Part I of Computing Careers 2026 — free sample
This book is a method, and 46 worked examples of that method.
The method measures any computing career across ten dimensions. The examples show the method working on real careers, with real numbers.
The method is the part that lasts. Numbers age. A way of thinking does not.
It is not a complete list of computing careers. No such list exists.
Here is the honest reason. Reliable career data does not exist for most of the world. It is good in the United States. It is fair in Western Europe. It thins out quickly after that, and for many countries there is nothing at all.
Any book that claims to cover every career in every country has filled the gaps with salary websites. Chapter 4 shows you why you cannot trust those numbers.
So this book does something different. Where the data is missing, it says so. Appendix B lists every gap by name.
This is version 1.0.0, dated August 2026.
A career book starts going out of date the day it is finished. That is why every figure here carries the date it was measured, and why Appendix F records what changed between editions.
Outside reviewers read this edition before release, and they found real errors. Appendix F ends with the full list of what they changed. It sits in one place rather than scattered through the chapters. Read it before you trust the rest. It is the fairest measure of how careful this book really is.
Chapter 20 lists the specific things that would prove this book wrong. An honest book tells you when to stop believing it.
Every correction to this edition, for as long as corrections are made.
This book is sold on the claim that every figure is checkable. So being told that a figure was wrong cannot be a separate product. If something here turns out to be wrong, the fix is part of what you already paid for.
Later editions and new material are a different thing, and they are priced separately.
There is no schedule attached to this, deliberately. A correction ships when it is found, not on a calendar.
This book carries a name that is not a person’s.
The name is not decoration. Chapter 5 sorts every source in this book into five grades. Grade 1 is the best: a body with the power to compel an answer, and nothing to sell you. That is the standard this book tries to reach and says plainly when it cannot.
That is a fair thing to be suspicious of. Chapter 4 tells you to ask who benefits if you believe a claim. It would be dishonest to ask that of everyone else and then refuse the question here.
So here is what you can check instead of a byline.
Nobody paid for this book to be written. It was not commissioned, sponsored or funded. No advance, no grant, no client.
No organisation named in this book has any relationship to it. Not the training providers, not the universities, not the employers, not the platforms, not the certification bodies. It criticises several of them by name, contacted none of them for approval, and gave none of them a say.
There are no affiliate links, referral codes or paid placements. Nothing here earns money if you click it, enrol in it or buy it. The book has one revenue source, which is people buying the book.
Every figure came from a published source, read directly. Appendix C lists all 200 of them with the grade each one earns under Chapter 5. Where a source could not be opened, or where a figure was taken from somebody else’s reporting of it, the record says so. Appendix B lists what is missing.
The same material comes with this book as data files, so you can sort and check it yourself rather than take the appendices on trust. The data folder says what is in each one.
Now the reason for the anonymity, plainly.
This book tells people not to buy things. It names crowded doors that expensive courses are still selling, and it says which certificates no longer work. That is easier to write, and harder to soften later, without a career attached to it.
The trade is real and it costs you something. You cannot look up the author’s record. You cannot weigh their experience against their claims. A named author with a reputation to lose is, genuinely, a form of accountability, and this book does not offer it.
There is no name on this book, but there is still a person behind it, and in a few places that person says I.
Those are the places where a judgement was theirs rather than a source’s. One is the experiment on a salary site in Chapter 5. Another is Chapter 20 admitting that nobody has ever tested this book’s main claim. In sentences like those you are trusting a person rather than a record, and the shift to I is how you can tell. Everywhere else the book speaks as a book, because everywhere else the claim rests on a source you can open yourself.
What it offers instead is the thing a name cannot give you. Every claim carries a source, a grade and a date. Chapter 20 lists what would prove the book wrong. Appendix B lists what it does not know before you find out yourself.
Check the book, not the author. That is what Part I is for, and it works on this book too.
Appendix B lists what this book does not know.
If you work in one of those fields, or one of those countries, you know things this book does not. Send them. Contributions with a source and a date go into the next version, with credit. Chapter 20 explains where to send them and what makes a correction usable.
The gap list is not an apology. It is where the next edition comes from. It is also the part of the method a reader can audit without taking anybody’s word for anything, which is the point.
Do not read this book from front to back.
It is in two halves. Book A is the method, and you read it in order. Book B is 46 careers measured the same way, and you look things up in it. Chapter 2 is six short routes, one for each kind of reader who opens this book. Find yours there and follow it. That takes two minutes, and it tells you which forty pages are yours.
The method. How to tell a real number from a marketing number, how to check any claim yourself, and how to check the people selling you a qualification. This part is free.
Priya and Marcus both earn $95,000 a year. They are the same age and have the same years of experience. Their lives are not alike at all.
Priya works at a consumer app company. She is on-call one week in three, which means she must stay reachable and fix problems at night. Forty other companies in her city hire for exactly the same skills, so she competes with thousands of people every time she looks for work. Her strongest framework came out three years ago, and in another two it will be a weak line on her CV. To stay employable she studies most weekends, and nobody pays her for those.
