In August 2026, three influential pieces emerged discussing a future dominated by artificial intelligence.

At the beginning of the month, Canadian author Cory Doctorow published an analysis based on the observations of a consultant who spent 18 months examining corporate AI projects, finding no viable use cases. Nine days later, Mark Zuckerberg promised

Despite the three differing visions presented, the authors agree on one point: technology is already delivering significant returns. Let's delve into the basis for this confidence, who is financing the radical restructuring of work processes, and why roles for writers, programmers, and designers are increasingly shifting from creators to overseers.

A Debate Without Common Ground

Zuckerberg's essay, "The Future is for Everyone," is built on three principles:

  • Empowering individuals as a source of wealth;
  • Innovation, rather than mere automation, as the goal of superintelligence;
  • A balance of power between individuals and institutions as the foundation of security.

The Meta CEO does not believe that security is achieved through the concentration of power.

His company, valued at $1.47 trillion, proposes:

  • A personal agent that works for the user around the clock;
  • A fully private mode where user data is inaccessible to the company (similar to WhatsApp encryption);
  • An auction mechanism for computing resources for those willing to pay for greater power.

The company also vows to resume publishing open-source models.

In contrast, Gates adopts a different tone, calling the transition to a new era one of the most turbulent periods in human history and noting that there is currently no preparation for it.

He identifies three risks:

  1. Jobs may vanish permanently.
  2. AI could enhance the potential for "causing harm."
  3. Machines may begin to replace human relationships.

His three proposals are:

  1. New institutions to manage the transition.
  2. A list of occupations that society agrees not to automate.
  3. A tax on AI tokens and robots.

Gates observes that it took twenty years for personal computers to reformat office work: initially, programmers had to create business software, then prices for the hardware dropped, followed by a lengthy retraining period for employees.

With language models, this lengthy process is unnecessary. Large Language Models (LLMs) can be utilized on existing devices and are accessible through natural language without the need for specialized interfaces. The technology adapts rather than the user; it can analyze training videos for new employees and acquire their skills.

Gates anticipates that AI will alter the balance of power in legal, customer support, medical, development, and manufacturing fields, with widespread transformation occurring in less than a decade.

Source: Meta, Gates Notes.

None of these essays scrutinize the assumption of technology's effectiveness; it is taken for granted and underpins all subsequent arguments about taxes, "reserved" professions, and the distribution of power over technology among corporations, governments, and individuals.

This verification was undertaken by a consultant whose firm spent 18 months working with optimistic and AI-friendly market participants.

Faith Over Metrics

Nikhil Suresh, who oversees sales and technical implementation at Hermit Tech, engaged in around 300 conversations with professionals worldwide—from niche specialists to executives from Fortune 500 companies. The results were detailed in his essay AI Mania Is Eviscerating Global Decision-Making.

In one instance, a top manager from a company with over $2 billion in revenue presented a technology strategy entirely centered around AI. It soon became apparent that this "innovator" had never used ChatGPT or any other AI tool. Suresh declines to name this individual, citing potential job loss due to such admissions.

Claims from Hermit Tech clients about a hundredfold increase in productivity have become commonplace. However, over 18 months, Suresh's team recorded a zero success rate for AI projects. Failures occurred not only in projects they were invited to but also in those they observed from a distance.

The gap between promise and reality is illustrated by statistics. Cortex, under ideal conditions, delivers accurate responses about 92% of the time—this was the assessment made by Snowflake employees during a presentation. Thus, at least eight out of every hundred responses are misleading, and someone must identify each error before it reaches a client or is included in a report.

Moreover, the effectiveness of human capital is gauged by the amounts spent on AI. Suresh labels this metric as "completely manipulable"—anyone can inflate it. In places where "token leaderboards" are implemented, the volume of generated text is reported.

One employee described their strategy:

  • They duplicate the work repository;
  • Task the model with rewriting code from one programming language to another;
  • They focus on the actual task simultaneously.

The duplicate project is unnecessary, yet token consumption rises, and the job is preserved. Why such schemes do not surface at the management level is the central question of the essay.

