Part I · Chapter 6

Reading AI-impact claims specifically

No topic in this book attracts more noise than this one. So it gets its own chapter.

One distinction organises everything. Learn it and most of the noise disappears.

Measured, or projected?

Measured displacement means somebody counted real jobs held by real people, before and after.

Forward projection means somebody estimated what might happen later.

Both appear as numbers in headlines. Only one is evidence.

Here is the test. Ask whether the number describes the past or the future. If it describes the future, it is a guess. The guess may be careful and well informed. It is still a guess, and you should never weight it like a measurement.

What has actually been measured

Three pieces of real evidence matter as of 2026. They agree with each other, which is why this book trusts them.

The Stanford payroll study

Researchers at the Stanford Digital Economy Lab studied real payroll records. The paper is by Brynjolfsson, Chandar and Chen, updated in August 2026.

The scale is what makes it useful. It covers between 3.5 and 5 million workers a month, using payroll data from ADP, a company that processes pay for a large share of American employers. These are not survey answers. They are records of who was actually on a payroll each month.

The finding: workers aged 22 to 25 in AI-exposed jobs saw a 19% relative employment decline. Experienced workers show no comparable gap.

The word relative matters. It does not mean 19% of young workers lost jobs. It means the gap against similar workers in less exposed jobs. Employment in the exposed group fell about 11%. In the unexposed group it grew about 10%.

One finding here matters more to you than the headline does. The gap comes from hiring, not from firing. Separations did not rise for young workers in exposed jobs. They fell. What narrowed was the way in.

The comparison group is what makes this study strong. It measures exposed jobs against unexposed ones, so a bad year for everybody does not read as an AI effect.

Read the controls carefully, though. The August 2025 version said the decline survived controls for company-level shocks. This version says only that the result is directionally consistent under those controls, and that precision varies. The headline number grew and the control behind it weakened in the same revision.

The authors are careful in a way the coverage is not. Their own summary says the results may be influenced by factors other than AI. They call the findings consistent with the hypothesis that AI has begun to affect entry-level hiring. That is the strongest honest statement available in 2026. It is weaker than almost everything you will read about it.

Four numbers, and why they are not rivals

You will meet several figures from this one study. People treat them as better and worse versions of one thing. They are not. Some measure different groups of people. Others are the same measurement in an older version of the paper.

About 20% is the fall for software developers aged 22 to 25, measured against their own peak in late 2022. It is an absolute fall in one occupation. It was Fact 1 of the November 2025 version, not a press exaggeration. It is also the figure closest to what this book is about. The August 2026 version restated its six facts, and this is no longer one of them. Quote it as a 2025 reading.

19% is the relative fall for young workers across all the most exposed occupations, with data to June 2026. Wider group. Different comparison. Different starting point.

16% is that same relative figure in the November 2025 version. 13% is the August 2025 version, which had data only to July 2025. Each revision added months and the estimate rose.

So “20% is the wrong version of 19%” is itself wrong. Both are in the paper. They answer different questions. Question three later in this chapter is compared against what? This is what happens when you skip it.

If you quote one of these, say which one, say against what, and say which version. A paper that is revised every year has no single number.

The Harvard study on seniority

Separate work by Hosseini and Lichtinger looked at companies adopting AI tools, and compared junior against senior employment inside them.

Same shape of result. The junior end contracted. The senior end did not.

Indeed job postings

Indeed Hiring Lab tracks job adverts, which move faster than employment data.

In October 2025 software development postings sat 36.4% below the level of February 2020. By July 2026 they had recovered about 15% from that low point, but remained roughly 27.5% below the pre-pandemic baseline.

The important part is not the recovery. It is the shape of it. 71% of the increase came from senior roles.

What has only been projected

Now the other category.

You will see claims that AI will remove half of all entry-level office jobs within a few years. You will see consultancy forecasts of tens of millions of roles gone by 2030. You will see vendor roadmaps promising that their tool replaces a whole team. None of these are measurements. They are statements about a future nobody has observed.