Marcus works on software for medical devices. His work must meet a safety standard called IEC 62304, and a regulator checks it. Learning that standard took him two years, and very few people have done it. When his company advertises a job, they wait months to fill it. His weekend study is close to zero, because the standard changes slowly. His knowledge grows more valuable each year instead of less.
Same salary. Opposite careers.
Compare the two on pay alone and they are identical. On every other measure of how a life actually goes, they are not close. This book is that comparison, done properly, 46 times.
Most career advice optimises one number, which is pay, and treats everything else as a feeling. Is the work interesting? Is the team nice? Will this last? People answer those with opinions, because nobody has organised them into anything you can measure.
That is backwards, because pay is the easiest thing to find out and one of the least useful things to compare. A salary tells you what a job pays this year. It tells you nothing about whether you can enter the field, whether you can stay in it, or whether it will exist in ten years. So this book measures ten things instead of one.
Every profile in Part VI scores a career on these ten dimensions, and here is what each one means.
Each one is scored 1 to 5, and 5 is always the good end for you. Five of these ten name a bad thing, so there the 5 means less of it. A 5 on AI exposure means the work is hard to automate.
1. Regional availability. Where the jobs actually are. Some careers exist in one country. Some exist everywhere.
2. Remote viability. Can you do this work from somewhere else? Some work cannot leave the building, for reasons of law, secrecy or safety.
3. Entry pathway. How people actually get in. Not the route the adverts describe. The route real people took.
4. Pay by region. What the work pays, in each place it exists, from the best sources available. One score cannot hold that spread, and Chapter 12 shows the same work paying very differently by country. So this number sorts careers against each other. It does not predict what you will earn.
5. Competition density. How many people compete for each opening. High pay with high competition can be worse than lower pay with none.
6. Cost to enter. What it costs you to reach your first job. Money, but also time and energy.
7. Cost to maintain and progress. What it costs every year after that, to stay employable and to move up.
8. AI exposure. How much of this work can a machine now do, and how fast is that changing.
9. Other disruption. Everything else that could remove the work. Offshoring, regulation, company mergers, a market disappearing.
10. What employers actually screen for. What decides who receives an offer. Often not what the job advert says.
Two of them are barely published anywhere. Two more are built entirely from people who succeeded. Both weaknesses are worth knowing before you use the scores.
Look again at dimensions 6 and 7: cost to enter, and cost to maintain.
You can find pay data for almost any job in ten minutes. You cannot find these: no government measures them, and no survey asks about them. They appear in no salary report anywhere, and they are the two that decide whether you last.
Priya and Marcus earn the same. Priya pays for her career every weekend, forever; Marcus paid two hard years at the start and very little since. Over ten years that difference is enormous, and it never appears in a salary comparison.
When you look at a career, ask both questions separately. What does it cost me to get in? What does it cost me every year after that? A low entry cost with a high maintenance cost is a trap, and it is the most common trap in computing.
Now a different warning, about dimensions 3 and 10.
Entry pathway says how people get into a role. What employers screen for says what decides who receives an offer.
Ask where both of those come from: job adverts, and the accounts of people who got hired.
Notice who is missing. Nobody anywhere measures the people who took the same route and did not arrive. They sent the applications, did the course, made the sideways move, and are not in any dataset, because nothing collects them.
This book names that problem when it belongs to somebody else. Chapter 7 says of a bootcamp’s own outcome figures that they cannot tell you anything about the people who paid, studied and never got in.
The same objection lands on these two dimensions, and it is fair. They describe routes that worked for people who are now inside. They cannot tell you the odds.
So read them as direction, not probability. “People reach security work from a service desk” is well evidenced. “You will reach security work from a service desk” is not, and this book never has the data to say it.
Chapter 30 maps these routes in detail and repeats the warning there, because that chapter is entirely built from arrivals. Appendix B records it as an open gap. It is arguably the most important thing this book does not know.
Here is the most important idea in this book. Read it twice.
A career is only as good as its worst dimension for your situation.
The dimensions do not average out. One bad score can make the other nine irrelevant.
Take a defence software job in the United States. It might pay very well, face almost no competition, and be nearly safe from AI, so nine dimensions look excellent. But the job needs a security clearance, and a clearance needs citizenship. If you are not a citizen, the score on that one dimension is zero.
Not low. Zero. And zero multiplied by every other strength is still zero.
The same job is superb for one reader and worthless to another. The job did not change. The reader did.
This is why the book cannot simply rank careers from best to worst. A ranking would have to pretend all readers are the same person. Instead, Chapter 17 shows you how to score these careers against your own limits.
The dimension that scores zero for you has a name in this book. It is your binding constraint.
Find it first. Not last.
Most people research careers in the wrong order. They read about pay, grow excited, and spend six months learning. Only then do they meet the thing that rules them out. The clearance. The degree. The visa. The city they cannot move to.