Suresh describes the situation as a coordination problem. Those who publicly affirm productivity growth retain their positions. Acknowledging the opposite would sound like an accusation of colleagues' dishonesty or incompetence—such a manager would be fired, and their replacement would likely continue the "party line."

Doctorow illustrated this in a column for The Nerve. In organizations with over 500 employees, promotions and job security were granted to those who claimed the transformative power of technology. Employees with legitimate objections were placed on layoff lists.

Readers of the blog with the title Head of AI at companies earning over $1 billion annually told Suresh that they cling to their fictitious positions solely for career advancement.

At one point, Hermit Tech's team stopped asking clients about their current AI initiatives. Once a project begins, discussing its progress honestly becomes impossible until it reaches a crisis point.

The sample consisted of the firm's clients, external projects monitored, and discussions outside of work tasks. Respondents were anonymized, and data collection methods were not published, nor was any independent verification conducted. Additionally, Hermit Tech has ceased all AI implementation work—this increases the observer's independence but also limits access to successful examples.

Job Cuts Under the AI Banner

In early 2026, Jack Dorsey announced layoffs of nearly 4,000 employees at Block—reducing the workforce from around 10,000 to under 6,000. Following the main session on February 26, the company's stock rose by about 20%.

About a month and a half later, Snap made a similar move. CEO Evan Spiegel cited advancements in AI as the reason for the restructuring, which affected around 1,000 employees—16% of the workforce. On April 15, the company's shares reacted with a 9% increase.

Shortly before Dorsey's announcement, OpenAI CEO Sam Altman gave an interview to CNBC, stating that large companies use artificial intelligence as a convenient explanation for layoffs that would have occurred regardless.

Goldman Sachs assessed the impact of AI implementation. According to the bank's calculations, the technology reduced the monthly job growth in the U.S. by approximately 16,000 positions and raised unemployment by 0.1 percentage points.

Block stands out for publishing measurable results of its implementation. By mid-April, each developer was making 2.5 times more changes to production code than in January. Failures following such updates decreased by 70% compared to the same period in 2025. The gross profit for the first quarter reached $2.91 billion, a 27% increase. The company raised its targeted adjusted operating margin to 26% from the previous 20%.

In March, the company quietly rehired some of the laid-off employees—several former staff members received offers to return due to a shortage of personnel for critical infrastructure.

In early August, Block exceeded analyst expectations for the second quarter and raised its forecast for 2026. However, shares decreased by 5% as investors remained skeptical about the returns from AI investments.

Source: ForkLog, SiliconANGLE, Block.

This indicates that the February stock price increase occurred before the release of fresh financial reports, while the August decline happened afterward.

Gates describes a mechanism that makes layoffs almost inevitable. A company that adopts technology saves on personnel costs and reduces product prices. In response to market share loss, competitors take similar actions. If major players hesitate, startups built without "excess" staff quickly fill the gap.

Companies are increasingly adopting AI regardless of whether returns are confirmed. The stock market reacts to layoff announcements the same day, while the first financial figures appear only after a quarter or two, during which time the tasks of laid-off employees are handled by those who remain employed.

The Reverse Centaur

In May 1997, Deep Blue defeated Garry Kasparov. A year later, the grandmaster proposed a format called "advanced chess": a human plays in tandem with a computer against another duo. It turned out that this partnership outperforms both the best programs and the strongest individual players. This led to the term "centaur"—the human remains the "head," while the machine acts as the "body."

Doctorow flips this concept. In his book, The Reverse Centaur’s Guide to Life After AI, he describes a model where decisions are made by an algorithm, and the physical work is done by an employee at a "superhuman, machine-like pace."

Examples from the modern economy include:

  • An Amazon delivery driver works under surveillance cameras that track whether they are distracted by their phone or singing while driving, while the algorithm assigns routes and time standards;
  • A programmer using an assistant like GitHub Copilot shifts from author to overseer of machine-generated code;
  • A lawyer or doctor signs off on a document prepared by a model, taking responsibility for someone else's mistake.

