How We Investigated AI Hype To Serve and Empower African Audiences

Lire en français

When we were kids, we used to watch the popular '90s cartoon series The Magic School Bus. The series featured an eccentric school teacher, Ms. Frizzle, her lizard companion, and a yellow school bus that could transform into any shape, size, or “thing” to take students on an educational adventure. Toward the beginning of each episode and before each adventure, Ms. Frizzle would exuberantly say: “Take chances, make mistakes, get messy!” 

Our investigation into the confusing and murky world of recruiting “AI trainers” and “AI tutors” completely embodied this spirit.

Fuelling the AGI Hype is a collaborative, data-driven investigation that revealed the strategy used by micro-tasking companies in recruiting digital workers to “train” Large Language Models like ChatGPT. It also exposed why this strategy is being employed, showing in granular detail how recruitment is used to generate human collateral, largely from the Global South, that could be leveraged at a moment’s notice. 

After speaking extensively with experts in economics, sociology, law, and human rights, among others, we developed the term “labor hedging,” which encapsulates the speculative and precarious nature of the AI industry; an industry based on potentialities rather than actualities.

Popular micro-tasking platform Mindrift—a Dutch subsidiary of formerly Russian-owned tech company Toloka—caught our attention back in July 2024. We noticed that Mindrift had been aggressively recruiting for AI tutors and trainers in South Africa through hundreds, sometimes thousands, of job postings on LinkedIn. Sources on the inside of Mindrift, however, were telling us there was little-to-no work. 

Our investigation found that while companies like Mindrift ran mass-recruitment campaigns on popular job posting sites, they did so knowing there was no promise of work (or any work available at all). In fact, they publicly acknowledged that these weren’t even jobs—they were “gigs.”

---

As a nonprofit journalism organization, we depend on your support to fund more than 170 reporting projects every year on critical global and local issues. Donate any amount today to become a Pulitzer Center Champion and receive exclusive benefits!

---

So, what was the point? This is the question that started our journey into the veritable maze that is the global micro-tasking industry. 

As we navigated everything from Mindrift’s recruitment pipeline and its star-studded client base, to the overall purpose of this grand operation, we frequently ran into dead-ends. We constantly faced new questions and had to change our hypothesis over and over again. We learned quite early on that we needed to become comfortable with being wrong. 

Building a strong foundation

Going into this we knew that data was the key to the AI industry. The more data a company had, allegedly, the more sophisticated their technologies could be. At this point, we did not know what superintelligence was, nor had we heard of AGI, so companies embarking on a ruthless quest for data (at any cost) seemed probable.

After tracking Reddit threads from digital workers who had applied for Mindrift jobs, and those that had worked for the platform, we arrived at our initial failed hypothesis: Micro-tasking companies, like Mindrift, were using the unpaid assessments that were part of their job application process as free labor to train AI models. This was a red-herring. As juicy as this would have been to uncover, it was too easy to guess and we didn’t find any evidence. Mindrift had also posted a blog specifically about this stating that the company does not use the assessments as free data to train AI. 

We pivoted, deciding to study Mindrift’s online behaviour and let the data do the talking. So, we started collecting and tracking job posting data from Mindrift to get a sense of the scale of their recruitment operation. We tried several methods of collecting this data but eventually agreed upon a daily collection of new data at the same time each day over a five month period.

During this period, we worked with Lighthouse Reports to develop a reporting plan and methodology. This was a critical step for us as it allowed us to generate a list of potential hypotheses and reporting questions, and to think concretely about how we were going to tackle them. 

Our mentor on the project, former Pulitzer Center AI Accountability Fellow and Lighthouse Reports journalist Gabriel Geiger, was an incredibly important part of this process, encouraging us to really strip the investigation down and think about the different types of stories that were possible. We needed to define a minimum and maximum story, outline our editorial goals (including who was in charge of each goal) and rank the goals according to priority level. 

All of this took time, and a lot of back-and-forth, but it helped us organize our project more clearly. Along with Gabriel, we built a solid reference point that we could always come back to if we felt like we were getting lost in the weeds. 

Our work with Lighthouse Reports helped us not only to define our methodology but also to uncover evidence that suggested Mindrift was hiring workers knowing there wasn’t any work available. 

So, our question became: Why? What was the end game here? What was the point of hiring hundreds of thousands of workers, only to have them sit for weeks and months at a time without anything to do?

Putting together the pieces of the puzzle

Finding sources for our investigation was a difficult hurdle to overcome. Digital workers exist within a very precarious and vulnerable system. Speaking with a journalist is extremely risky, and could jeopardize their job and subsequently their ability to make ends meet. 

We started cultivating our sources from very early on in this investigation. We were always transparent about who we were, what we were looking into, and made sure the workers we spoke to were aware of the risks it posed to them. With that being said, we did everything in our power to mitigate these risks. This includes using safe communication channels (Signal, Session, or similar encrypted messaging or email services), making sure they did not communicate with us through their laptop or computer while they were working (avoiding screen recording/worker surveillance tech), and only asking them to go on the record once we had established a clear trust between both parties.

We identified these workers based on previous reporting we had done on the topic, as well as by finding them through Reddit discussion threads, Facebook comments, LinkedIn searches, as well as different Whatsapp groups we had joined. Monitoring these different platforms played a huge role in our investigation, as it became a tool that enabled us to gain a better understanding of what was happening inside Mindrift and other similar platforms. 

