In Madagascar, the unsung heroes who keep artificial intelligence running
ChatGPT responds in a matter of seconds. Google recognises an image. Amazon anticipates a search. Artificial intelligence tools give the impression of autonomous, almost magical technology.
But behind every seamless response, every analysed image and every piece of filtered content, there is often human labour that nobody sees.
In Madagascar, thousands of people spend their days clicking, categorising, correcting, listening to, describing or verifying data. Their mission: to make algorithms more accurate, more reliable and more useful.
Some have a contract, an office and a team. Others carry out a succession of micro-tasks from their living rooms, paid a few pence. A few work alone, without social security cover, with irregular income and total dependence on foreign platforms.
Artificial intelligence is often portrayed as a revolution brought about by machines.
In Antananarivo, it takes on a different face: that of a human-centred, fragmented and fragile economy — but one that is also a source of local skills and ambitions.
In brief: Madagascar has become a major hub for data work in the field of AI. Workers there annotate images, text, audio and videos to feed global outsourcing chains. The sector creates opportunities for a digitally connected younger generation, but also raises questions about wages, social protection, transparency and the sharing of value.
What we really know about AI workers in Madagascar
The report Madagascar: the little hands of AI from ARTE Reportage, published in 2025, suggests that around 100,000 people are involved in this type of work every day in the country.
This figure should be treated with caution: it is an estimate derived from the report, not an exhaustive public census. It covers a highly diverse reality: employees of local companies, freelancers, workers on international platforms, subcontractors and people who use this work to supplement an insufficient income.
What academic research does confirm, however, runs much deeper: global AI relies on a largely invisible chain of human labour.
A study focusing on outsourcing between France and Madagascar shows that AI systems require people to annotate, classify, verify, transcribe and produce data. The researchers describe a global organisation in which the least visible and most repetitive tasks are frequently shifted to countries with lower labour costs. See the study published in Big Data & Society.
Artificial intelligence does not, therefore, replace human labour.
It shifts it. It fragments it. And often, it makes it difficult to identify.
How humans train artificial intelligence
An algorithm does not learn on its own to recognise a cat, a disease on a tomato leaf or a dangerous comment.
First of all, you need to show them examples. Lots of examples.
This task may seem simple when summarised as ‘clicking on images’.
In reality, it requires concentration, a detailed understanding of instructions, sometimes language skills, the ability to work at a fast pace and the capacity to make repetitive decisions for several hours on end.
The problem is that the less visible this task becomes in the final product, the easier it is to underestimate its value.
David, Elina, Dani: three realities behind the term ‘click worker’
In the ARTE report, three career paths show that the sector cannot be summed up in a single word.
David: supplementing his income, one job at a time
David, a father of three, carries out a series of micro-tasks from his living room between shifts at his local snack bar.
Her daily life illustrates the nature of on-demand work: tasks come in, earnings vary, and time is fragmented. Digital work does not necessarily replace an existing job; it is sometimes an addition to an already precarious economy.
It’s an appealing prospect: working from home, with just a computer and an internet connection.
But there’s a downside: if assignments dry up, if your account is closed, or if the platform changes its rules, your income can disappear without warning.
Elina: the more stable face of subcontracting
Elina, aged 25, works for a local company. She has a contract and a modern office environment.
Her career path highlights a key point: not all data-related work in Madagascar is informal. Well-structured companies recruit, train and organise teams dedicated to annotation, transcription or quality control.
This version of the sector can create skilled jobs and open up genuine career paths.
But it also raises a question: how much of the value created by these teams remains in the country, and how much flows back to the clients, platforms and technology companies based elsewhere?
Dani: the dream of independence, the reality of dependence
Dani works on his own. To access certain platforms that are unavailable in his country, he explains that he has purchased a foreign account.
He hopes to earn enough to build a house and provide greater stability for his family.
His story highlights the paradox of digital micro-work: whilst it can provide access to a global market, that access is subject to rules decided far from home.
Independence exists as long as the platform is willing to let you work.
Why Madagascar has become a key hub for AI outsourcing
Madagascar was not chosen by chance.
