The premise is straightforward enough to fit in a single sentence, and unsettling enough to require several paragraphs to absorb. Elon Musk's artificial intelligence company, xAI, is developing systems it calls "human emulators" — AI agents designed to perform white-collar work by mimicking exactly what a human does at a computer: moving a cursor, clicking buttons, reading screens, typing responses, navigating software. The agents do not use APIs or custom integrations. They use the computer the way you do. And inside xAI's own offices, employees have sometimes been unable to tell that the colleague responding to their Slack message or completing a task in their workflow was not a person. [1][2]

The disclosure came not through a press release but through an accident of candor. In January, Sulaiman Khan Ghori, a technical staff member at xAI, appeared on a podcast called Relentless and described the company's "Macrohard" project in detail that his employer evidently considered excessive. Ghori explained that xAI was building infrastructure to support "one million human emulators" — AI systems running on virtual machines, each capable of performing any digital task a human worker performs, at a cost he estimated at one to ten cents per hour. He was subsequently fired. [1][3]

What the Emulators Do

The Macrohard project — the name is a deliberate jab at Microsoft — takes a different approach from the AI copilot products that companies like Microsoft, Google, and Salesforce have been selling to enterprises. Those products sit alongside human workers, suggesting code completions, drafting emails, summarizing meetings. xAI's human emulators do not assist. They replace. The system observes how a human performs a task on a screen, learns the sequence of inputs, and replicates it autonomously. [2][3]

The scope is intentionally broad. Any work that involves a keyboard, a mouse, and a screen is within the system's target domain: data entry, customer support, content moderation, scheduling, report generation, basic analysis. The emulators do not need to understand the work in the way a human does. They need to perform the observable actions that produce the same output. The distinction between understanding and replication is, from the perspective of the employer paying for the output, irrelevant. [2]

The Internal Test

The most revealing detail from Ghori's disclosure was not the technology but the deployment. xAI had been testing the emulators internally — placing them into actual workflows alongside human employees without always disclosing their nature. Colleagues interacted with the emulators and, in at least some cases, did not realize they were not interacting with a person. The Information confirmed this characterization in its reporting. [1][2]

This is not a Turing test in the philosophical sense. The emulators are not pretending to be conscious. They are performing bounded tasks within structured software environments where the range of expected behaviors is narrow enough that competent execution is indistinguishable from human execution. A customer support response that resolves the ticket is a resolved ticket, regardless of who — or what — typed the reply.

The Cost Structure

Ghori's cost estimate — one to ten cents per hour — deserves scrutiny. The figure presumably reflects the marginal computational cost of running a virtual machine with the emulator software, not the development cost of building the system. Even if the all-in cost is ten times higher, it remains a fraction of the cost of a human worker performing the same task. A customer support representative in the United States earning $18 per hour costs an employer approximately $25 per hour after benefits and overhead. An emulator at one dollar per hour — ten times Ghori's upper estimate — represents a 96 percent cost reduction. [3]

Brownstone Research, which covers AI enterprise applications, published an analysis of the Macrohard project noting that xAI's ambition of one million human emulators would, if achieved, represent the equivalent of a major corporation's entire workforce running on infrastructure that fits in a data center. The labor cost savings at scale would be measured not in millions but in billions annually. [3]

The Companion Story

The timing is not coincidental. As this paper reports today, 58,121 tech workers have been laid off in 2026, with one in five cuts explicitly blamed on AI. The layoffs represent companies eliminating positions that AI can now perform. xAI's human emulators represent the system those positions are being eliminated for. The supply side and the demand side of the AI labor displacement story are now both visible, and they point in the same direction.

Musk has not commented publicly on the Macrohard project since Ghori's departure. xAI's official communications have focused on Grok, the company's chatbot, and on its data center expansion. The human emulator program operates, for now, in the space between disclosure and acknowledgment — confirmed by reporting, visible in leaked details, but not yet part of xAI's public narrative.

The emulators, by all accounts, continue to improve. Their coworkers, by all accounts, continue not to notice.

-- DAVID CHEN, Beijing