In 2024, an Egyptian freelancer began comparing answers written by artificial intelligence. Each assignment contained a question and two responses, and the worker was asked to judge which response was more accurate, relevant, complete, or safe.
“We were asked to decide which response is more accurate,” they told The Cairo Report.
All workers who spoke to The Cairo Report for this article were granted anonymity because, like many in the AI industry, they had signed non-disclosure agreements restricting what they could say publicly about their work, including their working conditions and terms of employment.
The work came through Populii, a crowdsourcing platform owned by India’s technology services multinational Tech Mahindra, whose business is to organize people who collect, transcribe and evaluate data for companies developing AI.
The freelancer had started taking online work through small transcription teams and WhatsApp groups in 2021. Assignments arrived once or twice a week. Then, a year later, they moved onto Clickworker, a crowdsourcing platform through which they accessed Microsoft’s Universal Human Relevance System (UHRS), where they would help train the AI-powered Bing search engine.
This is the occupational adaptation Egypt’s government has been promoting for some time in its push to cultivate cheap outsourcing labor. In September, Ahmed El-Zaher, head of the Information Technology Industry Development Agency (ITIDA), described an outsourcing industry advancing from call centers toward engineering, research and AI. Its exports reached 5.2 billion US dollars (around 277.5 billion Egyptian pounds) in 2025/26, he said at the time. AI, the agency argued, would transform jobs and create opportunities in more sophisticated work.
Finance Minister Ahmed Kouchouk reinforced that message on November 16, albeit using a higher amount of dollar exports (7 billion dollars) in the outsourcing industry, telling students at a university, “I am proud of our creative young people, who can compete globally and export digital services”, and called for “greater support for technology businesses and entrepreneurship”, presenting the sector as an opportunity for further export growth.
But the freelancer’s experience shows what can happen within its less visible market for AI data: workers acquire new capabilities while employers retain the freedom to buy almost none of their time.
On UHRS, once their best source of assignments, the worker said they “can’t even make 10 dollars now.”
The contradiction taking place in the AI industry in Egypt is best understood inside production. AI developers need examples of language and behavior, judgments about correct answers, and people who resolve errors that automated systems can’t reliably settle. Data companies turn these activities into projects with instructions, payment units, and quality controls.
In all of these projects, workers obtain assignments from intermediaries organizing the work, rather than through lasting employment with the businesses using the resulting systems.
Their expertise can be highly specific too. For example, Perle, a San Francisco AI data company, recruits Saïdi Arabic transcribers who understand complex Upper Egyptian speech and can distinguish its regional vocabulary, grammar, and pronunciation from the more familiar Cairene Arabic.
For this specific project, workers are required to have resided in the area for their whole lives or to have been ‘deeply immersed’ there, and must identify speakers, overlapping speech and hesitations, and decide how spoken dialect should be written. Payment is per task or audio minute, while work volume rises and falls with client projects.
In another part of the industry, Nasdaq-listed data engineering company Innodata advertised in Cairo for experienced customer service workers. They would check AI customer support responses against product catalogs, company procedures, and detailed scoring rules. The freelance engagement was advertised for approximately two months, with six hours’ daily availability being the ask.
Some projects divide judgment further. Welo Data, a supplier of AI training data, lists Cairo among the locations for its ‘Andromeda Delta’ project, in which, at 7 dollars an hour, workers resolve disputed English transcripts and word-level timestamps.
Under this specific division of labor, first-pass annotators flag ambiguities before adjudicators make final decisions under a common style guide.
It is no surprise, however, because for centuries capital has divided production into specialized operations and combined the people performing them into a collective worker, using their cooperation to produce a capacity that can’t be supplied separately. Yet because the employer directs that cooperation, its productive power appears as an attribute of the capital employing them, and workers ultimately encounter their own combined capabilities as an authority standing over them.
Welo’s annotator and adjudicator gigs illustrate such a relationship. One identifies an ambiguity, and another resolves it according to common instructions. Both contribute more or less expert judgment, but the firm determines the rules connecting their judgments into a finished product.
Likewise, both can become more skilled in their separate functions while the company retains control over the whole. The contributor may need considerable expertise to perform a narrowly assigned function while remaining dependent on an organization they do not control, in fact, an organization they barely come into direct contact with. Thus, capital can, and does, cultivate particular skills while reserving total command over their combination.
On ‘Convo’, a project accessed through Outlier, the contributor platform operated by US company Scale AI and its affiliates, the freelancer who spoke to The Cairo Report recorded 10-minute conversations about daily routines, work, and mundane household activities.
“First we were just two, then three and four and at the end of the project we were five in the conversation,” they said, adding that none of them had ever spoken for more than ten minutes in their lives or seen each other’s faces, but were asked to “talk like we were friends”.
While this specific project required workers to produce speech rather than grading machine-written answers, both activities supplied the industry with human capacities it needed. ‘Convo’, the freelancer recalled, “paid so good too.”