Turn it around. Start with the thing most likely to eliminate options, and filter on that before you look at anything else. If you cannot leave your country, remove every region-locked career on day one. What remains is smaller, and every option in it is real.
A short list of possible careers beats a long list of impossible ones.
You now have the framework, and Chapter 21 shows you how to read a profile that uses it.
The rest of this Part teaches you to check any claim about a job market, including the claims in this book. Part II describes what happened to computing work by 2026. Part IV helps you choose, and Part VI applies these ten dimensions to 46 careers.
One thing to carry with you from here. When somebody tells you a career is good, your first question is not “how much does it pay?” It is “good for whom, and on which dimension?”
A number can come from a serious institution, be reported accurately, and still be close to useless. That is what the four cases below have in common, and why writing about computing careers reads as either too negative or too positive. You have probably repeated at least two of them.
In 2025 a chart spread across the internet. It showed that computer science graduates had 6.1% unemployment, and computer engineering graduates had 7.5%. People used it to argue that studying computing was now a mistake.
The chart was real and so was the source, a Federal Reserve Bank of New York dashboard that is a serious institution. Nothing in it was invented, and it was still close to useless.
The numbers came from survey data, and every survey carries uncertainty, because it asks some people rather than all people.
Good statistics report that uncertainty as a confidence interval, which is a range showing how sure you can be. The Economic Innovation Group checked this one, and for computer engineering the real range ran from roughly 4% to 11%.
The true figure might be 4%, which is excellent, or it might be 11%, which is bad. The chart showed 7.5% as if it were a fact.
A single number without a range is a rumour. Learn that sentence; it will protect you more than anything else in this book.
The same dataset contained another number: about 90% of computer science graduates were employed.
Both numbers are true at once. Unemployment counts only people actively looking for work, while the employment rate counts everyone. They answer different questions and leave very different impressions, and nobody shared the second one, because it was not alarming.
The chart used data collected in 2023, published later. People shared it in 2025 as a description of the market that year.
Two years is a long time in this field, and Chapter 9 explains what changed in between.
The source was excellent and the number was real. The conclusion was still wrong, because the number arrived with no range, no context and no date attached.
You will meet these figures if you read anything about mainframes.
COBOL, they say, runs 70% of the world’s business transactions. It handles 95% of cash machine withdrawals. There are 220 billion lines of it still running.
These numbers appear everywhere, and something strange happens if you follow them backwards.
Most articles cite a Reuters piece from 2017. The Reuters piece took the figures from Rocket Software, a company that sells mainframe products. Rocket Software’s figures trace back to a survey by a firm called DataPro. That survey ran in 1997, and it asked 421 companies.
So a claim about the whole world’s banking system in 2026 rests on 421 companies answering a question three decades ago. Nobody has measured it properly since.
This is a citation chain: article cites article cites report cites survey, and each step looks respectable. The number gets stronger as it travels, because each new article states it with more confidence than the last.
The number is not stronger. It is just older.
When a figure matters to your decision, follow it back until you find the original measurement, and then ask three things about it. How many people? What year? Who paid?
Often the chain simply stops, and there is nothing at the end of it.
No, and this matters: Chapter 25 recommends mainframe work.
The point is narrower than that. COBOL genuinely runs critical banking systems, but the specific percentages everyone quotes are not evidence. The career is real. The statistic is decoration.
You can believe a claim and still reject the bad number used to support it.
For years, one figure dominated every article about cybersecurity careers: the world needed 4.8 million more security professionals, and there was a huge gap.
ISC2, a large security membership body, published it in 2024, a 19% rise on the year before, and journalists repeated it constantly.
In 2025, ISC2 published its workforce study again. This time it had a record 16,029 respondents, so the study was bigger and better than before.
The 4.8 million figure was not in it.
ISC2 did explain it. Respondents in 2024 and 2025 had put critical skills ahead of headcount, and the study says that is why the gap estimate is missing. The new framing was about skills rather than bodies. It reported that 95% of teams lacked at least one specific skill, and that 59% called the gap critical or significant.
Note what the organisation did here, because it is the honest move and it is rarer than the alternative. It did not quietly drop the number and hope nobody noticed. It said why it had stopped publishing it, and its stated reason is the same reason this chapter is about to give you.
When an organisation stops publishing its own headline number, pay attention, because that silence tells you something.
Here it tells you something useful for your career. “Millions of open jobs” was never the right description. Employers do not want any 4.8 million people; they want specific skills they cannot find. Those are different problems, and they lead to completely different advice for someone entering the field.
Chapter 25 builds on this: the security shortage is real, and it is not a shortage of bodies.
Now the most useful question in this chapter. Who profits if you believe this?
The COBOL shortage. Article after article warns of a coming crisis as mainframe programmers retire, so look at who publishes those warnings. Training companies that sell COBOL courses. Staffing firms that place mainframe contractors. Consultancies that sell modernisation projects.
