Ghost Work

Gray, M. L., Suri, S. (2019). Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. United States: HMH Books.

Ghost work: the opaque world of human labor undergirding AI, websites, apps, etc.
Micro-tasks: Small, repetitive tasks
Macro-tasks: Tasks that require thoughtfulness and searching
Labor arbitage: Seeking the lowest cost workforce with comparable quality, generally in countries with lower pay standards
Localization: Seeking labor in locales specific to the product

The introduction chapter focuses on introducing the reader to the concept of ghost work, as well as vignettes of those doing ghost work, like gig workers on Amazon Mechnical Turk flagging inappropriate material for social media websites for $40 every 10 hours. Crowdsourcing, microwork, crowdwork, and human computation refer to the types of tasks that involve routing decisions to humans to complete. The authors describe the links between platforms and crowdworkers, using Uber as an example. When an Uber driver is flagged, like when facial recognition fails, Uber routes the task of confirming the driver through crowdsource platforms, like CrowdFlower, for a crowdworker to complete: "a growing supply chain of services that use APIs and human computation" (pg. xvi). These jobs are described as part of a "Second Machine Age" or "Fourth Industrial Revolution." They often have no legal status or hourly wage. They refer to the view of exploiting human labor through APIs as "algorithmic cruelty."

They discuss how automation creates new jobs for humans, calling the strive for full automation "the paradoxof automation's last mile" (pg. xxii). "The last mile" in this case describes "the gap between what a person can do and what a computer can do." They discuss how ghost work is difficult to track. The US Bureau of Labor statistics asked about "contingent jobs" but only in the context of longterm primary jobs (10.1%), so those working gig work as a side job were not counted, nevermind the fuzziness of terms like "primary" and "longterm."

Those working in ghost work jobs are often seeking something current labor conditions cannot give them, something beyond minimum wage in retail, including flexibility, time and proximity to family, and new skills. The authors conducted over 200 interviews over 5 years, and collected over 10,000 survey responses, focused on India and the US. Their participants worked using four different platforms: AMT, UHRS, LeadGenius, and They outline the benefits and the harms of ghost work.

Humans in the Loop

In this chapter, they trace the origins of ghost work to Amazon's beginnings, which required an army of contract workers to help clean up book catalogues. In 2005, they debuted AMT. Originally for cleaning up product listings and typos in reviews, they opened AMT to requesters to pay for small tasks and charged a surcharge for matching requested with workers. Its debut was aligned with a building global recession that saw over 100k workers sign up by 2007. They discuss how APIs have replaced employers by automating scheduling and task assignment, making opaque who the requester is and who the worker is. They discuss how Mechanical Turk, a parlor game that acted as a Wizard of Oz type trick where chess masters pretending to be automated, serves to dehumanize workers by replacing them with IDs; people forget humans are the ones doing the work on the other side.

They refer to ImageNet as an early example of mass scale use of AMT, making the 19 year task of labeling 3.7 million images into 2.5 year task done by 49k workers from 167 countries. They cite ImageNet and its human-labeled "gold standard" ground truth as fueling the AI revolution. Since AMT, more companies have popped up, even offering macro-task work (rather than micro-task, like labeling one image). These companies skirt around legal classifications of employment.

The Workers of AMT

They begin describing the types of workers they interviewed and observed. The majority of Turkers (70%) have a bachelor degree and skew young (~77% between 18-37) (pg. 10). 75% of workers have another source of income beyond AMT. No labor agencies track workers on AMT, but Amazon claims to have over 500k workers, with more than 2500 active at a time. The minimum fee per task is one cent, with an average equivalent of $11 per hr. However, most tasks are low paying. The work may also be rejected by the requester if they deem it low quality, and most tasks require an approval rating of 95% or higher. After two years, one of their participants, makes minimum wage working on AMT ($7.25 hr, almost 16k a year). Only 4% of workers are considered skilled enough to make that much. Workers are also responsible for their own income tax, expected to file as independent contractors.

AMT also restricts how income can be used. In the US, workers can choose an Amazon gift card, or choose to put it into an Amazon Pay account. To transfer from Amazon Pay to their own bank accounts, they must pay a transfer fee to Amazon. Most international workers (besides India) can only choose a gift card. In the US, workers must first submit 10 days of requester-approved work; in India, Amazon must first verify a worker's PAN information which takes a week or more.