Employers see the benefit as follows: ten editorial staff members are replaced by three equipped with AI. The former workload remains, with the added task of reviewing machine drafts.

Source: The Reverse Centaur’s Guide to Life After AI.

Here lies the flaw: reviewing each result costs almost as much as creating it. Employers only benefit from savings because no one pays for the hours spent on this task.

Top executives making implementation decisions do not become reverse centaurs; they occupy higher positions in the hierarchy, and AI serves as a tool for managing people.

This creates a divergence evident in almost every company: enthusiasm at the upper levels, resistance at the lower.

The distinction between the centaur and its inverse is not based on the quality of technology but rather on the right to choose. Doctorow illustrates this using a warehouse example: a forklift is useful on its own, but problems arise when an employee no longer decides when and how to use the device.

This explains the polarized views on AI. Proponents of the technology typically choose when to engage with the model, while opponents are compelled to use it.

Zuckerberg promises superintelligence to everyone on the planet. However, obtaining a tool and having the freedom to use it as one wishes are two different matters, and the latter is not addressed in his essay.

Checking someone else's output may seem like a compromise: the job is preserved, but the nature of the work changes. Yet, this role is also expected to be temporary.

A Closing Window

The occupational vulnerability index yields an unexpected result: the greatest productivity gains from AI are likely to come from jobs that are also at the highest risk of significant reductions.

Writers and authors face a 57% vulnerability. Programmers and digital interface designers are at 55%. Roofers, orderlies, and dishwashers face less than 1%. At risk are 9.3 million jobs and $757 billion in annual income for American households, distributed unevenly across the country.

In this new reality, three categories of workers can be identified:

  • Disappearance—new hires for such positions are almost non-existent;
  • Verification—jobs remain, but the creator becomes an overseer;
  • Empowerment—the worker decides when to utilize AI tools.

Data regarding salary allocations supports this classification. The Stanford Digital Economy Lab updated its research Canaries in the Coal Mine? in August, utilizing current ADP data.

There is no mass substitution across the economy as a whole. However, the employment rate for young professionals aged 22–25 in AI-optimized occupations is 19% lower than that of their peers in less affected jobs—a gap that has widened from 15% over the past year.

Experienced workers do not exhibit such discrepancies. Employment declines are not due to layoffs but rather to a slowdown in hiring. This primarily occurs in areas that rely on codified knowledge rather than experience and intuition, which cannot be reduced to a set of rules.

In sectors where AI complements human effort, no employment declines among youth are observed.

Reviewing machine-generated results represents an intermediate state between augmentation and replacement. How long this phase will last is described plainly by Gates: a significant shift will occur when AI begins to produce nearly flawless results. At that point, systems could operate without human oversight, and companies would have strong economic incentives to remove overseers from the process.

In simple terms, an overseer is only needed until the error rate justifies their salary. Eight errors per hundred make those costs worthwhile. One in ten thousand does not.

"Reservoir," Tax, and Dividend

Gates' father passed away from Alzheimer's disease in 2020. Professional caregivers provided round-the-clock care, anticipating his needs based on various subtle cues.

This observation led to the idea of a Human Reserved list—activities reserved for humans. While machines can perform tasks, society refuses to assign them to machines. The analogy is a natural reserve: land can be developed, but society rejects such plans because the losses outweigh the benefits.

An example from the essay is a robot informing a patient of an incurable illness. This is technically feasible today, and Gates suggests isolating such tasks from machines.

Alongside the ethical argument, Gates presents a more practical economic rationale. When automation displaces a construction worker who has spent their entire life in the industry, they will be retrained. However, it is impractical to suggest that a 55-year-old mason should take a position as a caregiver in a nursing home, hoping that this new job will appeal to them. Reserving certain professions makes sense for those who have no viable transition options: there are no related vacancies available for their qualifications.

The labor market partially validates the rationality of this approach. Caregivers, orderlies, and home health aides are the least affected by AI: their work requires physical presence. Employment among youth in these professions, according to ADP data, is not declining but rather increasing.