The shortage may well be real, but every organisation making the loudest noise about it earns money from your belief in it.
Quantum computing. For years, projections circulated that quantum would need two million workers by 2025, or 500,000 new hires by 2025.
Then somebody counted. QED-C, an industry body, measured the actual global quantum workforce at about 16,500 people in its 2026 report.
Not two million. Sixteen thousand five hundred.
The projection was not slightly optimistic. It was wrong by a factor of more than a hundred.
QED-C’s 16,500 is a count of people who exist. The two million was a guess about people who might exist one day.
Both get printed as numbers. Only one is a measurement.
Chapter 6 applies this specifically to claims about artificial intelligence, where the same confusion does the most damage.
You now have a checklist, and it works on any number about jobs, including every number in this book.
1. Is there a range? A single number with no interval is a rumour.
2. Where does it actually come from? Follow the chain to the original measurement. Check its size and its date.
3. Who benefits if I believe this? Training companies, staffing firms and software vendors fund most shortage claims.
4. Is this a count or a projection? Counting is measuring. Projecting is guessing with a spreadsheet.
Four questions. Ten seconds each. They will save you years.
Every source of career data falls into one of five grades, and the grade tells you what a number from it is worth. Lower is better. Chapter 4 judged four bad numbers one at a time; this is the system that saves you doing that.
These organisations measure things themselves, publish their method, and have no product to sell you.
Government statistics offices. The Bureau of Labor Statistics in the United States. Eurostat in Europe. The OECD and the International Labour Organization. National statistics offices in most countries. Federal Reserve research banks.
Also here: academic research surveys with published methods. The CRA Taulbee Survey, which counts computing degrees and faculty in North America. The National Science Foundation’s Survey of Earned Doctorates. Government immigration authorities publishing their own visa rules and thresholds.
Grade 1 sources are slow, usually one to two years behind. That delay is the price of getting it right, and it is worth paying.
Large surveys run by organisations with a real interest in the answer, but with a published method and a stated sample size.
The Stack Overflow Developer Survey. The JetBrains developer ecosystem survey. ISC2 and SANS in security. NACE and Handshake on graduate hiring. QED-C on quantum.
These are useful and often current, with two weaknesses. The people who answer choose to answer, so they are not a random sample. And the organisation usually has a view it would like confirmed.
Read them, and read the method section before the results.
This grade is unusual, and it is the one most people underrate.
These are records of money that actually changed hands, checked by somebody with a reason to check.
Labor Condition Application filings. A United States employer sponsoring a foreign worker must file the wage in public and sign it. Lying is a crime. In several fields these filings are the closest thing to the truth about pay.
Payroll platforms that verify. Firms like Howdy publish salary data taken from payroll they run themselves. Vetted contractor sites such as Lemon.io publish rates people were really paid.
Levels.fyi and similar sites, which check submissions against offer letters.
Grade 3 has one known weakness: selection bias. The data only covers people who used that route. Wage filings cover sponsored workers, not everyone. A contractor platform covers contractors who passed its test.
So grade 3 numbers are accurate about a specific group. Your job is to check whether that group is your group. Chapter 25’s Epic profile shows this failing in practice, where excellent wage data described the wrong population entirely.
Salary websites. Glassdoor, Payscale, Salary.com, ZipRecruiter.
People type in their own pay and nobody checks it. The site mixes different jobs under one title, and different countries under one average.
They are also, for most of the world, the only pay data that exists at all.
So this book uses them, and labels them every time, under one rule. A grade 4 number gives you a direction, never an amount. If it says senior pays more than junior, believe that. If it says senior pays $127,431, believe only the first digit.
Here is an exercise that takes ten minutes and will change how you read salary sites forever.
While researching Chapter 26, I asked one large salary site what three different specialists earn in the United States. A precision agriculture software engineer. An energy grid software engineer. A government software engineer.
Three unrelated fields, in three different industries, checked in three different months of 2026.
The site gave the same answer to all three. $147,524 a year. Not a similar number. The identical number, down to the last dollar. The middle band was identical too, $120,000 to $173,000 in every case.
Then I asked the same site a fourth question. What does a software engineer earn in the United States?
$147,524. Middle band $120,000 to $173,000.
Now you know what happened. The site has no data about precision agriculture software. When you ask it a question it cannot answer, it does not say so. It quietly hands you the national average for all software engineers, wearing the job title you typed.
A specialist figure that equals the general figure is not a specialist figure.
This matters more than it might seem. The whole reason you look up a niche career is to find out how it differs from the ordinary one. On that exact question the number is silent, and it is silent in a way that looks like an answer.
Check it before you trust any pay figure for a narrow field. Look up the general job title as well as the specific one. If the two agree closely, you have learned nothing about the specialism.
One caution about my own evidence. The general software engineer page is the only one of the four I could open and read directly. So treat this as a method to repeat, not a fact to accept from me. That is the whole point of this chapter.
Pages published by a company that sells the thing the page recommends.
