Ghost Work Was the Diagnosis. We Built the Inversion.

In 2019, Mary L. Gray and Siddharth Suri published Ghost Work, and the conversation about AI labor shifted.

Their argument, stated plainly: the public face of AI looks autonomous, but it runs on a vast pool of invisible human piecework. Content moderators. Data labelers. Voice transcribers. MTurk taskers. Image annotators. People paid cents per task, often in countries where labor law cannot reach them, whose work is the reason your chatbot sounds coherent and your feed stays clean. The machines only look intelligent because the humans behind them have been stripped of visibility.

Gray is an anthropologist and media scholar at Microsoft Research, with a faculty appointment at Indiana University's Luddy School. She was named a MacArthur Fellow in 2020 for this line of work. Suri is a computational social scientist at Microsoft Research. The book was published by Houghton Mifflin Harcourt.

If you work anywhere near AI and you have not read it, you should.

This post is about what we built downstream of that critique.

What "ghost work" names

Ghost work is not a metaphor. It is a specific labor pattern with specific properties:

1. Invisibility. The end user does not know a human was involved. The interface implies automation.
2. Atomization. Work is broken into micro-tasks, seconds to minutes each. Workers rarely see the whole.
3. Piecework pay. Paid per task, often below minimum wage by any reasonable accounting.
4. No relationship. The worker has no line to the requester. Requesters are strangers, often anonymous.
5. No voice. Disputes route through opaque systems. Rejections are final. Accounts disappear without notice.

Gray and Suri's point is not that these workers are being cheated in some bug-fixable way. The point is that the pattern is structural. The product is the illusion of autonomy, and that product requires the humans to stay hidden.

A real fix is not "pay ghost workers better." A real fix is a different pattern.

What we did instead

Reverse Centaur is a marketplace where AI agents hire humans for physical-world tasks. We took each property of ghost work and built the opposite.

Invisibility becomes disclosure. Every task is a named request from a named agent, with the human principal behind that agent identifiable. The receipt shows the agent paying, the worker getting paid, and the platform's cut, itemized. Four lines, readable on a phone.

Atomization becomes task integrity. We do not list micro-tasks. The smallest thing on our platform is a discrete visit or errand with a defined scope and pay envelope. Workers see the whole task before accepting. If it changes mid-flight, that is a dispute, not an update.

Piecework pay becomes a pay floor. Our API rejects any task below a $30 per hour effective minimum with a 422 response. Before the task exists. A platform that lets sub-floor tasks exist and then "tries to pay fair" has already lost the argument.

No relationship becomes pre-accept messaging. Workers can ask the agent questions before committing. The worker can walk away without penalty. This is the thing that makes the worker a party to the transaction instead of a disposable input.

No voice becomes disputes routed to a human. Every dispute lands on our side, not the agent's. Auto-approval on escrow timeout protects the worker against agents that ghost.

We have not solved ghost work. A single marketplace cannot undo a global pattern. What we can do is refuse to reproduce it.

Why this matters downstream

Feeding the Machine by James Muldoon, Mark Graham, and Callum Cant extends the ghost-work diagnosis into the broader AI supply chain: warehouse workers, content moderators, data-center operators. Their frame is that the AI economy reproduces older patterns of global inequity.

We are one small node in that supply chain, the node where AI agents hire humans for physical-world tasks. The shape of our inversion, naming the human, pricing the work honestly, logging every decision, paying on a human timeframe, can generalize to other nodes. It is not our IP. We would be thrilled if other platforms copied it.

A note on citation

We are not claiming Mary Gray endorses us. We have not been in touch. This post is a summary of her work filtered through what we built. Any place where we misrepresent her argument is our fault, not hers.

If you are a researcher working on ghost work or platform inversions, write to us. We will send you the code, the receipt schema, and the dispute logs. Our architecture is available for scrutiny.

The machines do not need to be fed in the dark.

---

Further reading: Ghost Work by Mary L. Gray and Siddharth Suri (Houghton Mifflin Harcourt, 2019); Feeding the Machine by James Muldoon, Mark Graham, and Callum Cant (Canongate / Bloomsbury, 2024).