It is evident that the composition of such a reserve will differ in each country. Japan lacks young workers to care for the elderly, so robot caregivers will emerge there sooner than in countries with a surplus of available hands.

The second thesis can be validated mathematically. Employers contribute to the budget from employee earnings. When a company purchases a robot, it immediately counts that expense against its taxable income.

Gates sees this as a distortion: the fiscal system itself encourages employers to replace humans with machines. To rectify this, he proposes a tax on AI tokens and robotic equipment—this could help to balance disparities and fund retraining and social guarantees.

Gates had previously articulated a similar idea in 2017, noting that a worker performing a $50,000 job in a factory pays income tax and social contributions, while a robot in the same position incurs no tax. At the time, this suggestion was considered peculiar.

The potential scale of future budget deficits is illustrated by calculations from the AI Futures Project, a non-profit organization founded by former OpenAI researcher Daniel Kokotailo.

In the scenario AI 2040: Plan A, the employment level in the U.S. plummets to 12% by the end of the decade after 2032. By 2033, 60 million AI agents will be operating continuously in companies at speeds 20 times greater than humans. This document serves not just as a prediction but as a recommendation for how Washington and Beijing could slow down the race and delay the emergence of superintelligence until 2040.

To compensate for lost earnings, the authors propose a "citizen's dividend"—distributing a significant portion of revenues from computational resources and robot licenses to all adult Americans.

Neither the tax on AI tokens nor the list of reserved occupations has been adopted by any jurisdiction. Gates himself lists questions for which he currently has no answers:

  • Who decides what to reserve for humans;
  • By what criteria to select professions;
  • How to prevent companies from circumventing regulations;
  • What to do about international trade if one country allows robots to produce goods while another prohibits it.

All of this requires institutions that currently do not exist. However, some mechanisms are already functioning.

Without Regulatory Approval

In Germany, a retraining subsidy program has been in place since 2019. Its implementation did not require a new agency or international agreement.

Paragraph 82 of the Social Code mandates that the Federal Employment Agency cover the costs of retraining for currently employed individuals. The subsidy amount depends on the size of the enterprise: firms with fewer than 50 employees are reimbursed up to 100% of course costs, companies with 50 to 500 employees receive half, and large employers receive a quarter. Additionally, the agency compensates part of the salary that the employer continues to pay during training.

The conditions for receiving social assistance are strict. The employer must submit the application before training begins, the program must exceed 120 hours, and it must be state-accredited. Starting January 2026, supplementary payments will only apply where training genuinely removes a person from their job. Evening courses and self-study will no longer be subsidized.

This mechanism addresses only one issue—financing a career change. The pace of work and the timing of algorithm engagement remain unregulated.

The Stanford Digital Economy Lab's research indicates that employment holds steady where technology complements tasks rather than substitutes them. Doctorow explains the same pattern differently—through the worker's right to choose when to apply the tool.

The value of an employee increases not with the volume of generated text but with the ability to distinguish effective uses of the model from unsuccessful ones. Doctorow refers to this skill as discernment. However, discerning has become increasingly challenging: new tools often operate outside the user's awareness.

AI agents execute action sequences without displaying intermediate steps. Interfaces that once allowed tracking of the process are gradually disappearing. Regulations that would require developers to disclose intermediate steps have yet to be established.

The debate over the future of AI largely revolves around access to the technology. Zuckerberg promises to distribute superintelligence to all, while Gates proposes taxing it and reserving some roles for people. Yet neither approach clarifies who sets the pace of work and decides when to engage the model.

The answer is clear. The authors of both essays face no risk of being relegated to the role of "human appendages for machines." It is those lower in the corporate hierarchy who become reverse centaurs.

An employee who checks the model's output is needed only as long as the model makes mistakes, and Gates acknowledges this. Nevertheless, companies continue to restructure their workforces as if the "gatekeeper" will remain indefinitely.

Billionaires are debating who will gain access to the technology. But the more pressing issue is who will be subservient to it.