Training providers describing skill shortages. Bootcamps publishing graduate salary claims. Staffing firms describing talent scarcity. Certification bodies measuring the value of their own certifications.
Treat these as advertising, because that is what they are. This book cites them only as examples of claims to distrust.
Even good numbers mislead when you compare them wrongly. Five traps do most of the damage.
Gross pay is before tax. Net pay is what reaches your bank account.
The gap between them varies enormously by country. Comparing a gross salary in one country with a net salary in another produces a meaningless answer.
Base salary is your fixed annual pay. Total compensation adds bonus and company shares.
In most jobs these are close, and in some they are not remotely close. A quantitative researcher may file a base salary of $190,310 while actually earning three times that. Chapter 26 covers this.
Always ask which number you are looking at. Recruiters quote whichever is larger.
In Poland, a large share of developers work on B2B contracts rather than employment contracts. They invoice as a business, and roughly 38.5% of the market works this way.
A B2B rate looks much higher than an employment salary. It also carries no paid holiday, no sick pay, no notice period, and the worker pays their own social contributions. So the headline numbers are not comparable, and many articles compare them anyway.
In Brazil, employment under the CLT labour code costs an employer roughly 1.6 to 1.8 times the salary, once mandatory contributions are counted.
This means a Brazilian employer paying the same total cost as an American one offers a much smaller salary. The difference is not stinginess. It is law.
$30,000 in Manila and $30,000 in Zurich are not the same money.
Purchasing power parity compares what money buys locally rather than what it converts to. On that measure, many salaries in lower-income countries are far better than they look, and some Western salaries are far worse.
Neither measure is wrong. Use nominal figures when comparing what you could send abroad, and purchasing power when comparing how you would live.
Everything above ranks sources by who published it. That is one axis, and on its own it will mislead you.
The second axis is whether the figure answers your question. Call it fit.
A source can be perfect on the first axis and useless on the second. The United States government publishes exactly what its own digital service pays in Washington DC. That is grade 1, audited, and impossible to argue with. It is also one small programme in the most expensive city in the country, so as an answer to “what does government software pay?” it is worse than a mediocre average.
Now the uncomfortable direction. Suppose a grade 4 aggregator actually surveyed machine learning engineers. It may tell you more about machine learning pay than a grade 3 contracting rate for developers in general. Worse provenance, better fit.
A precise measurement of the wrong thing beats nothing, and loses to a rough measurement of the right thing.
This book got that wrong and had to be told. It printed a verified contract rate for senior developers as the pay figure for five different roles. Partly because the rate was grade 3 and the alternatives were grade 4. Grade 3 about the wrong people won over grade 4 about the right ones. Chapter 23 explains what replaced it.
So ask both questions, in this order.
1. Who measured it, and what do they sell? That is the grade.
2. Did they measure the people I am asking about? That is the fit. Check the population, the country, the seniority and the year.
A figure needs to pass both. Where this book has a strong source covering the wrong population, the profile marks it with a dagger and says so. The grade does not get to speak for the fit.
Every figure in Part VI carries a grade.
Some careers have no grade 3 pay data. This book says so plainly and lists them in Appendix B. It does not fill the space with grade 4 numbers.
Hold this book to that rule. Every number here should carry a source, a date and a grade. If one does not, that is a mistake. Appendix B tells you how to report it.
One distinction clears most of the noise around AI and jobs. It is whether a number counts something that happened, or estimates something that might. No topic in this book attracts more noise, which is why it gets a chapter of its own.
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. That guess may be careful and well informed. It is still a guess, so never weight it like a measurement.
Three pieces of real evidence matter as of 2026. They agree with each other, which is why this book trusts them.
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%, and 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.
You will meet several figures from this one study, and people treat them as better and worse versions of one thing. They are not: some measure different groups of people, and 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, and they answer different questions. Question three later in this chapter is compared against what?, and 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.
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 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 but the shape of it. 71% of the increase came from senior roles.
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, and 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?
Here is the honest difficulty, and most articles skip it.
Three things happened to technology hiring at the same time.
All three push hiring down, and 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.
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.
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.
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 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.
Training is the largest payment most readers of this book will make before they earn anything at all. It is also bought with the worst information in the book. A degree, a bootcamp, a certificate, a course: Chapter 5’s five grades of evidence all get spent here.
Almost every number you can find about a course comes from the people selling the course.
By Chapter 5’s system that is grade 5. Treat it as advertising.
This is not a claim that schools lie, and some publish very carefully. It is a claim about who holds the data. A school knows how many students finished and how many found work, and nobody else does. So the school decides what to publish, and it decides how to count.
That is the whole problem. You are not reading a measurement. You are reading a choice about how to measure.
Here is the most common trap. It needs no dishonesty at all to work.
A school publishes a placement rate. A rate is a fraction. The top is students who found work. The bottom is the group being counted. The bottom is where the argument lives.