Chapter 4 showed how badly this can go. Quantum computing was projected to need two million workers by 2025. When somebody counted, the real figure was about 16,500.

Projections are not worthless. A careful projection from a good model is worth reading. But treat it as an opinion with numbers attached, and check the four questions from Chapter 4. Especially question three: who profits if you believe it?

The problem that makes this genuinely hard

Here is the honest difficulty, and most articles skip it.

Three things happened to technology hiring at the same time.

1. Companies over-hired massively in 2021 and 2022, then corrected.
2. Interest rates rose sharply, which made investors demand profit instead of growth, which cut hiring budgets.
3. AI tools arrived and companies adopted them.

All three push hiring down. All three happened together. So when hiring falls, you cannot simply credit AI.

This is called a confounding problem. Several possible causes move together, so you cannot separate their effects by looking at the total.

There is a tempting way to settle this, and it is wrong. It goes: young workers fell, experienced workers did not, so the cause must be something new.

Young hiring is more cyclical than experienced hiring. That is one of the steadier findings in labour economics. In any downturn, in any industry, the first thing an employer stops doing is hiring beginners. Existing staff are kept because replacing them costs money. Experienced hires are preferred because they produce sooner.

So an age-shaped hole is the ordinary shape of a hiring freeze. On its own it is not evidence of anything new.

What does carry weight is the comparison between occupations. Young workers in exposed jobs fell against young workers in less exposed jobs. Same months, same firms. That comparison holds the business cycle still while you look.

The age gap is the pattern. The exposure gap is the evidence. This book rests on the second one.

The argument is still live

Two papers came out in early 2026. They pull in opposite directions, so this book prints both.

February 2026, the same Stanford team. They asked whether interest rates explain the fall. If rates were the cause, the jobs hit hardest should be the ones that react most to rates.

They are not. AI-exposed jobs react less to rates than jobs like building work. That is the opposite of what the rates story needs. It makes the AI reading stronger.

The same paper also weakens it. With their strongest controls, the fall only shows up clearly from 2024. Anything before that may be something else.

January 2026, the Economic Innovation Group. This is the strongest argument against the whole idea.

They counted job adverts rather than payrolls. Adverts in the most exposed jobs peaked in March 2022. Then they fell hard, for the rest of that year.

ChatGPT came out in November 2022. So the fall started more than six months before the tool existed. On their reading, jobs went later because hiring had stopped earlier.

Where this book lands. Both are partly right.

The exposure gap holds. Young workers in exposed jobs really did fall against young workers in other jobs.

The clean story does not hold. Something was already slowing before the tools arrived. The tools then landed on the group with the least cover.

That is a smaller claim than the headlines make. It is the one the evidence supports. Chapter 20 says what would change it.

Four questions for any AI-impact headline

Use these on every claim you meet, including the ones in Chapter 10 of this book.

1. What exactly was measured? Jobs held? Job adverts? Company statements about future plans? These are three very different things, and adverts are the weakest.

2. Over what period? A single quarter tells you almost nothing. Hiring is seasonal and noisy.

3. Compared against what? A fall means nothing without a comparison group. The Stanford study works because it compares exposed jobs against unexposed ones in the same period.

4. Who paid for it? A study funded by a company selling AI tools, and a study funded by a body representing workers, will find different things. Neither is automatically wrong. Both need the question asked.

What this means for you

The measured evidence supports one specific conclusion, and not a broader one.

Entry-level work in exposed occupations has become harder to obtain. Experienced work has not. The recovery that has happened is flowing mainly to senior roles.

That is a serious problem, and Chapter 10 explains the mechanism behind it. It is not the same as “computing careers are finished”, and anyone telling you that is going beyond the evidence.

It is also not the same as “nothing has changed”, and anyone telling you that is ignoring the evidence.

The truth is narrower, stranger and more useful than either. Part II describes it.