Hidden humans are powering the AI ​​economy

Since January, Tina Lynn Wilson of Hamilton, Ontario, has been freelancing for a company called DataAnnotation.

The 45-year-old woman says she loves the work, which involves checking responses from AI models for grammar, accuracy and creativity. It requires analytical skills and an eye for detail – and she finds some interesting projects too, like choosing the better of two poetry samples.

“Because this is a constructive response, there will be no fact-checking involved. You must explain… what the better answer is and why.”

The work Wilson does is part of a vast, yet not well-known, network of gig workers in the emerging AI economy. Companies like Outlier AI and Handshake AI employ them as “artificial intelligence trainers”, and contract with larger AI platforms to help train their models.

Some data annotation work in other parts of the world is less paid – even exploitative – but there is a wide range of jobs in training, caring for and perfecting AI. It seems this is the labor the tech giants prefer not to talk about. And as models advance, they’ll require more specialized training — meaning companies may soon no longer need many of the humans who helped make them what they are today.

human expertise

We often hear that today’s generative AI is trained on large amounts of data to teach it how human thoughts typically go together. Sometimes called pre-training, this is just the first step. In order for these systems to produce responses that are accurate, useful, and non-invasive, they need to be further refined, especially if they are going to work in narrow areas in the real world.

This is called fine-tuning, and it depends on human expertise. This is basically gig work: done on a per-assignment basis, with no guaranteed hours. Canadian AI trainers we spoke to earn around $20 an hour, although some of the more specialized tasks may pay around $40. Still, incompatibility can be a problem.

“You can’t rely on this as your main source of income,” said Wilson, who described his work as a generalist. That said, many other commentators also consider it an extra effort.

Reinforcement learning from human feedback a kind of fine-tuning It depends on the people evaluating the AI ​​output.

Wilson’s work involves evaluating how “human” an AI response feels.

“This is especially true for voice responses,” she said. “‘Is this something a human being would want to hear?'”

So, when ChatGPIT or the cloud seems extremely human, it’s because humans have trained it to do so.

“It’s still the output of a software product,” said Brian Merchant, a tech journalist specializing in labor and digital technology. “You need quality assurance of the output of a commercial, for-profit, software product.”

Outlier AI has more than 250,000 active contributors in 50 countries, said Fiorella Riccobono, a spokesperson for its parent company Scale AI. Eighty-one percent have at least a bachelor’s degree, he said. The company was not able to provide Canada-specific numbers.

possibly changing market

There are signs that the market for this work is changing, with less demand for the generalized labor performed by people like Wilson. Scale AI recently laid off generalist employees in Dallas, According to Business InsiderIn a shift toward more technical training. Meanwhile, newer, more advanced models have come from Chinese firm DeepSeek Automated the reinforcement process,

“demand “This has increased significantly as AI systems become more complex, thanks to contributors with expert knowledge and advanced degrees,” Riccobono said.

Eric Zhou, 26, was one of these elite workers. After studying materials and nanoscience at the University of Waterloo, he freelanced part-time for Outlier AI for about a year. There, they evaluated prompts and AI responses about undergraduate-level physics and chemistry and corrected the answers.

“It’s a lot of fun if you’re just solving science problems,” he said. “So that problem-solving part, I really enjoyed.”

However, they found that tasks could take longer than the time allotted by the company, so a task listed as $20 for an hour of work could take longer, with no additional pay.

There appears to be no shortage of Canadians working in specific fields who want to supplement their income by improving AI, including many of Zhou’s friends.

This means employees feel they can be constantly replaced, he said.

‘Digital sweatshop’

Yet, AI as a whole depends on the global supply chain during the training process, much of which is outsourced to workers in low-wage countries. This may mean fine-tuning the data, but a lot of the work is in labeling the data, which can be difficult.

The number of people working in this sector is estimated to be in lakhs. Companies have been accused of taking advantage of lax labor laws in regions such as East Africa and Southeast Asia.

“There are a variety of digital sweatshops anywhere from the Philippines to Kenya, where workers are essentially turning these data sets into products that can be used by AI,” said James Muldoon, co-author of the book. feed the machineAbout the hidden costs of human labor and empowering AI.

He says the work can be brutal, as he found in field work in Kenya and Uganda, where people worked up to 70 hours a week for little more than a dollar an hour, under conditions he called “truly appalling.”

While many interpreters had ambitions to work more meaningfully in a technical field, they said they found themselves stuck in “really boring, excruciating” tasks.

AI companies typically don’t focus on human labor powering automation. Tech journalist Merchant says these companies usually want their product to appear “magical, powerful, and futuristic.”

“It’s very rare that your job is completely taken over by machinery, especially in an industrial setting,” he said.

“Usually you have a technology that can help you part of the way.”