For several months, digital workers inside Mindrift were telling us that there was hardly any work, but sporadically hundreds of people were being added to the company’s Discord. At this point, not even our sources understood what the company was doing. 

We had a critical breakthrough in mid-June when news broke that Meta was acquiring 49 percent of Scale, the so-called “leading player in the AI data industry.” The business deal meant that Meta was leveling up in the AI industry, and they were also getting Scale’s twenty-something CEO, Alexandr Wang, to run their new “superintelligence” unit. 

We kept rolling different ideas around in our heads; the quest for “superintelligence” meets a big data labeling company, what did it mean? The article spoke about a “shift” in the data work industry, where models were progressively requiring teams of “expert humans” as they became more complex, or developed (simulated) “reasoning” capabilities. This was an “aha!” moment. 

Companies like Mindrift were vigorously recruiting “expert” AI trainers and tutors to respond to the industry’s push towards superintelligence, or Artificial General Intelligence (AGI). Or, so we thought. But something still didn’t add up. If Big Tech players like Meta and OpenAI were talking about AGI, why didn’t the digital workers inside micro-tasking companies have any work?

Learning from the experts

There is only so much that we as journalists can learn from our sources and through social media monitoring. Especially when reporting on AI technologies, there are going to be people out there who know the industry better than we do and have access to empirical research and case studies that can speak to our investigative hypothesis and questions. Consulting researchers and experts in the AI industry who were not affiliated with Big Tech helped us make sense of our on-the-ground reporting. 

It was a conversation that Kathryn had with Antonio Casilli, who is quoted in the story, that really made the story what it is today. When discussing our (incorrect) hypothesis that the industry had changed to further the goal of achieving AGI, Casilli said he refused to legitimize AGI because it doesn’t exist—it’s all hype, he said.

Instead, advertising jobs en masse without work lined up was all a play to signal to investors and potential clients that they had the capabilities, the potential, to fuel a client's drive towards developing superintelligent AI technologies. To add more context to our story, Casilli connected us to economists who could help us understand what this was called in practice: speculative capitalism. 

Engaging with researchers completely changed the game for our investigation, because finally things were lining up. What we were hearing from digital workers inside Mindrift was, in a sense, being clarified by what researchers were sharing with us. Together, they corroborated one another’s experiences and analyses. Up until this point, we had many puzzle pieces in the form of anecdotal evidence from our sources, but after speaking with researchers, it’s as if the puzzle pieces could finally be put together so we could see the full picture. 

With new perspectives from the researchers, we did further reporting using Mindrift’s own website and materials and found further evidence to bulletproof the claims made in our investigation.

Highlighting the human stories

For the first time since we started the investigation, we were finally right about something, or at least we had a hypothesis that we could prove and a question we could answer. However, we were missing a crucial element in our story: a voice, and that had to come from a digital worker who was agreeable to go on the record. Without a voice with whom readers could connect and engage, the story was dry and colour-less. 

From the beginning, we always wanted to keep the investigation focused on the needs of an African audience, and how these issues impact African digital workers. We promised each other that our investigation would serve our local communities. As a result, we made a concerted effort to find African digital workers and activists who would serve as guides in the story. 

Through one of the researchers we consulted, we connected with Joan and Ephantus from the Data Labeller’s Association (DLA), based in Nairobi. Joan and her team from the DLA were so excited to speak with us and their stories and contributions to the investigation really strengthened the piece. It was their stories, their first-hand experiences of being exploited and abused by Big Tech, that ultimately gave the reader a reason to care about any of this. 

Unfortunately, the majority of journalism on AI technologies and Big Tech is centered around the Global North and often excludes Global South, but more specifically, African perspectives. We had seen the reporting from the Global North at the intersection of labor and AI that consistently portrayed African (or Global South) digital workers as invisible cogs in a neo-colonial machine. It was the same story, over and over and over. This type of reporting is not empowering for the people in the story itself, but reinforces the same picture to audiences again and again. 

Our investigation sought to empower and uplift the digital workers who so bravely spoke to us, and offer new perspectives and new stories apart from what had been previously reported on. Instead of once again telling the story about the “invisible workers behind AI,” we told the story of African digital workers who are co-creators of the AI future. 

This was a tough investigation. It challenged us as both journalists and people. In the end, we stayed true to the words of Ms. Frizzle: We took chances even when we were filled with doubt; we made countless mistakes but chose to use them as learning opportunities; and, perhaps most important of all, we fully embraced the messiness of the investigative journey.

Tips for journalists:

1. Keep a record of everything you see online. There are tools that can help with this but ideally you want to be taking screenshots that include the date and time so you have a record of any changes or evidence during the course of your reporting. 
2. Build rapport with sources early on before asking them to speak on the record, even anonymously. Building trust is key, and that takes a tremendous amount of time and effort.
3. Form early connections with independent researchers, academics, or civil-society organizations that are working within the space you are reporting on. They know more than you ever will. 
4. Interrogate data in different ways to observe as many patterns as possible before deciding on the most unique or powerful patterns to showcase to your audiences. 
5. Don’t be afraid to fail. Pivoting is part of investigation—we chase leads until they run cold, we shake ourselves off, and we try again.