The Bangladeshi freelancers quietly training the world's AI

It has been nearly four years since ChatGPT introduced large language models (LLMs) to the public. Today, these models can write emails, draft code, generate media, and assist with a wide range of business processes. Yet, behind each response sits a human, a data annotator or AI trainer who tests the model, ranks its answers, and corrects its mistakes. A growing number of people working in this industry are Bangladeshi.

So, what does a data annotator actually do?

Tausif Niloy, who worked at Meta as a data annotator between 2024 and 2026, described his responsibilities, "I used to do data labelling based on golden set prompt constraints, exploiting AI models to generate violating responses, using multiple attack vector types, including code injection, false premises, and so on." 

His work primarily involved model testing and red-teaming — the practice of working against the AI to test system resilience.

Other popular tasks, such as tagging images and audio, are more foundational; they are done so that a model learns what it is looking at. Transcribing and rating chatbot replies are also routine requirements. 

This is how a model learns which answers people prefer, flagging harmful outputs so it learns what not to say. Stack enough of those corrections together, and you get the fluency users now take for granted.

The market for this work is expanding fast. According to Straits Research, the data annotation tools market was valued at around $2.37 billion in 2025 and is projected to grow to nearly $30 billion by 2034. Meanwhile, the World Bank estimates that between 154 million and 435 million people worldwide work in various AI-training related online gigs, including image labelling and content review.

While there is clearly a demand for these jobs, certain routine tasks within the sector are shrinking due to automation. This raises a vital question: Will the work Bangladesh provides survive the rise of automation?

Bangladesh's case for establishing a foothold in this digital space rests on its large, English-proficient workforce, which uses the internet to perform IT-enabled work for foreign clients. In 2021, the country's IT workforce accounted for around 15% of the global online labour market. According to figures previously reported by TBS, the Bangladeshi government estimates there are around 650,000 freelancers in the country who contribute to the economy through digital spaces, earning approximately $725 million during the 2024–25 fiscal year.

Nabeel Mohammed, an associate professor of Computer Science at North South University and an expert on computer vision and natural language processing, spoke to TBS about whether Bangladesh's AI training industry could mirror the Business Process Outsourcing (BPO) sector.

"AI data work parallels previous BPO work, using lower-cost capacity for repetitive tasks," he said. "However, working with AI data requires higher education and tech knowledge. It will become yet another outsourcing wave if we cannot use our AI data work as a foundation for higher-value output."

Mohammed framed the present opportunity brought forth by the data annotation sector, noting, "These jobs can be a good initiation into the AI supply chain, as long as workers and their leaders create pathways to improve skills, enabling higher-value outputs for local and global needs."

An opportunity, though, is only worth what it leads to — and some cases paint a completely different picture. In 2025, a report by Time Magazine found that American technology companies pay outsourcing firms as much as $12.50 per hour for AI annotation work, while Kenyan workers receive as little as $2. This wage arbitrage affects workers in Bangladesh too, as LinkedIn job listings frequently offer similar low rates.

Furthermore, in September 2024, a Kenyan court ruled that Meta could be held legally responsible for 184 outsourced content moderators from Kenya, leading to a $1.6 billion lawsuit. The wage gap that these companies utilise is no accident. It is baked into the system, and routine work is frequently routed to the lowest-cost location.

Like many industries, data annotation has a low skill floor but a high skill ceiling. Basic labelling and rating are easy enough to pick up, which allows Bangladesh to enter the market at scale. However, because these workers are easily replaced, they have less negotiating power, which keeps wages down. 

Conversely, the ceiling — encompassing red-teaming, specialist reviews and policy work — pays well and remains insulated from automation. The issue is that much of Bangladesh's capacity and training efforts are geared towards teaching simple, repetitive tasks.

Nabeel Mohammed sees the human cost of this approach. "We have some really talented individuals locked into routine tasks that do not always translate into significant career progression," he observed. 

"Given the right environment and time, these individuals can produce better, higher-value output for us. AI training can be very useful. However, real value comes from doing practical work. Not all training programmes can provide that experience, leading to a false sense of skill."

There is, however, an even greater risk than a mere race to the bottom. Some annotation tasks are now trivial due to the rapid increase in AI capabilities, meaning Bangladesh has entered a field where the floor could disappear quickly. 

"Data annotation skills in demand today can become less important in months, not years," Mohammed warned, painting a stark reality. "Unless we aim higher, the risk is not a low-wage group but rather an unemployed one."

Aside from the unvarnished macroeconomic truths, how does this work play out for the individual? Can data annotation culminate in a career, or does it remain a stopgap? That depends largely on securing full-time, skilled positions. 

Tausif Niloy indicated that data annotation could lead to other opportunities within a company, noting, "One can advance to project coordinator roles. The most lucrative possibility is moving on to the policy-making team."

On the other hand, Fardeen Zareef, who worked as an AI annotator from 2024 to 2025, found his employer's promises to be unfounded. "The company I worked at said that they would offer ways of advancement, but we never really got to see the fruit of it because they started layoffs a year in," he said.

Despite these hurdles, is the industry worth pursuing? For a freelancer working from Bangladesh, the pay is still slightly better than most local alternatives. For those who bring skills beyond the minimum threshold, the prospects can be worthwhile.

For Bangladesh as a nation, however, everything depends on what is built on top of the foundation, according to Mohammed. 

"Human input is still relevant. However, the skill level of the human in the loop increases as AI models become more capable. Our AI training and data annotation industry will remain relevant only as long as we provide the required value. But where do we add that value?" he asked. "We can focus on providing it towards our homegrown products and services. That will depend on policy, environment, investment, R&D, and the whole ecosystem."

His recommendations reach far past the industry as it exists today. 

"We can aim to move from training to be AI consumers to gaining the training and exposure needed to be AI providers," he said. "A large number of courses exist that profess to teach the effective use of AI tools. Many of these skills are at risk of becoming irrelevant. We should aim to become providers of AI, specifically types suitable for the Global South. This requires a public-private partnership and practical collaboration between industry and researchers."