Ghost Work

Gray, an anthropologist, and Suri, a computer scientist, document what they call the paradox of automation's last mile: every automated system leaves tasks that require human judgement, and those tasks are contracted out at piece rates to workers who are deliberately invisible.

The workers are not a transitional stage. The authors argue the arrangement is structural: as automation advances it generates new last-mile work rather than eliminating it.

Their finding is that automation does not shrink the human remainder. It keeps relocating it.

Devmark’s reading — not a quotation from the book

The research covers workers in the United States and India across platforms including Amazon Mechanical Turk, and combines ethnography with platform data.

What emerges is specific: unpredictable income, no recourse against rejected work, accounts deactivated without explanation, and informal worker forums that do the support work the platforms do not.

The final section argues for treating this as employment with obligations rather than as a marketplace of tasks, and sets out concrete proposals including portable benefits and a right to appeal.

The book is from 2019 and the category has grown enormously since, including the annotation labour behind current language models. It reads now as an early and accurate account of an industry that has become much larger.

Written in 2019, and it describes the labour arrangement that current AI systems were built on.

Devmark’s reading — not a quotation from the book

The authors' central concept is a paradox: every advance in automation creates new human tasks at its edges — labelling, verifying, correcting, handling the cases the system cannot. The demand is real and continuous, but because it is framed as temporary scaffolding for a system that will soon be fully automated, it is never treated as a job.

The work is permanent. It is only ever described as temporary.

Devmark’s reading — not a quotation from the book

Tasks appear on a platform, are claimed, are completed in minutes, and are paid per unit. The authors document the consequences carefully: substantial unpaid time spent searching for work worth taking, no visibility of the requester, and no way to build a relationship or a reputation that transfers between platforms.

The most damaging asymmetry they document is the requester's unilateral ability to reject completed work without explanation and without payment, while retaining the output. The worker has no appeal, and a rejection rate affects access to future tasks. It is the mechanism that most clearly marks this as something other than a market between equals.

Against that, the fieldwork records extensive informal organisation: forums, messaging groups and shared tools where workers warn each other about bad requesters, circulate information about which tasks pay, and help newcomers. The authors treat this as evidence of an occupational community forming without any of the structures that normally produce one.

Losing an account — through an automated decision, an error, or no stated reason — removes a person's entire income with no notice and no process. The authors return to this because it is where the absence of employment status has the sharpest consequence, and because the platforms' scale makes individual review uneconomic by design.

The fieldwork spans the United States and India, and the comparison is structural rather than decorative: the same platforms, the same tasks and the same rating systems produce very different meanings depending on local labour markets, what the income is supplementing, and who in a household is doing it. It is the strongest argument in the book against treating platform work as one phenomenon.

Read it as the empirical answer to a question the industry prefers rhetorical: when a product says it is automated, who is actually doing the part that is not. The fieldwork is specific, the workers are quoted at length, and the scale is larger than almost any reader expects.