Good pay for an assignment in the AI industry seems to regularly coexist with increasingly precarious income. An Egyptian coder, who also spoke to The Cairo Report, described Outlier coding rates of 12.5–30 dollars an hour, depending on the project. They had nevertheless been assigned language work paying 7.5 dollars. “It’s a nice side income, but you can’t rely on it at all as a main income,” they said.
For the first freelancer, the payment system itself had changed: UHRS paid per task, whereas Populii and Outlier “usually paid per hour”, but that transition didn’t establish a working day they could count on.
But this is also an arrangement that has been around for a couple of centuries. When an employer pays an hourly rate without undertaking to provide a day’s or week’s employment, the worker can receive the stated price for every purchased hour and still lack enough employment to secure subsistence, and the employer continues to profit from the hours purchased while leaving the worker to deal with the interruptions.
An hourly rate prices a fragment of the working day, while a livelihood depends on how many fragments the buyer purchases, how regularly, and how much unpaid preparation each requires.
Other contracts price the output instead, such as Appen, an Australian-listed AI-data company, which advertises its Egyptian Arabic ‘Poliwag’ transcription project through CrowdGen at the staggering pay of 45 dollars an hour.
However, further down the advertisement, the unit becomes clear: an hour of audio transcribed. Workers must capture hesitations, unfinished sentences, timestamps, speaker changes, and background sounds. The price measures the recording delivered, whatever time producing its transcript consumes.
Acceptance determines whether that activity becomes payable, as one Poliwag contributor told The Cairo Report, saying that when they worked on the project over the summer, they had submitted approximately 20 hours of work before they were removed from the project and their submissions disappeared from the dashboard.
Support, which is run by an automated AI large language model, said “payment required passing quality review.”
“It’s difficult to understand how I could qualify for work and then have no accepted produced work or visible submissions at all,” they said.
An OECD report produced jointly with the EU research agency Eurofound describes how this authority can become less visible without becoming less concentrated. Its review of research on algorithmic management notes that software can reduce layers of middle management and fragment work while centralizing knowledge and control. Instructions, monitoring, and evaluation pass through a system instead of a supervisor standing over the worker, which ultimately leads to an organizational hierarchy that may look flatter even as the information needed to direct production remains concentrated above.
This gives the concept of piece wages a particularly concrete application. How much authority has passed to a contributor who must exercise independent judgment, but cannot determine the acceptance criteria, inspect the decision rejecting their work, or preserve access to the record through which payment is claimed? These platforms require expertise while retaining authority over whether its exercise even counts as work deserving payment. Thus, a worker’s discretion becomes one of the capacities the company purchases, while the power to recognize and remunerate its use remains with the buyer.
Still, Egypt’s economic conditions make access to these contracts attractive, as the private sector minimum wage is currently at 7,000 pounds a month, while consumer prices were still rising by 12.7% annually in October 2026, so dollar earnings do in fact offer substantial relief from diminishing domestic wages and rising living costs regardless of the nature of exploitation or poor treatment these workers face in the AI industry.
A medical graduate seeking Outlier work discussed with The Cairo Report their earlier experience with Appen, saying “I didn’t like that they’d started paying just 3 dollars an hour, but I eventually started finding different projects that pay 7 dollars which is slightly better”.
Another call center worker searching for supplementary work explained the pressure: “The salary literally goes toward paying debts,” they said, adding that they had tried finding transcription gigs on Outlier and found no projects available. “I started wondering if I’m not qualified.”
The market translates insufficient demand for their labor into doubt about their personal inadequacy for AI jobs.
ITIDA translates these labor market conditions into an advantage for multinational companies, actively promoting talent-cost savings of up to 60% against “other leading regional offshoring hubs.”
But, to be fair, this isn’t a zero-sum game, as technical advancement can fulfill both promises for a time, and more capable workers can remain cheap and intermittently employed.
The promise of better jobs also depends on which parts of production Egypt is being incorporated into. A new paper by the World Bank titled ‘Shifting Sands’, published on November 13, looks at how technological change is reshaping the relationship between exports and employment. Drawing on research covering 26 economies, it distinguishes occupations in which AI could make workers more productive from those whose tasks it could replace.
Using Egypt’s employment and production data, the study finds an above-average concentration of export-linked employment in the first category. Within business services, however, Egypt falls below the study’s benchmark in both categories, meaning that the economy-wide potential for AI to assist workers doesn’t translate into the same specialization in the services sector officials are promoting.
So, workers don’t need to be directly replaced to face pressure from automation. If AI makes competing suppliers elsewhere more productive, what happens to an Egyptian industry marketed on its ability to supply skilled people cheaply? Buyers will obtain lower prices through productivity improvements in one country and seek lower labor costs in another.
The pressure on Egyptian workers can therefore arrive through the terms of the next contract, even while their own work remains necessary, and their incorporation into more advanced production can renew the commercial incentive to keep them inexpensive.
Uneven and combined development helps explain why this combination persists. Advanced production does not require every social relationship to pass through an orderly sequence of modernization. Imported techniques combine with existing relations and can reinforce them. Applied here, employment insecurity is commercially useful to sophisticated production: companies can draw upon qualified people without maintaining a permanent position for each one.
