What will remain of professions in the AI era
In August 2026, three notable texts were published about a future dominated by artificial intelligence.
At the start of the month, Canadian writer Cory Doctorow published an analysis of a consultant’s observations after a year and a half inside corporate AI projects, finding no working use cases. Nine days later, Mark Zuckerberg promised a personal superintelligence for every person on the planet. Two weeks after that, Bill Gates began his essay with an admission: there is no plan for entering the new era.
Three visionary perspectives, three incompatible pictures, but the authors agree on one point: the technology is already delivering significant returns. We examine what underpins that confidence, who pays for the overhaul of workflows, and why writers, programmers, and designers are increasingly cast as checkers rather than authors.
A debate without a common baseline
Zuckerberg’s essay The Future is for Everyone rests on three principles:
human empowerment as the source of prosperity;
invention, not mere automation, as the purpose of superintelligence;
a balance of power between people and institutions as the basis of safety.
The Meta chief does not believe safety is achieved by concentrating power.
The company, with a $1.47 trillion market capitalization, proposes:
a personal agent working for the user around the clock;
a fully private mode in which data are not accessible even to the company itself (modeled on WhatsApp encryption);
an auction mechanism to allocate compute to customers willing to pay for more capacity.
The company also says it will resume releasing open-source models.
Gates strikes a different tone: he calls the transition to a new era one of the most turbulent periods in human history and says no preparation is underway.
Three risks:
Jobs will disappear permanently.
AI will expand opportunities to “cause harm.”
Machines will start to substitute for human relationships.
Three proposals:
New institutions to manage the transition.
A list of occupations society agrees not to automate.
A tax on AI tokens and robots.
In Gates’s view, it took personal computers twenty years to reformat office work: first programmers had to write business software, then manufacturers cut hardware prices, and after that it took time to retrain employees.
Large language models LLMs need none of that. They run on existing devices and require no interface training — you address them in natural language. The technology adapts more than the user does. It can watch the same training video new employees see and absorb their skills.
Gates expects AI to change the balance of power in law, customer support, medicine, development, and manufacturing — a broad transformation in less than a decade.
Source: Meta, Gates Notes.
None of these essays tests the assumption of effectiveness. It is taken as a given, and it underpins all subsequent claims: on taxes, a “reserve” of professions, and how power over the technology should be distributed among corporations, states, and individuals.
A consultant took on the testing after his firm spent eighteen months working with market participants who were optimistic and friendly toward AI.
Belief over metrics
Nikhil Suresh, who leads sales and technical delivery at Hermit Tech, held about 300 conversations with professionals worldwide — from niche specialists to executives at Fortune 500 companies. He summarized the results in AI Mania Is Eviscerating Global Decision-Making.
In one episode, a top manager at a company with more than $2 billion in revenue presented a technology strategy built entirely around AI. It soon emerged the “innovator” had never opened ChatGPT or any other AI tool. Suresh does not name him — such admissions can cost people their jobs, he says.
Claims from Hermit Tech clients of 100x productivity gains became routine. Yet over eighteen months, Suresh’s team recorded a zero share of successful AI projects. Failures included both engagements the firm worked on and projects it observed from the outside.
Numbers highlight the gap between promises and outcomes. Cortex, in an ideal configuration, answers correctly in about 92% of cases, according to Snowflake staff at a presentation. That means at least eight answers in every hundred mislead. And someone has to catch each error before it reaches a client or goes into a report.
Beyond that, “human capital efficiency” is measured by how much is spent on AI. Suresh calls this metric “fully gameable” — anyone can inflate it. Where “token leaderboards” are in place, teams report by the volume of generated text.
One such employee described his tactic:
make a copy of the working repository;
ask a model to rewrite code from one programming language to another;
work on the real task in parallel.
The twin project serves no purpose, but token usage climbs and the job stays. Why schemes like this don’t surface at the leadership level is the essay’s central question.
Suresh describes a coordination problem. Those who publicly confirm productivity gains keep their jobs. Admitting the opposite sounds like accusing colleagues of lying or incompetence — that leader is fired and replaced by someone willing to continue the “party line.”
Doctorow spelled this out in a column for The Nerve. In organizations of 500 people or more, those who declared the technology transformative received promotions and protection from cuts. Employees with well-founded objections landed on layoff lists.
Blog readers with the title Head of AI at companies with more than $1 billion in annual revenue told Suresh they cling to a fictitious role purely for career advancement.
At some point, Hermit Tech stopped asking clients about ongoing AI initiatives. Once a project starts, it can no longer be discussed candidly — until a crisis hits.
The sample combined the firm’s clients, outside projects it watched, and off-the-record conversations. Respondents were anonymized, the data-collection method was not published, and there was no independent verification. Hermit Tech also declined all AI implementation work — which makes the observer more independent but reduces access to successful cases.