The country has a French-speaking population, a tradition of outsourced services, a pool of young graduates and labour costs that are attractive to foreign companies. Antananarivo is home to a significant number of BPO firms Business Process Outsourcing able to provide remote services: customer relations, transcription, data processing, moderation or annotation.
Researchers who have surveyed French companies and Madagascan subcontractors emphasise that data annotation forms part of a longer history of outsourcing digital tasks. Their analysis in Le Monde highlights, in particular, the role of the French language, existing infrastructure and pay gaps.
In other words: AI did not create this geography of work. It has accelerated it.
Before AI, there were already call centres, outsourced administrative services and remote working platforms. With AI, new tasks have been added: teaching a machine to see, listen, sort, write and respond.
The real issue: the AI value chain
When a user pays for a subscription to an AI tool, they see an interface. They do not see the workflow that made it possible.
However, there are often several layers behind a generated response:
- expensive infrastructure: servers, data centres, chips and energy
- researchers and engineers who develop the models
- companies that collect, clean and organise data
- subcontractors recruiting teams
- workers who carry out the most repetitive tasks
- platforms that organise, rate and pay for these assignments
The further down this chain you go, the more invisible the work generally becomes.
And it is becoming increasingly difficult to know who is responsible for the conditions under which it is produced.
A comparative study focusing on AI production in Brazil, France, Madagascar and Venezuela describes these supply chains as mechanisms capable of reproducing long-standing inequalities between countries that design the technologies and countries that provide part of the labour required for their operation. Read the study Global Inequalities in the Production of Artificial Intelligence.
The key word here is not just ‘technology’.
It’s “distribution”.
Who generates the revenue? Who bears the risks? Who owns the data? Who can progress to better-paid roles? Who remains merely carrying out a task without knowing either the end customer or its future use?
A real opportunity, but a fragile model
To reduce Madagascar to a “click factory” would be as unfair as it is incomplete.
The sector can contribute:
- earning a living in a challenging job market
- a first digital experience
- skills in languages, data, quality control and collaborative tools
- a gateway to international markets
- salaried jobs in certain local companies
- an opportunity to gradually progress towards roles in supervision or project management
However, this opportunity remains precarious if the following conditions are not met:
- predictable earnings
- a contract that’s easy to understand
- social security cover
- the right to appeal in the event of account closure
- transparency regarding tasks and clients
- protection against offensive content
- prospects for career development
- an opportunity to put the skills you’ve learnt into practice through local projects
The danger would be to celebrate ‘flexibility’ without seeing what it might conceal. Deliberate flexibility provides autonomy.
Imposed flexibility shifts the risk onto the worker.
The often invisible risks of data processing
Salary is the most visible aspect of the debate. But it is not the only one.
Unstable income
On micro-task platforms, a person may be paid per task, rather than for time spent. If the instructions change, if the task is rejected or if demand slows down, earnings plummet.
The worker therefore bears the business’s commercial risk, without benefiting from its stability.
Inadequate social security
Without a local employment contract, standard protections may be lacking: holiday entitlement, insurance, pension, support in the event of illness, recourse in the event of a dispute, or compensation upon completion of an assignment.
A lack of transparency regarding the end use of the data
A person can annotate images, categorise posts or evaluate texts without knowing in which final product their work will be used.
This lack of transparency is problematic, particularly when the data is used in sensitive systems such as surveillance, recruitment, credit, security, image recognition or content moderation.
A potential psychological burden
Certain forms of moderation or assessment expose workers to violent, hateful or traumatic content.
Not all data-processing tasks carry this risk. But where it does exist, piecework pay is not enough: support, limited exposure and genuine guidance are required.
Total dependence on platforms
An account may be suspended. A rule may change. A geographical area may be blocked. A platform may reduce the number of available assignments.
The worker remains connected to the whole world, but has no control over its rules.
The story we mustn’t tell: “AI creates jobs, so everything’s fine”
To say that AI creates jobs is true. To say that this is enough to prove that it drives progress is false.
The real question is the quality of the jobs created.
A digital job can be a gateway to independence, training and entrepreneurship. It can also become a form of globalised precariousness, in which highly skilled people carry out essential tasks without recognition, without stability and without access to decision-making power.