Stagnant relative surplus population likewise includes people within the active workforce whose employment was extremely irregular, participating in production and simultaneously belonging to its available reserve. Platform qualification makes that availability administratively convenient without creating a corresponding obligation to purchase enough labor to sustain the qualified person.
Egypt’s unemployment rate ostensibly fell to 5.8% in the second quarter of 2026, alongside an increase in employment, but those gains don’t tell us whether an employed person obtains enough paid hours to live on.
Someone must carry the intervening costs, and human capacity to work exists in a person who must eat, find housing, and maintain their abilities between assignments. Other jobs, savings, borrowing, and household support finance that continuity. The company can purchase skilled labor intermittently because those costs continue to be met elsewhere. Where they can’t be met, the worker bears the exhaustion and debt.
This is the class relationship concealed by a story about Egypt entering the knowledge economy. As Sameh Naguib argues in ‘Marxism & the Global South’, international inequalities must be examined through production and class relations. Here, an Indian multinational, an American platform and Egypt’s information technology agency participate in organizing a profitable supply of labor.
An Egyptian owner could preserve precisely the same arrangement.
We don’t have to imagine what such an arrangement would look like, since the state is an active participant in that market. Digital Egypt for Investment, established as an ITIDA-owned company and which promotes itself as “accelerating national growth”, advertises a full-time annotation position in the New Administrative Capital. Applicants need a bachelor’s degree and good Arabic; their work includes identifying keywords and named entities in documents and evaluating and validating AI outputs under the machine learning team’s supervision.
Bel Masry brings the question of “national growth” into sharper focus. Developed by the Communications Ministry’s Applied Innovation Center, the government platform offers transcription, translation and synthesized speech, free to individuals, institutions, and developers. The ministry presents its Egyptian development as securing “digital sovereignty over national data.”
Sure, that represents an expansion of domestic technical capacity, but how does “sovereignty over national data”, that is, over a technological resource, become power for the people employed to build and maintain it? Free access establishes who may use a service, state control establishes where authority over it resides, but neither, by itself, establishes the terms on which its workers participate in deciding how it is produced or—at the very least—provides workers with economic justice.
In the private sector, the industry’s proposed infrastructure reveals another aspect of this grossly unequal relationship. In one case, Egyptian developer Income and Dubai-based Sustainable Infra Capital are planning Project Atlas, an AI data-center campus near Alexandria whose first phase would cost around 2.15 billion dollars, excluding tenants’ servers and chips. But the consortium wants to finance roughly 70% through debt, and Income’s head of strategy and digital infrastructure said that securing finance depends on “long-term binding offtake agreements”, which, in investor-speak, are contracts committing customers to purchase capacity.
Those negotiations are yet another example of how differently uncertainty can be organized within the same industry. The developers, at home, seek a contractual claim on future sales before committing to construction. But why should continuity of demand become a condition for financing computing infrastructure, while the freelancers supplying human judgment must accept its absence as an ordinary condition of work?
Well, a long-term customer contract can make anticipated revenues bankable, allowing investors to borrow against income that has not yet arrived, while the contributor’s qualifications secure no comparable commitment to purchase their time.
Capital’s expansion here therefore depends on two arrangements operating together: predictable revenues sufficient to support long-term investment, and a supply of labor whose purchase remains intermittent. And while the uncertainty obviously has consequences for both, the power to negotiate protection against it is distributed along class lines.
ITIDA, and the state in general, constantly celebrate the development of Egypt’s “human capital”, which is a vague term that blurs a discrepancy on which the outsourcing arrangement depends, because even when a worker possesses the capacity to work, they do not, in any way, control the conditions under which that capacity can earn them a living.
That is precisely why labor power is a commodity, not capital in the worker’s hands, supplying an income only through its repeated sale. Once purchased, it functions as part of the employer’s productive capital, and acquiring more knowledge improves what the worker can sell, but it doesn’t give them any form of meaningful command over its employment.
Egypt’s training strategy can therefore succeed on its own terms while reproducing the dependence these workers describe. More people can qualify for sophisticated assignments, and firms can gain access to a larger supply of skilled labor, but neither development requires those firms to provide each qualified person with enough paid work to live on. Training, in the end, expands productive capacity while leaving the decision about its use with the buyer.
The freelancer who began transcribing in 2021 has already undergone the promised transformation. They learned to evaluate AI and contributed to several platforms.
But that didn’t seem to make a difference in his ability to sustain making a living. “Unfortunately it’s been dry from May [2026] till now,” they said.
The state’s own measure of achievement makes this unusually clear through its ‘ITIDA Gigs’ scheme, which awards a completion certificate after a participant lands a paid gig, or, to put it more accurately, after their skills have found a buyer.
For the worker, however, that sale must be repeated under conditions they do not, and cannot under current arrangements, determine.
At the point where the state can certify success, the worker must begin looking for work again, and again, and again.