Most published rates count only graduates who were looking for work. That bottom number leaves out three groups. It leaves out students who dropped out. It leaves out students who finished and did not look. It leaves out students who never answered the survey.
Work through a real example. App Academy publishes its whole funnel, which makes it both a good example and a fair one. Most schools publish only the last step.
Of every 100 students who start, about 80 graduate. Of those graduates, 98% were looking for work. Of the ones looking, 91% found work in the field within 180 days.
The headline is 91%.
Now multiply the three numbers. 80 students graduate. About 78 of them look for work. About 71 find it. So roughly 71 of the original 100 got a job in the field.
Both numbers are true, and they answer different questions. You are asking the first one, and the school is answering the second.
So ask every provider the same question, in these words. Percent of what? If the answer is not “of everyone who enrolled”, do the multiplication yourself. If they will not give you the numbers to multiply, you have your answer.
Ranked, best first. This is Chapter 5’s hierarchy applied to one industry.
An outside body checks the numbers and publishes them on a fixed definition. The Council on Integrity in Results Reporting runs the only widely used standard of this kind. Member schools report outcomes at 90, 180 and 360 days, on a defined bottom number, and an independent auditor checks them.
Very few schools take part. That is itself the useful signal. You do not need to know the exact membership. You need to ask one question: does this school submit its numbers to an outside audit? A school that does has chosen to be checked. A school that does not has chosen not to be.
The next section is about these. A consumer-protection agency can compel documents. It sees the internal numbers, not the marketing ones. When such an agency publishes what it found, that is grade 1 evidence. It is the only kind in this book where the publisher could force the answer.
Websites that compare schools mostly reprint what schools tell them. The site did not measure anything. A number that starts at grade 5 does not improve by being copied.
Grade 5. Read it to learn what the school wants you to believe.
Three cases follow. All three are United States actions, and that matters. I come back to it at the end of this chapter.
BloomTech, April 2024. The Consumer Financial Protection Bureau acted against the school and its chief executive. The agency found the school had advertised job placement of 71% to 86%. Its own internal figure was close to 50%.
The same action covered how students paid. The school sold agreements requiring anyone earning over $50,000 to pay 17% of their income before tax. That ran for 24 payments, up to a cap of $30,000. The school said these were not loans and carried no debt. The agency found they were loans, with a finance charge averaging about $4,000. One missed payment could put a student in default.
The school was permanently banned from consumer lending. Its chief executive was banned from student lending for ten years.
Prehired, November 2023. The same agency shut this provider down. It sold a job guarantee, again funded by income-share agreements. The order made it stop operating, pay $4.2 million back to students, and cancel about $27 million of outstanding agreements.
Career Step, 2024. The Federal Trade Commission settled over deceptive job-placement claims aimed at military families and veterans. The settlement was worth $43.5 million in cancelled debt and cash. Eight months later the Commission paid out $15.5 million, to 42,794 people.
Hold that last pair of numbers next to each other. The settlement headline was $43.5 million. The money that reached people was $15.5 million, and it arrived in March 2025. A settlement is not the same as a refund. Neither one returns the years those students spent.
An income share agreement is a deal where you pay a share of your future income instead of paying tuition now.
Sellers describe it as risk-free, and the argument sounds fair. If you do not earn, you do not pay. So the school only wins when you win.
Read the BloomTech finding again before you accept that. The regulator found those agreements were loans. It found they carried a real cost in money. It found the income level that started repayment was low enough to catch ordinary outcomes, not only good ones.
Check three things in any such contract, before you sign anything.
The Federal Trade Commission’s standing advice on this is one sentence. Do not pay for the promise of a job. It says that anyone who asks you to is a scammer.
Any offer built on a job guarantee deserves that sentence read twice. Chapter 16 sets out what a parent should refuse outright.
A certificate is a different product, and it fails in different ways.
A course sells you teaching. A certificate sells you a test. The test either passes or it does not, so a certificate cannot lie about its own outcome. What it can do is oversell what that outcome is worth.
Some certificates are cheap, stackable and honest about their place. A stackable certificate is one that counts towards the next one up. CompTIA publishes such a ladder: A+, then Network+, then Security+, then a specialist one. Each higher certificate renews the ones below it, so the cost of staying current falls as you climb.
That whole entry ladder costs a few thousand dollars and takes months. A bootcamp costs three to five times as much. This does not make one better than the other. It makes them different bets, and Chapter 15 compares them properly.
The honest limit is this. A certificate proves you passed a test. It does not prove you can do the work, and employers know that. Chapter 31 sorts the certificates that move you into a different tier of job from the ones that only break a tie.
In the 2024 Stack Overflow developer survey, 82% of working developers said they learn using online resources. Only 49% said they learn in school. And 66% of them hold a university degree.
Read all three together. Most working developers taught themselves something from free material. Most of them also have a degree. Both facts are true at once, and no course advertisement will show you both.
