In the early stages of artificial intelligence development, models like ChatGPT or competitors like Cohere relied heavily on large teams of low-wage workers to help distinguish simple facts, such as identifying whether an image contained a car or a carrot. However, as AI models have evolved, they now require a more sophisticated approach, involving highly skilled human trainers with specialized expertise—ranging from historians to scientists, many holding advanced degrees.

Cohere co-founder Ivan Zhang explained the shift: “A year ago, we could hire undergraduates to teach AI how to improve. Now, we need licensed physicians, financial analysts, and accountants to help guide the models in their respective fields.” As AI becomes more sophisticated, the need for expert human trainers has grown accordingly.

To meet these new demands, Cohere, valued at over $5 billion, collaborates with a startup called Invisible Tech. Cohere, which focuses on AI for businesses, is one of OpenAI’s main competitors. Invisible Tech, in turn, has become a critical partner for many AI companies, including AI21 and Microsoft, helping to train models to minimize errors, or “hallucinations,” which occur when AI generates incorrect information.

“We have 5,000 people across more than 100 countries, many of whom hold PhDs or Master’s degrees,” said Francis Pedraza, founder of Invisible Tech. These trainers earn up to $40 per hour depending on the complexity of the work, with companies like Outlier paying as much as $50 per hour, and others, like Labelbox, offering up to $200 per hour for experts in fields like quantum physics.

Invisible Tech was initially founded in 2015 to help companies like DoorDash digitize their delivery menus through workflow automation. However, in early 2022, OpenAI approached Invisible with a new challenge: to reduce AI hallucinations ahead of the ChatGPT launch. At that time, early versions of ChatGPT were prone to generating false information. “OpenAI needed a partner to provide reinforcement learning with human feedback,” Pedraza explained.

While OpenAI did not comment on the collaboration, the partnership between Invisible Tech and OpenAI has led to a significant role for Invisible in the AI world. Invisible is now a key AI training partner for many major generative AI (GenAI) companies, including Cohere, AI21, and Microsoft. While Cohere and AI21 confirmed their partnerships, Microsoft did not.

Generative AI, which creates new content based on previously acquired data, occasionally struggles to distinguish between true and false information, leading to hallucinations. In 2023, for instance, a Google chatbot provided inaccurate information in a promotional video about which satellite first captured images of a planet outside Earth’s solar system. Reducing these hallucinations is crucial for AI’s credibility, especially in business settings.

AI companies are acutely aware that hallucinations can undermine the reliability of GenAI systems, so they are working on multiple strategies to reduce these errors. One of the most effective methods involves using human trainers to teach AI the difference between fact and fiction.

Since their initial collaboration with OpenAI, Invisible Tech has become one of the most prominent training partners in the AI industry. Pedraza notes that these AI firms face two major costs: the first being computing power, and the second being the need for high-quality human training to fine-tune the models.

Source