least as much as it tells you about your job.

"Task" and "job" are different nouns, and almost every scary headline swapped one for the other. This is the quiet hinge the whole confusion turns on. Microsoft's own researchers drew the line and then watched the world erase it: a high overlap between AI and the tasks inside a role is not the same as that role disappearing.³⁷ Amodei's Jevons reframe is really just this same distinction in optimistic clothing — automate 90% of the tasks and the person doesn't vanish, they get redeployed onto the 10% that didn't automate, which expands to fill the day.⁶ Whether that redeployment is a promotion or a slow squeeze depends entirely on the role, and the aggregate number can't tell you which. Gates's "for now" was pointing at the same seam from the other side: coding, energy, and biology aren't safe because AI can't touch their tasks — it can — but because those roles are mostly judgment wrapped around the tasks, and judgment is the residue automation leaves behind.⁵

The entry level is where task and job collapse into the same thing — and that's the real signal under the noise.Here's why the layoff layer and the structural layer can both be telling the truth. A junior role is, often, a bundle of exactly the well-specified, checkable tasks that AI does first: the research memo, the first-draft contract, the boilerplate code, the data pull. Senior roles are judgment with tasks attached; junior roles are tasks with judgment still to be learned. So AI doesn't need to eliminate the occupation of "analyst" to thin out the job of "first-year analyst" — it can absorb the bundle that the first year used to be. The structure looks stable because the occupations still exist. The entry-level hiring data sags because the on-ramp to those occupations was made of the most automatable work in the building. Two layers, one mechanism — though, again, AI is one driver of that squeeze and not the whole of it.

And that points at the one consequence nobody's reversal addresses. If you dissolve the bottom rung because it's cheap to automate this quarter, you also weaken the path by which juniors become the seniors whose judgment was just declared irreplaceable.¹³ The forecasts argued about how many jobs. The number that will actually matter is how many people are still allowed to start.

The note I'd pin to all of it

I went into this expecting to find out who was right — the bloodbath camp or the it's-fine camp. What I found instead is that the question was badly posed. The forecasts were the weakest instrument available. The aggregate structural data was measuring something too slow to have moved yet. The only place something was unambiguously happening was the seam between task and job, at the entry level, where the cheap-to-automate work lives — and that's precisely the place the headline numbers are worst at seeing.

So the next time someone with a model to sell puts a percentage and a deadline on your profession, the useful move isn't to believe them or to dismiss them. It's to ask which noun they're using, which layer of measurement they're standing on, and what the slow structural data says underneath the quote. A year of watching the experts reverse themselves taught me to trust the forecasts less and the seams more. The forecasts changed. The data didn't get the memo. The seam — quiet, unglamorous, badly measured — was the only one telling the truth the whole time.

A note on method: quotes are kept short and attributed; figures are as reported by the outlets below, several of which are secondary write-ups of primary interviews (Axios, Financial Times, CNBC, Fortune) and the underlying arXiv paper. Dates and titles were re-checked against primary sources where possible; two figures the original draft mis-dated (Suleyman's "18 months," from February 2026 not May; and the framing of Amodei's Jevons remarks as a clean reversal) have been corrected here.