That survey has a hole in it, and it is a large one. People answer it because they already work in software. It cannot tell you anything about the people who paid, studied, and never got in. Nobody measures those people. That is exactly why a placement rate matters so much, and exactly why the bottom of the fraction matters more than the top.
Everything above assumes a private provider. A university is not the same problem, and the difference is worth knowing, because it runs the opposite way to what most people expect.
For bootcamps there is almost no independent outcome data. For degrees, in several countries, there is a lot of it, and it is grade 1.
Some governments match graduate records against tax and employment records. They then publish what people from a named course at a named university actually earned, one year later and five years later. That is not a survey, and it is not self-reported. It is tax data.
Where that exists, it beats every other source in this chapter. It has no survey response problem, because nobody opts out of tax. It counts everyone who finished, so the bottom of the fraction is fixed for you.
So the first question about a degree is not “is this university good?”. It is: does my country publish course-level graduate earnings, and what does it say about this course? Appendix G answers the first half for you, country by country.
Two warnings about those figures.
They measure the students who went in, not only the teaching. A course that admits only the highest-scoring applicants will report high earnings whether or not it taught anyone anything. The number tells you where graduates end up. It does not tell you why.
They are also old by the time you read them. A five-year figure describes people who started eight or nine years ago. Chapter 4 makes this point about dates in general, and it applies here with force.
Now the part of this chapter that took longest to establish and is the least comfortable to write.
Eleven publishers were confirmed against their own primary source. The United Kingdom, the United States, Canada and France. Ireland, the Netherlands, Poland and New Zealand. Chile and Colombia. And the shared register systems of Norway, Denmark, Sweden and Finland.
Then look at where none was found. Nigeria. Pakistan. Bangladesh. Indonesia. Vietnam. The Philippines. Egypt. Kenya. Ukraine. Turkey. Argentina. South Africa.
Five more have something close and not the same thing. Mexico publishes graduate outcomes that are not matched to tax records. Brazil has a national employment register, and whether it can be read by course was not established. Australia and Singapore run graduate surveys, which leave out everyone who did not answer. India’s institutions report their own placement medians, which is exactly the self-reported figure this chapter warns you about.
Read those lists together. The strongest evidence about the value of a degree exists mostly in rich countries. It is largely absent from the countries sending the most people into computing, and from most of the countries this book is written for.
That is not a small inconvenience. It means two readers of this chapter, doing everything right, end up with different qualities of answer. One opens a government file and reads what people from their exact course actually earned. The other has the university’s own placement figures, self-reported by the institution being judged, so the honest instruction differs by where you live.
If your country is on the first list, use it. It outranks everything else in this chapter and it takes twenty minutes.
If it is not, do not substitute a salary website, because that is worse than having nothing and believing nothing. Lean instead on what your country does publish. Entrance-exam data. The questions in this chapter, put to the institution in writing. The employer-side evidence in Book B. Chapter 8 shows you how to find your own country’s primary sources.
One caution about that second list. Absence of a finding is not proof of absence. Those searches ran in English, and a dataset published only in a national language could have been missed. What the book claims is that it did not find one, which is a statement about this edition rather than about your country. If you find one, Chapter 20 explains how to send it back.
One more test, and it cuts across every kind of training.
In HackerRank’s 2025 survey of 13,732 developers, managers and recruiters, 77% said that most assessments do not match the skills the role needs. Two thirds said they prefer practical challenges to abstract coding problems.
Read that as a buying instruction. Hiring is moving towards asking people to do job-like work. So the training worth paying for is the training that leaves you holding job-like work.
At the end of the course, what do you have? If the answer is a certificate and nothing else, you have bought a claim. If the answer is three finished things you built, that a stranger can look at and test, you have bought evidence.
Note who published that survey: HackerRank sells technical assessments, and a finding that current assessments are wrong is useful to them. The sample and the method are both stated, so it stays at grade 2, but read the conclusion with the seller in mind. Chapter 5 taught you to do that, and it applies to sources this book likes as much as to ones it does not.
Ask these before you pay, and ask them in writing.
A provider that answers all five plainly may still be wrong for you. A provider that will not answer them has already told you what you need to know.
Three honest limits.
All three cases above come from the United States, and that is no longer the whole picture. Appendix G lists a consumer regulator with power over training providers in seventeen other countries.
In eleven of them, something has happened to a provider. Only eight of those were fines. One national body punished thirty-one coaching institutes over what they claimed about selection and placement.
In five countries no action was found at all. In one more, the regulator looked and found nothing wrong.
Do not read that as proof that providers there behave better. Several of those regulators publish almost nothing in English, and settle most complaints by talking to both sides. That leaves no record for anyone to find. The regulator, the law and the complaint route all exist anyway, and Appendix G lists them.
What regulators do also changes over time, and with governments. These cases say what happened once. They do not promise that the same scheme would be stopped today, in any country. What protects you is the method in this chapter. Do not wait for a regulator to arrive.
