Outlier AI, a platform that brings SMEs to help build GenAI products in the form of training data, is in the news for the wrong reasons. Exposing a huge flaw in the whole training gamut of AI models, the company has been accused of exploiting workers from developing countries to train AI models.
The surprising bit is that Outlier AI is a subsidiary of unicorn Scale AI.
Indian Workers Everywhere
Scale AI, a data platform company backed by tech giants like Meta and Amazon, and even partnered with OpenAI, was recently valued at $14 billion. Its subsidiary, Outlier AI, has been allegedly operating in non-ethical ways and not paying its users.
Users on the remote platform work on hourly wages (or project-basis) for tagging and annotating data, among other things. However, many of them on Reddit have spoken up about the platform’s delayed onboarding, non-payment of wages, etc. And some have even questioned the authenticity of the company.
Interestingly, the kind of project openings on the platform invites not just coders but also people with minimal technical expertise—many from India. There are openings for regional language writing experts in Gujarati, Kannada, Malayalam, Punjabi, Tamil and other languages.
Coders for Less Pay
Outlier AI has also listed a number of openings for coding experts, which is accessible for specific countries, including India. The qualifications expected however compete with that of an actual developer in any tech company.
For instance, proficiency in either Python, Java, C, C++ and similar programming languages is a requisite, and expertise in Swift, Ruby, Rust, Go, and others are listed under preferred qualifications. Further, a bachelor’s degree in computer science and similar fields is preferred.
While discussions on stealing jobs from Indians are doing the rounds, where Indian coders in the US are losing their jobs to coders in India, companies such as Outliers are probably exploiting the coder workforce from India and other developing countries.
In the case of Outlier, the wages are determined by the user’s country of operation. In a recent discussion, a user revealed that despite him having a masters degree and having worked on other AI projects, the platform only offered $7.5/hour as he belonged to a third-world country. Whereas, a user from the US easily earned $40/hr.
Source: Reddit
Big-Tech Companies are All Doing It
“Outlier gets paid when they are hired by big corporations to train their LLMs. It provides the human labour to do that training, which must be done on a massive scale (ironic) and for which task workers are the best solution,” explained a user on Reddit to those trying to understand the structure of Outlier.
While this is just one example, there are many large-scale tech companies already employing startups to train AI models through crowdsourcing, especially from developing countries.
Scale AI’s another subsidiary, Remotasks, runs on a similar model. The platform already has close to 2.5 lakh users. However, this company has also been in the limelight for exploiting users and not paying them. It has employed users from Kenya, Nairobi, and other African countries, besides some from the Philippines too.
Source: X
Interestingly, OpenAI is considered one of Remotask’s customers. Similarly, there are other outsourcing companies for data annotation where users earn as little as $1.2/hour.
Future of Data Annotation
In the days leading to last year’s AI Senate hearing, data workers had requested the lawmakers to protect their rights as they felt their contributions had been sidelined. The workers also pointed out how such a move will prevent them from being exploited by big-tech companies such as Amazon.
Interestingly, a number of companies in India, such as Karya and NextWealth, are employing people from rural areas to help with data annotation and in the process create some form of employment. Gig workers in India engage in microtasks to train AI models, positioning the country as a central hub for data annotation, with the potential to reach a global market worth $8.22 billion and a workforce of one million by 2028.
While this happens on one side, discussions on data annotation dying have also come up. The possibility of auto annotation with minimal human intervention may be a possibility. However, until we get to that point, AI model training employing people from less developed countries will continue.
