AI-proof jobs are not a list of titles: what holds up is the part of any job that means being accountable, choosing, and knowing what never got written down. Dmitry Kargaev wrote this page from chapters 7 to 10 of his book Don't Replace Me: A Survival Guide to the AI Apocalypse.

This site, dontreplace.me, belongs to the book's author, so everything said here about the book comes from its maker. The book's core claim is that "AI takes over tasks much faster than it takes over jobs, and it does it unevenly, company by company, desk by desk." Its first chapter turns that into advice: fear tasks, not jobs.

What published lists of AI-proof jobs say

Three sources that name specific titles are listed below, with what each says and what kind of source it is.

Taste: a choice you sign (chapter 7)

Chapter 7 of Don't Replace Me argues that taste, meaning the choosing and finishing of one output among many, is the part of creative work that holds up, because anyone can now generate options. Rule 7 reads: taste is a choice you sign.

The chapter opens with Jason Allen, who took first place and $300 at the Colorado State Fair in August 2022 with a Midjourney piece. Press reports cited in the book say he prompted and revised at least 624 times, generated about 900 images and spent more than 100 hours. The Copyright Office review board refused to register the work in September 2023. Allen sued in September 2024, and the book says it could not find a ruling. Its summary of the law is that only the human-made parts are protected: "The machine's raw output is a public commodity, like tap water: anyone can fill a glass, and nobody can register one."

The exercise takes 30 minutes. Ask an AI for 20 versions of something small, such as a subject line, cross out the embarrassing ones, edit one, and write a sentence on why you picked it.

Being the one who answers for the work (chapter 8)

Chapter 8 argues that what does not automate is accountability: someone has to stand behind the work, take the blame and show up at personal cost. Rule 8 is: be the one who answers for it.

The chapter concedes that AI can comfort. It cites a Dartmouth trial of Therabot, published in NEJM AI in March 2025, with about 210 adults over four weeks. The book reports depression symptoms down 51% and anxiety down 31%. It also quotes Heinz: "no generative AI agent is ready to operate fully autonomously in mental health". On the book's account, comfort was never the protected part.

Its case is Jake Moffatt's, who booked a flight in November 2022. Air Canada argued its chatbot was "a separate legal entity that is responsible for its own actions." Tribunal member Christopher Rivers rejected that in a ruling of February 2024, and about CA$812 was awarded. The book's line: "A model can't be fired, sued or embarrassed, and the client knows it."

The exercise is an "it's on me" script for when AI-made work has a mistake. Say what was wrong, give the correct version, say it is yours, say how you checked the rest and what changes. Mention the tool once as a fact, never as the culprit.

This came from a book.

Don't Replace Me

200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff.

Get the Book →

Writing down what nobody wrote down (chapter 9)

Chapter 9 argues that the knowledge worth protecting is tacit knowledge that never got typed, and that you should write it down yourself, on your terms. It cites Polanyi, 1966: "we can know more than we can tell". Rule 9 is: write down what nobody wrote down.

The chapter opens with IBM Watson at MD Anderson, announced in 2013. The book reports a cost of about $62 million and a contract extended 12 times. An audit in February 2017 judged the money and the paperwork, and the auditors said they were not judging the science. IBM disputed the picture the audit drew.

The urgency, in the book's telling, is that AI tools now remember: "any job-security plan built on its amnesia went out with the old title."

The exercise is a two-week room log. After each meeting, write three lines: what was decided but is not in the minutes, who reacted how, and one thing you now know that is in no document. On the second Friday, mark each line with a check, safe to give an AI, or a lock.

Deciding what to build (chapter 10)

Chapter 10 argues that choosing the goal and the trade-offs stays with people, even though agents can now plan: they plan toward a goal someone typed. Rule 10 is: decide what to build.

It begins with Kasparov against Deep Blue, game 6, on 11 May 1997. Human-plus-engine "centaur" chess, the book says, worked for about fifteen years before engines stopped needing humans. A 2005 freestyle event was won by amateurs Cramton and Stephen, rated 1685 and 1398. Work differs from chess because nobody hands you the goal: "Chess is a game made entirely of referee."

The exercise is to write the brief before the prompt, in five lines: Done looks like, I'm trading, I'm not trading, Who's disappointed, I'll know it worked when. Paste the first three into the prompt and judge the output against the fifth. Do it three times.

Exposed jobs: one task a computer can check (chapters 3 and 19)

Chapter 3 of the book says jobs that are one checkable task are in real trouble, and that entry-level work is where it shows now. Its pile test takes 15 minutes. List the five tasks that ate most of your hours, then score six yes or no questions: whether a chatbot could draft the task, whether a computer can check it, whether a stranger could do it in a month, whether you are in the first three years of your career, whether nobody outside your team asks for you by name, and whether nobody asks your opinion. A score of 0 to 2 is the bundle pile, 3 to 4 the changing pile, 5 to 6 the exposed pile.

The book cites the Stanford paper "Canaries in the Coal Mine?" by Brynjolfsson, Chandar and Chen. Its August 2025 version found a relative drop of about 13% for ages 22 to 25 in exposed jobs, and an August 2026 revision with data to June 2026 found about 19%. The authors do not claim causation.

Chapter 19 separates two situations. Task automation, which the book says covers most people, is a reason to keep going. Industry disruption, where AI makes your employer's product free, is a reason to move sideways while you are still employed. Its 15-minute exercise scores your tasks 1 to 4 and your industry 1 to 4, plots them on a grid labelled chill, pivot, level up and run, and sets a six-month reminder to retake it.

The four traits at a glance

The traits from chapters 7 to 10 fit in one table.

TraitWhy it holds up, per the bookChapter rule
TasteGeneration is cheap for everyone, so the person who chooses and can say why stands apartRule 7: taste is a choice you sign
AccountabilitySomeone has to stand behind the work, be blamed and show up at personal costRule 8: be the one who answers for it
Unwritten knowledgeTacit knowledge never got typed, and it stays yours to judgeRule 9: write down what nobody wrote down
Choosing the goalAgents can plan, but toward a goal someone typedRule 10: decide what to build

What is not known

The sources state several limits.

The book behind this page

Don't Replace Me: A Survival Guide to the AI Apocalypse is a second edition rewritten for November 2026, with 22 chapters that each end in one rule and one exercise. The Amazon listing read on 11 November 2026 shows a Kindle ebook at $2.99, a 174-page paperback at $14.99 and a hardcover at $24.99. The introduction is a free download on this site. There is nothing to buy after the book.