One more limit. No primary source told me how many schools currently submit to an outside audit. So this chapter does not tell you a number. It tells you to ask the school, which works whatever the number is.
This book will age, as every book of numbers does. What follows is how to check any career, in any country, yourself. After this you do not need this book to be your only source, which is the point.
Four steps. About thirty minutes for a career you know nothing about.
Step 1. Count the jobs. Search one job title on two or three job sites for your country. Count roughly how many results appear. Then narrow to entry level and count again.
The second number matters more than the first. A field with many senior openings and almost no junior ones is a field you cannot enter, whatever its total looks like.
Step 2. Read twenty adverts properly. Not the pay. The requirements.
Write down what appears in most of them. That list is what employers actually screen for, and it is usually different from what courses teach.
Step 3. Find one grade 1 or grade 3 number. Use the hierarchy from Chapter 5. One good number beats ten salary-website numbers.
Step 4. Ask who is selling. Search the career name with the word “shortage”. Look at who publishes the results. If they sell training or staffing, adjust your belief downwards.
Your national statistics office. Slow, boring, reliable. Measures employment and pay by occupation. One to two years behind.
Bureau of Labor Statistics Occupational Outlook Handbook (United States). Ten-year projections by occupation, plus current pay. The projections are projections. The current figures are solid.
Eurostat for European Union labour data.
Indeed Hiring Lab. Measures job adverts, not jobs. Fast, and useful for direction. Remember that an advert is not an opening, and one opening can produce several adverts.
Stack Overflow Developer Survey. Measures what people who answer surveys say. Good on technology use. Weaker on pay.
Levels.fyi for large technology company pay, checked against offer letters. Covers a narrow slice of the market extremely well.
Your country’s immigration authority, directly, if a visa matters to you. Never trust a blog for a threshold. These change, and the official page is the only one that counts.
Most career writing assumes you live in the United States. This book does not. So here is where to start, wherever you are.
Every country with a working government runs a labour force survey. It is usually free and open to anyone. The websites that want to tell you about salaries almost never mention it.
| Where you are | Start here |
|---|---|
| United States | Bureau of Labor Statistics, and O*NET for what jobs involve |
| European Union | Eurostat, then your own national office |
| United Kingdom | Office for National Statistics |
| Germany | Destatis, and the Federal Employment Agency for vacancies |
| France | INSEE · Italy ISTAT · Spain INE · Poland GUS |
| Canada | Statistics Canada · Australia ABS · New Zealand Stats NZ |
| India | Periodic Labour Force Survey, from the statistics ministry |
| Brazil | IBGE · Mexico INEGI |
| Nigeria | National Bureau of Statistics · South Africa Stats SA |
| Philippines | Philippine Statistics Authority · Indonesia BPS |
| Japan | e-Stat portal · South Korea KOSTAT · Singapore Ministry of Manpower |
| Anywhere at all | ILOSTAT, run by the International Labour Organization |
Now the part that matters more than the list.
A primary source will answer a coarser question than the one you asked.
Two real examples. ILOSTAT publishes occupation data at the second level of the international classification. That means it can tell you about software and applications developers as a group. It cannot tell you about Epic analysts, because no statistical system in the world has a category for them.
India’s labour force survey now publishes monthly, which is fast for official data. What it publishes monthly is participation and unemployment. Occupation and earnings detail sits in the annual report, which arrives later.
So expect three things from a primary source. Broad categories, a delay of months or years, and total honesty about both. That is the trade. You are swapping precision for the certainty that nobody is selling you anything.
Use the official figure as your anchor and the fast sources as your direction. Suppose a job board says a role pays double the official figure for its occupation group. One of them is wrong, and it is usually not the government.
Finding your own country’s page is a twenty-minute job you do once. Do it before you finish this chapter, and write the address down somewhere permanent.
Adverts are the fastest signal available. Read them as a group, not one at a time.
Three things to track.
Volume over time. Search the same title every month and note the count. The direction matters more than the number.
Seniority mix. Count how many say junior, graduate or entry level. Compare against senior, staff or principal. This single ratio would have warned you about the entry-level squeeze a year before the articles did.
Requirement drift. Compare adverts from this year against last year, where you can find them. Requirements creep upward quietly. A role that asked for two years now asks for five. Nobody announces this.
Someone tells you a career pays a certain amount. Here is how to test it.
If a claim survives all six, treat it as probably true.
Search engines will show you what is optimised, not what is true.
Search any career with the word “salary” and the first page will be salary websites and training companies. That is grade 4 and grade 5 material, arranged by whoever spent most on being found.
Two habits fix this. Add the name of a statistics body to your search. And go directly to the sources you bookmarked rather than searching fresh each time.
Apply all of this to Part VI.
Every profile carries its sources, dates and grades. If a number looks wrong to you, follow it. If you find an error, Appendix B explains how to report it, and Appendix F records what changed in each version.
A book that teaches you to distrust weak sources has to accept the same treatment. Chapter 20 lists the specific findings that would prove this book wrong.