AI companies seek out more AI experts to train smarter models

Major AI companies are phasing out cheap data labelling workers in Africa and Asia for skilled, higher-paid professionals to help train smarter models.

Before, data labelling workers were primarily given simple annotation tasks. However, AI companies have come to realize that improving their models requires even larger volumes of data. As a result, workers are expected to work even faster and complete hundreds of tasks each day, hence their shift to industry experts.

Scale AI, Turing, and Toloka are hiring more experts

Developing “reasoning” AI systems, including OpenAI’s o3 and Google’s Gemini 2.5, has hastened the shift from low-wage workers in countries like Kenya and the Philippines to more skilled individuals.

Companies such as Scale AI, Turing, and Toloka are already employing leading specialists in fields like biology and finance to support AI teams in generating more refined and complex training datasets.

Olga Megorskaya, CEO and co-founder of Toloka, even commented, “The AI industry was for a long time heavily focused on the models and compute, and data has always been an overseen part of AI. Finally, [the industry] is accepting the importance of the data for training.” 

Scale AI, Turing AI, and Toloka have all seen increased investor interest since their recent shift in strategy. Meta’s $15 billion investment in Scale AI in June raised its valuation to $29 billion. In March, Turing AI secured $111 million at a $2.2 billion valuation, and in May, Bezos Expeditions led a $72 million investment in Toloka.

Turing pays its experts about 20% more than their current pay

Joan Kinyua, head of the Data Labelers Association in Kenya, explained that labellers are now being asked to carry out tasks that depend on their understanding of local languages and cultural nuances.

The organization has also seen increased quality assurance roles where humans review AI-generated content. As OpenAI, Anthropic, and Google work toward creating models that could exceed human intelligence, the priority is shifting to data accuracy and expert analysis.

Jonathan Siddharth, the co-founder and chief executive of data labelling company Turing AI, also claimed that to improve AI models, it’s necessary to use training data from real human usage, especially in complex tasks, and to understand how the models break down in those scenarios.

He even noted that a fully advanced AI system may outperform not just physicists but also become more intelligent than all top experts in all the fields needed to build it.

He added that Turing compensates experts with salaries 20–30% above their current earnings. While AIfirms dedicate only about 10–15% of their budgets to data, compared to the vast sums poured into computational resources, it still translates to significant financial investment.

Toloka’s Megorskaya also argued that features like chain-of-thought, which illustrate how AI models solve problems step by step, are developed through demonstrations by human experts who break down problems into smaller components.

KEY Difference Wire: the secret tool crypto projects use to get guaranteed media coverage