There are not only robots behind artificial intelligence (AI): at the end of the chain, there are often workers from southern countries. Recently a Time investigation revealed that Kenyan workers paid less than three euros per hour were responsible for ensuring that the data used to train ChatGPT  did not contain discriminatory content .

AI models indeed need to be trained, by mobilizing an extremely large mass of data, to teach them to recognize their environment and to interact with it. This data must be collected, sorted, verified and formatted. These time-consuming and low-value tasks are typically outsourced by tech companies to a host of precarious workers, usually located in countries in the global south .

This data work takes several forms, depending on the use cases of the final algorithm, but it can for example involve surrounding the people present in the images captured by a video surveillance camera, to learn how to algorithm to recognize a human. Or manually correct errors produced by an automatic invoice processing model.

We propose, through a survey conducted between Paris and Antananarivo, capital of Madagascar, to look into the identity of these data workers, their roles and their working conditions, and to suggest ways to enrich the discussions around the regulation of AI systems.

Artificial intelligence, a globalized production

Our research supports the hypothesis that the development of artificial intelligence does not mean the end of work due to automation , as some authors suggest, but rather its displacement in developing countries.

Our study also shows the reality of “French-style AI”: on the one hand, French technology companies rely on GAFAM services to access data hosting services and computing power; on the other hand, data-related activities are carried out by workers located in the former French colonies, in particular Madagascar, thus confirming the already old logic in terms of outsourcing chains. The literature also compares this type of industry with the textile and mining sector .

An initial observation guided our investigation work: the conditions of production of AI remain poorly known. Building on previous research on “digital labour” , we sought to understand where and how are the algorithms and datasets needed to train them shaped?

Integrated within the Digital Platform Labor research group , our work consists of analyzing outsourcing relationships between French artificial intelligence companies and their subcontractors based in French-speaking African countries and revealing the working conditions of these Malagasy “data workers”, who have become essential to the operation of intelligent systems.

Our survey began in Paris in March 2021. First, we sought to understand the view that French companies producing AI had on these activities related to data work, and what were the processes implemented to ensure the production of data sets of sufficient quality to train the models.

We spoke with 30 founders and employees operating in 22 Parisian companies in the sector. One result quickly emerged from this first fieldwork: the majority of data work is outsourced to service providers located in Madagascar.

The reasons for this concentration of outsourcing flows to Madagascar are multiple and complex. However, we can highlight the low cost of skilled labour, the historical presence of the business services sector on the island, and the high number of organizations offering these services.

During a second part of the survey, first conducted remotely, then on site in Antananarivo, we interviewed 147 workers, managers and directors of 10 Malagasy companies. At the same time, we sent out a questionnaire to 296 data workers located in Madagascar.

Digital jobs: a precarious solution for educated urban youth

First, the field reveals that these data workers are integrated into a broader sector of business service production, ranging from call centers to web content moderation and writing services for optimization. site visibility on search engines.

The data from the questionnaire reveal that this sector mainly employs men (68%), young people (87% are under 34), urban and educated (75% have completed higher education).

When they evolve within the formal economy, they generally occupy a permanent position. The lesser protection offered by Malagasy labor law compared to French labor law, workers’ ignorance of the texts, and the weakness of intermediary bodies (unions, collectives) and company representation nevertheless accentuate the precariousness of their position. They mostly earn between 96 and 126 euros per month, with significant wage differences, up to 8 to 10 times higher for team supervisory positions, also occupied by local Malagasy workers.

These workers are located at the end of a long chain of outsourcing, which partly explains the low wages of these skilled workers, even with regard to the Malagasy context.

The production of AI indeed involves three types of actors: the data hosting and computing power services offered by the GAFAMs, the French companies which sell AI models and the companies which offer computer services. annotations from Madagascar, each intermediary capturing part of the value produced.

The latter are also generally very dependent on their French clients, who manage this outsourced labor force almost directly, with dedicated middle management positions within Parisian start-ups.

The occupation of these management positions by foreigners, either employed by client companies in France, or by expatriates on the spot, represents a significant obstacle to the possibilities of career development offered to these workers, who remain blocked in the echelons bottom of the value chain.

Companies profiting from postcolonial ties

This industry benefits from a specific regime, the “free zones”, instituted in 1989 for the textile sector. From the beginning of the 1990s, French companies settled in Madagascar, in particular for digitization tasks related to the publishing sector. These zones, present in many developing countries, facilitate the installation of investors by providing tax exemptions and very low tax rates.

Today, of the 48 companies offering digital services in free zones, only 9 are run by Malagasy people, compared to 26 by French people. In addition to these formal enterprises, the sector has developed around a mechanism of “cascading sub-contracting”, with, at the end of the chain, informal enterprises and individual entrepreneurs, treated less well than in formal enterprises. , and mobilized in the event of a lack of manpower by companies in the sector.

In addition to the cost of labor, the outsourcing industry benefits from well-educated workers: most have gone to university and are fluent in French, learned at school, on the Internet and through the network of Alliances françaises . This French learning institution was initially created in 1883 in order to strengthen colonization through the extension of the use of the language of the colonizer by the colonized populations.

This diagram is reminiscent of what the researcher Jan Padios calls “colonial recall” . The former colonized countries have linguistic skills and cultural proximity with the countries giving the orders from which the service companies benefit.

Make AI workers visible

Behind the recent explosion of commercialized AI projects in northern countries, there is a growing number of data workers. While the recent controversy around “smart cameras”, provided for in the bill relating to the Paris Olympic Games, has mainly focused on the risks of generalized surveillance , we believe it is necessary to better take human work into account. essential to the training of models, as it raises new questions relating to working conditions and respect for private life.

Making the involvement of these workers visible means questioning globalized production chains, well known in the manufacturing industry, but which also exist in the digital sector. As these workers are necessary for the operation of our digital infrastructures, they are the invisible cogs in our digital lives.

It is also to make visible the consequences of their work on the models. Part of the algorithmic biases indeed reside in the work of data, yet still largely invisible to companies](https://milamiceli.com/wp-content/uploads/2021/10/GROUP2022_CRv1.pdf). A truly ethical AI must therefore go through an AI work ethic .

Author Bios: Clement Le Ludec works in Digital sociology and Maxime Cornet is a PhD student in the sociology of AI both at Télécom Paris – Institut Mines-Télécom