What The Hell America!

In her new book, Empire of AI: Dreams and nightmares in Sam Altman’s OpenAI, Karen Hao writes about the Silicon Valley firms competing to be the first to reach artificial general intelligence and the damage they are inflicting on workers in poorer nations. 

The book comes at a time when OpenAI, Google, Meta and others are fast-tracking their AI projects amid rising competition from Chinese firms. Their models are built by a vast hidden workforce in nations such as Kenya, the Philippines and Venezuela that tag data, moderate content, and train and test AI models. 

These workers have complained of precarity and exploitation, including low wages, psychological trauma from moderating disturbing content, and being fired for unionizing. Tech companies are also accused of extracting the creative labor of artists whose copyrighted works are used to train large language models without their consent.

In her book, Hao — who was the first journalist to gain extensive access to OpenAI in 2019 — argues that this route to AI development is not the only way. As Pope Leo XIV warned this month of the challenges AI posed to human dignity, justice and labor, Hao spoke to Rest of World about the choices made by AI firms and alternative paths for AI development. 

The interview has been edited for length and clarity.

Data annotation firm Sama and its client Meta are facing a series of lawsuits from content moderators in Kenya and Ghana on exploitative working conditions. We’ve seen similar reports on OpenAI. What makes these workers vulnerable to the extractive practices you describe as “data colonialism”?

The countries where the digital workers live are still wrestling with the legacies of colonization. Many of them have underdeveloped political institutions and heavily volatile economies. When companies from the Global North enter these countries, they can easily find workers who are well-educated and have good internet connectivity—and are desperate for any job opportunity, and willing to work for any price. 

The governments of these countries are also in a tough position. Some don’t function well and are not acting in the best interest of their citizens. But even the ones that are trying to bring more economic opportunity have little power in stipulating the conditions under which foreign tech companies should operate. They’re willing to allow the companies in to secure the investment and be part of the global technology chain.

These labor practices happen so far away, and that’s by design. The companies choose places that are hidden from the consumers that they sell to.

Data annotation firms are often bound by legal contracts that limit what they can say, allowing Big Tech companies to distance themselves legally and ethically from their workers. What makes working conditions for these digital workers so challenging?

Content moderation for generative AI exposes workers to psychologically traumatic content. Regardless of the benefits they’re provided, whether it’s pay or mental health counseling, the work itself is extremely damaging to anyone who does it for long periods of time. It’s hard to imagine a version of this work that would not be harmful or debilitating.

AI companies have created this veneer of inevitability that, in order to have technological progress, workers will need to do this kind of labor. 

But that trade-off is false. Psychologically traumatizing work is not necessary for progress; it serves only a particular vision of AI development that uses colossal models trained on the whole internet, and that needs an army of workers to filter out its worst content.

Many research organizations outside Silicon Valley advance alternative notions of AI: well-scoped systems trained on well-curated datasets that do not encompass all of the content in the world. In that vision of AI, the need for psychologically compromising work could be significantly minimized or even eliminated.

Generative AI is replacing the clickwork that many migrants in countries like Venezuela relied on. What does that mean for these workers?    

The AI industry has had a long-standing ambition to replace clickworkers with AI models, but edge cases still require humans in the loop. Even so, the work remains precarious because the industry continually moves to different populations and geographies in search of cheaper, more compliant labor.

That’s why Venezuelan workers began falling off platforms [like Appen]. Improving economic conditions made the workers demand higher pay. 

Then, the industry shifted from image-based to language-based tasks, prompting companies to seek out English-speaking workers in Kenya, the Philippines, and India.

The AI industry shifts to different parts of the world whenever it encounters an obstacle. With so many countries undergoing economic or political crises, the labor pools for exploitation are abundant.

Artists, writers, and musicians in the U.S. have filed lawsuits against AI companies for training their models on copyrighted works. Writers and publishers in Singapore have rejected their government’s plan to train LLMs on their work, even as Big Tech firms lobby governments to allow use of copyrighted material under “fair use.” How does this impact artists whose work is used without consent or compensation?

I’ve spoken to artists who say that they do not have economic opportunities anymore. Concept artists, who generate visual ideas for film and advertising, have been especially hit hard. Karla Ortiz, one of the first concept artists who filed a lawsuit against generative AI companies, has described these jobs as solid middle-class jobs that you could build a life around. 

But now, concept artists are not getting jobs, or getting extremely underpaid jobs, because their work was taken by tech companies that created models to replace them. The fact that they’re rapidly losing economic opportunities shows definitively that AI-generated works are functioning as direct substitutes of their original work.

What are some examples of alternative visions for AI that can produce ethically sourced systems—for those who create the data and the workers who train the models?

We can develop AI systems with data that are ethically sourced, labeled, and curated. 

A Maori media organization in New Zealand developed a speech recognition tool for the Maori language with data given by their community members, with consent. They also developed a license requiring any organization seeking to access their data to sign a legally binding agreement that governs how they will use, protect, and secure the data. This ensures the data is not fed into applications that might ultimately harm the Maori community.

In generative AI, the BigScience initiative was an open-source effort involving hundreds of scientists around the world to develop a large language model using archival data from governments and public institutions and donated data. There are a lot of these kinds of models out there. But tech companies have chosen not to adopt any of them. And they have been so effective in completely hijacking the public’s imagination for what AI should look like and how it should be developed.