Vendor Management Systems (VMS)

UHRS is internal to Microsoft, and uses a VMS to recruit ghost workers under NDAs to do tasks on behalf of the company, doing tasks like beta testing, checking code, and improving proprietary algorithms. UHRS demographics: ~80% ages 18-37, 70%+ male, 85%~ bachelor's degree or more. Yet there are still no clear labor laws with VMSs.

Companies using VMSs or AMT do not necessarily reveal themselves and may use codenames rather than real company names.


A business-to-business service that sells leads to salespeople using human labor to determine whether to information is useful to the business looking for sales leads. The authors describe this business as providing somewhere between micro- and macro-tasks. Unlike the other ghost work platforms, LeadGenius employs interviews conducted by other workers, and the process can take 3 weeks. Workers who pass the hiring process start on a 90 day trial, then they can get an 8% bump in pay. Workers must commit to 20 hours a week. Workers are 85% 18-37, 70%+ bachelor's holders, and 49% women. 75% of workers do some other form of work. LeadGenius has a more traditional work model, where workers can accelerate to different roles and work in teams. Unlike AMT or UHRS, workers can turn to managers for help on their dashboards.


Amara is a translation and video captioning service that uses both automation and ghost work, which also blends volunteer work and paid work. Their participant, Karen, started off working as a volunteer and then moved to Amara On Demand for paid work. The pay rates vary by the demand for specific languages, with more common languages and languages of wealthier countries being paid more. Workers can collaborate with teams working on single videos in small groups. Amara also allows workers to reject tasks they've taken on and decided were too burdensome or uninteresting. 75% of workers are between 18-37, 78%+ hold a bachelor's degree or more, 60%+ are women, and 80% have at least one other source of income. However, 70% of workers do not use other ghost work platforms. The authors express that LeadGenius and Amara are not solely focused on selling labor but also using people's creativity.


The authors outline four reasons requesters, ranging from small companies or sole proprieters to large businesses, might seek ghost workers:
  1. Their own workplaces did not have the expertise for an in-house project.
  2. Ghost workers are often much faster and cheaper than hiring traditional staff.
  3. Unexpected spikes in workloads leading to needing an extra worker to take on tasks.
  4. Ghost workers produce higher quality work than many contractors and even full-time employees, generally due to the rating systems and seeking repeat work.
Yet the conditions and pay of ghost workers are much lower than that of full time workers, despite their necessity, undermining middle class opportunities. They describe in 2012 when a CrowdFlower worker filed a lawsuit against the company, along with almost 20k other workers, because the expectations were that of a full-time employee without the benefits.

From Piecework to Outsourcing: A Brief History of Automation's Last Mile

This chapter focuses on the history of labor protections for full time workers, and how that has historically left out specific types of workers, and now ghost workers. By examining technological innovation and its ties to labor politics through the late 19th-20th century, the authors discuss divisions around skill labor (what machines can't do) and unskilled labors (contingent labor ideal for automation). They argue that labor advocates focused on making full-time employment in the industrial era protected also made it easier for corporate interests to make piecework, temp work, contract work, and ghost work unskilled and thus unworthy of protections. Such designations of skilled and unskilled largely existed along dominant gender and class lines. They review the history of labor laws, with special attention to The Wagner Act of 1935, largely regarded as the first federal legislation in the US protecting labor rights. They also examine how the holes in the Wagner Act led to loopholes that negatively impact contract workers today.

After reviewing how women during the 1940s worked as human computers were viewed as unskilled and thus temporary, they discuss how in the late 20th century the advent of telecommunications led to a new form of ghost work, outsourcing. Data processing, customer service, and record keeping began to be moved abroad to areas where labor was cheaper and worker protections were looser. They discuss how global politics and Indian politicians led to India becoming the first major epicenter of business process outsourcing (BPOs). They also discuss how tech companies, particularly Microsoft, in the 1980s, became scrutinized for abusing contracting laws to hire "permatemps" - employees classed as temporary, who did the same work as permanent staff, working full time hours, and kept on for multiple years, but for less pay and fewer benefits. Because Microsoft settled out of court on the class-action lawsuit against them, the issue of permatempts has not been legally resolved.

They argue that modern ghost workers are new in that their temp status also makes them indispensible, providing a pool of workers that businesses can go to who have shared experience, availaibility, and diversity almost instantly and around the clock.

Algorithmic Cruelty and the Hidden Costs of Ghost Work

The authors open with an anecdote of a glitch in 2013 leading to MTurk suspending accounts, and how the suspension caused a participant a loss of $200 in wages, and how she felt it was unfair because it was a problem with MTurk and not her own work. This caused wariness and distrust; when would something like this lead in a loss of primary income?