Layoffs under the AI banner
In early 2026, Jack Dorsey announced layoffs of nearly 4,000 employees at Block — shrinking headcount from about 10,000 to under 6,000. After the close on February 26, the stock rose about 20%.
About six weeks later, Snap took an analogous step. CEO Evan Spiegel cited progress in AI as the reason for restructuring. Roughly 1,000 employees — 16% of staff — were cut. On April 15, the stock rose 9%.
Shortly before Dorsey’s announcement, OpenAI head Sam Altman told CNBC that large companies use artificial intelligence as a convenient explanation for layoffs that would have happened anyway.
Goldman Sachs estimated the scale of the effect from AI adoption. According to the bank, over a year the technology reduced monthly U.S. job growth by roughly 16,000 and raised unemployment by 0.1 percentage point.
Block stands out by publishing measurable outcomes. By mid-April, each developer was landing production code changes at 2.5 times the January rate. Post-release failures fell 70% versus the same period in 2025. First-quarter gross profit reached $2.91 billion, up 27%. The company raised its target adjusted operating margin to 26% from 20%.
In March, the company quietly rehired some of those laid off — several former employees were invited back to their previous roles amid staffing gaps for key infrastructure.
In early August, Block beat analyst expectations for the second quarter and raised its 2026 guidance. The shares still fell 5% as investors questioned returns on AI investments.
Source: ForkLog, SiliconANGLE, Block.
Thus, the February share-price increase came before fresh results; the August decline came after.
Gates describes a mechanism that makes layoffs almost inevitable. A company that adopts the technology saves on staff and cuts prices. Competitors respond to lost market share by doing the same. If major players hesitate, startups built without “extra” employees take their place.
Companies are adopting AI faster regardless of whether returns are proven. The market reacts to a layoff announcement the same day, while the first reported numbers arrive a quarter or two later — and in the meantime, the remaining staff do the work of those who were let go.
The reverse centaur
In May 1997, Deep Blue defeated Garry Kasparov. A year later, the grandmaster proposed “advanced chess”: a person plays in tandem with a computer against the same kind of duo. The tandem beat both top programs and the strongest solo players. The term “centaur” was born — the human remains the “head,” the machine the “body.”
Doctorow flips the figure. In the book The Reverse Centaur’s Guide to Life After AI, he describes a model where the algorithm makes decisions and the employee does the physical work “at an inhuman, machine pace.”
He draws examples from today’s economy:
an Amazon delivery van driver works under recognition cameras — the system logs whether he looks at his phone or even sings while driving, while an algorithm sets routes and time standards;
a programmer with an assistant like GitHub Copilot turns from author into a controller of machine-written code;
a lawyer or doctor signs a document prepared by a model and is liable for someone else’s error.
The employer’s math goes like this: ten editorial staff are replaced by three equipped with AI. The previous workload remains, plus checking machine drafts.
Source: The Reverse Centaur’s Guide to Life After AI.
This is where the math breaks: checking each result costs almost as much as creating it. The employer’s savings arise only because the time spent on checking is not paid.
Top managers who decide to adopt AI do not become reverse centaurs: they are higher in the hierarchy, and for them AI is a tool to manage people.
A split appears in almost any company: enthusiasm at the top, resistance at the bottom.
The difference between a centaur and its inversion is not the technology’s quality but the right to choose. Doctorow explains it with a warehouse example: a forklift is useful in itself, but harm begins where the employee no longer decides when and how to use it.
This also explains polarized views of AI. Proponents usually choose when to invoke a model. Opponents use it under duress.
Zuckerberg promises a superintelligence for every person. But getting a tool and controlling it are different things, and the latter is not discussed in his essay.
Checking others’ output looks like a compromise: the job remains, only the content of the work changes. But even this role is time-limited.
A closing window
The Occupational Vulnerability Index offers a surprising result: the biggest productivity gains from AI are in the very occupations most at risk of being cut.
Writers and authors — 57% exposure. Programmers and digital interface designers — 55% each. Roofers, orderlies, and dishwashers — under 1%. At risk are 9.3 million jobs and $757 billion in annual income for U.S. households, distributed extremely unevenly across the country.
In the new reality, a worker’s position falls into three categories:
elimination — new hires in such roles are rare;
checking — the position remains, but the author becomes a controller;
augmentation — the employee decides when to use an AI tool.
Payroll data back this taxonomy. In August, Stanford Digital Economy Lab updated its study “Canaries in the Coal Mine?” using current data from ADP.
There is no economy-wide mass substitution. But employment among 22–25-year-olds in occupations with high exposure to AI optimization is 19% below peers in less affected fields — up from 15% a year earlier.
Experienced workers do not show a comparable lag. Employment is falling not because of layoffs but because hiring slows. This occurs mainly where work relies on codified knowledge, not on experience and intuition that cannot be reduced to fixed rules.
Where AI complements people, there is no decline in youth employment.
Checking machine output is the interim state between augmentation and replacement. Gates describes the horizon plainly: a serious shift will come when AI delivers virtually error-free results. From that point, the system can run without a human overseer, and companies will have strong economic incentives to remove the checker.