The debate on AI is often dominated by two extremes:
- “AI will destroy all jobs”
- “AI will create opportunities for everyone”
Reality is more demanding.
AI is transforming jobs, creating new professions and shifting certain tasks. But the benefits are not automatically distributed fairly.
Without rules, without transparency and without the local capacity to create its own tools, innovation can reinforce the very dependencies it claims to overcome.
In Madagascar, too, AI is being developed to meet local needs
The ARTE documentary does not merely show the unseen work of the annotators.
It also showcases the other side of the country: engineers and entrepreneurs who want to use AI to tackle local problems.
Fitahiana and Fahasoavana are developing an artificial intelligence application dedicated to agriculture, capable of helping to detect plant diseases.
The idea seems simple. Its significance is immense.
AI that can identify a plant disease from an image can help a farmer take action sooner, better protect their crops and minimise losses. Provided, of course, that it is adapted to the crops, languages, internet connectivity and the realities on the ground.
This is where a key difference comes into play.
An AI developed entirely elsewhere may treat Madagascar as a source of labour or data.
AI developed in Madagascar can also become a problem-solving tool for the country’s farmers, businesses, teachers, healthcare workers and citizens.
The future isn’t just about teaching machines thanks to Madagascar.
It also involves giving Madagascar the means to build its own machinery, develop its own practices and establish its own businesses.
What companies using AI should change
Talking about responsible AI without mentioning the workers who develop it no longer makes sense.
Any organisation that designs, purchases or deploys an AI system should be able to answer these questions:
- Who prepared, annotated or verified the data?
- In which countries was this work carried out?
- Are the people concerned employees or paid on a piecework basis?
- Do wages provide a decent standard of living locally?
- Is there any protection against traumatic content?
- Can subcontractors be audited?
- Do workers have any recourse in the event of a dispute?
- Are sensitive data and operations properly regulated?
- Does any of the value created benefit local ecosystems?
Transparency does not slow down innovation. It ensures that innovation is not shrouded in mystery.
How this study changes our perspective on ChatGPT and generative AI
The next time an AI answers a question instantly, there’s one thing to bear in mind: its speed doesn’t mean that human labour has disappeared.
Perhaps it has simply become invisible.
Artificial intelligence is not limited to a single model, an interface or a promise of productivity. It is a global chain comprising data, infrastructure, economic decisions and workers.
In Madagascar, this chain takes on a very tangible form: tens of thousands of people are trying to find an income, a skill, stability or a path to the future there.
Some see it as an opportunity. Others find it brings new insecurity.
And in between, one question remains unanswered: will the AI of tomorrow merely extract labour and data from the Global South, or will it finally enable the wider distribution of knowledge, power and value?
Perhaps this is the true test of so-called ‘artificial’ intelligence.
Not just what it can do. But what it enables humans to become.
FAQ: The unsung heroes of AI in Madagascar
What do AI workers do in Madagascar?
They can annotate images, classify text, transcribe audio, verify data, moderate content or evaluate the responses of automated systems. These tasks are used to train, test or improve digital tools and artificial intelligence models.
Why do companies outsource data annotation?
Annotation requires a great deal of manual labour, which is often repetitive and difficult to automate. Companies outsource this work to subcontractors or platforms, particularly in countries where labour costs are lower and where the necessary linguistic and digital skills are available.
How many people work in AI in Madagascar?
The ARTE report Madagascar: the little hands of AI suggests that nearly 100,000 people are involved in this type of activity on a daily basis. This figure should be understood as a journalistic estimate, as there is no unified public census covering all employees, freelancers and platform workers.
Are all ‘click workers’ in precarious employment?
No. Some people are employed by local businesses and have a contract, a working environment and prospects for career progression. Others work on a freelance or piecework basis, with more unstable incomes and limited social security cover.
Can AI create local opportunities in Madagascar?
Yes. Beyond simply acting as subcontractors, Malagasy engineers and entrepreneurs are developing tools tailored to the country’s needs, particularly in agriculture. The challenge is to enable these local skills to also create products, businesses and value locally.