Simply put, a checker is needed only so long as the error rate justifies their salary. Eight errors per hundred can warrant the cost. One per ten thousand does not.
A “reserve,” a tax, and a dividend
Gates’s father died of Alzheimer’s disease in 2020. Professional caregivers looked after him around the clock, anticipating his needs from various, implicit cues.
From this came the Human Reserved idea — a list of activities reserved for people. A machine can do the task, but society refuses to hand it over. The analogy is a nature reserve: land could be developed, but society declines because losses exceed gains.
An example from the essay is a robot telling a patient about a terminal illness. Technically feasible now, and precisely the sort of task Gates suggests isolating from machines.
Alongside the ethical case, Gates makes a more prosaic economic one. When automation displaces a lifelong construction worker, he is sent to retrain. But it is impractical to offer a 55-year-old bricklayer a nursing-home aide role in hopes he will find it fulfilling. The reserve should cover professions from which there is nowhere comparable to go: adjacent vacancies matching the person’s qualifications no longer exist.
Labor-market data partly support this approach. Care aides, orderlies, and home health aides are among the least exposed to AI: the work requires physical presence. Youth employment in these roles, according to ADP data, is rising, not falling.
Obviously, each country would draw the reserve differently. Japan lacks young workers to care for the elderly, so a robot caregiver will appear there sooner than in countries with a surplus of available labor.
The second thesis is tested by arithmetic. Employers pay contributions on wages into the public budget. The purchase of a robot is expensed immediately and reduces the tax base.
Gates sees a distortion: the tax system itself nudges employers to replace people with machines. He proposes fees on AI tokens and robotic equipment to smooth the imbalance and fund retraining and social guarantees.
He made the same point in 2017, noting that a factory worker doing $50,000 worth of work pays income and payroll taxes, while a robot doing the same job is taxed nothing. The idea was considered odd at the time.
The scale of the potential fiscal gap is illustrated by a calculation from the AI Futures Project, a nonprofit founded by former OpenAI researcher Daniel Kokotajlo.
In the AI 2040: Plan A scenario, U.S. employment after 2032 collapses to 12% by the end of the decade. By 2033, companies continuously run 60 million AI agents operating at twenty times human speed. The document is less a forecast than a set of recommendations on how Washington and Beijing could slow the race and delay superintelligence until 2040.
To compensate for lost earnings, the authors propose a “citizen dividend” — distributing most proceeds from compute and robot permits to all adult Americans.
No jurisdiction has adopted either a tax on AI tokens or a list of reserved occupations. Gates himself lists unanswered questions:
who decides what to reserve for people;
by what criteria to select professions;
how to stop companies from circumventing the rules;
what to do about international trade if one country lets robots produce a good and another bans it.
All this demands institutions that don’t yet exist. Some mechanisms, however, already work.
Without a regulator’s permission
Germany has offered reskilling subsidies since 2019. Implementing the program required neither a new agency nor an international agreement.
Paragraph 82 of the Social Code requires the Federal Employment Agency to pay for retraining of people already in work. The subsidy depends on company size: firms with fewer than 50 employees are reimbursed up to 100% of course costs, companies with 50 to 500 workers receive half, and large employers a quarter. In addition, the agency reimburses part of the wages the employer continues to pay during training.
The eligibility rules are strict. The employer applies before the course begins, the program must exceed 120 hours, and it must carry state accreditation. Since January 2026, top-ups are paid only where training truly pulls the person away from the workplace. Evening courses and self-study are no longer subsidized.
The mechanism solves a single task — it pays for a career change. It does not regulate work tempo or when to invoke an algorithm.
The Stanford Digital Economy Lab study finds employment holds where technology complements tasks rather than replaces them. Doctorow explains the same pattern differently — through the worker’s right to choose when to apply the tool.
An employee’s value grows not with the volume of generated text but with the ability to tell effective model use from ineffective. Doctorow calls this skill discernment. It is getting harder: new tools are moving out of the user’s line of sight.
AI agents execute chains of actions without showing intermediate steps. Interfaces through which the process could be monitored are gradually disappearing. No rules currently require developers to disclose intermediate steps.
The debate over AI’s future focuses mostly on access. Zuckerberg promises to hand superintelligence to everyone; Gates would tax it and leave some activities to people. Neither approach explains who sets the pace of work and decides when to switch on the model.
The answer is in plain sight. The authors of both essays face no scenario in which they risk becoming “a human appendage to a machine.” Those who turn into reverse centaurs are lower in the corporate hierarchy.
A human checker is needed only while the model errs, and Gates acknowledges this. Yet companies are still reorganizing headcount as if the “gatekeeper” will remain forever.
Billionaires are arguing over who will get the technology. For now, a different question is being decided — who will submit to it.
Follow ForkLog on social media
Found a mistake in the text? Select it and press CTRL+ENTER