Speaker A
I definitely worked on ChatGPT. I definitely worked on Gemini. I'm shown, like, honestly, some things that humans should never really witness.
Explores the hidden workforce powering AI like ChatGPT, revealing the challenges faced by data annotators and AI trainers in a precarious job market.
Key Takeaways
- AI development depends heavily on a hidden, precarious workforce often paid low wages.
- Many AI data workers are highly educated yet face unstable employment and financial hardship.
- The AI job market is shifting from international low-wage labor to more US-based, skilled workers.
- There is a significant disconnect between the profits of AI companies and the conditions of their workers.
- The rise of AI could exacerbate economic inequality and labor precarity on an unprecedented scale.
Summary
- The video investigates the hidden workers behind AI systems such as ChatGPT and Gemini, focusing on data annotation and AI training jobs.
- It highlights the growing AI job crisis, emphasizing the rise of a new workforce often overlooked and underpaid.
- Many workers are college graduates struggling to find stable employment, taking on AI training jobs with low and unstable pay.
- The workforce includes both international low-wage workers and increasingly US-based professionals with advanced degrees.
- Companies like Mercor and Scale AI act as intermediaries, connecting AI firms with distributed workers globally.
- Workers face job insecurity, pay cuts, and a lack of transparency, often having to accept whatever contracts are offered.
- The AI industry’s rapid growth contrasts with the financial struggles of its data workers, many relying on public assistance.
- Experts warn about unprecedented inequality driven by AI labor dynamics and the precarious nature of these jobs.
- The video includes personal stories from workers like Jen and Ozzy, illustrating the emotional and financial toll.
- Despite the high revenues of data work startups, the workforce remains vulnerable and undercompensated.
Full Transcript — Download SRT & Markdown
Speaker A
[Karen Hao] The AI job crisis is here. But it's not what you've been told.
Speaker A
[Anderson Cooper] You've said AI could spike unemployment to 10 to 20% in the next 1 to 5 years.
Speaker A
Yes. Tech CEOs talk about a future in which AI systems will soon do away with the need for human labor.
Speaker A
But this rhetoric conceals a more sinister reality. While AI is triggering layoffs, its development is also driving the rapid growth of a new kind of worker.
Speaker A
[Worker] I could set alarms for specific jobs. And if that alarm went off in the middle of the night, I was up, and I would work until the tasks ran out.
Speaker A
It's a hidden workforce which Silicon Valley first recruited in low-wage countries. With college graduates making up a record share of the unemployed, data annotation and AI training are among the fastest growing jobs in the US.
Speaker A
So I set out to talk to these workers. I wanted to know: how is AI actually affecting work in this country?
Speaker A
So I'm essentially just working as an anxious rabbit the entire process because, like, will this end tomorrow?
Speaker A
And how do we make sure AI works for all of us, rather than so many of us having to work for AI?
Speaker A
[Daron Acemoglu] The kind of inequality that we're talking about here could be something we've never experienced.
Speaker A
[Jen] I was talking to my mom, and I just start crying. I’m like... I can't see where this money is coming from.
Speaker A
[Karen] Jen is an Ivy League PhD graduate from a small southern town. Jen is not her real name. Like most people we talked to, she asked to be anonymous for fear that tech companies could retaliate.
Speaker A
Well, I highly doubt they want this information public. Last year, as she was wrapping up her degree, she started looking for a job.
Speaker A
You start off like, this is my goal, I know I'm qualified for it. It just keeps getting lower and lower and lower.
Speaker A
Over the past year, I've applied to over 200 roles, and I got maybe three callbacks.
Speaker A
[Karen] Graduates today are facing the worst job market in years. To make ends meet, Jen got on food stamps, moved in with her sister, and started working for $15 an hour as a cashier and substitute teacher.
Speaker A
Then she saw a job posting on LinkedIn for $55 an hour. [Jen] I think the role I saw was philosophy intelligence analyst.
Speaker A
I'm looking and I'm like, “Well, why wouldn't I be able to do that?” After applying, she got a job interview—but not with a human.
Speaker A
Let's start by discussing the languages you speak. I am fluent in English and Chinese, and I consider English to be my primary native language.
Speaker A
Thank you for sharing that. While reporting this story, I got a recruiting message from the same company.
Speaker A
So I did an interview to see what Jen's experience was like. I feel like my English is really excellent because I spend all of my time using it.
Speaker A
It sounds like your job as an investigative journalist really enriches your use of the English language.
Speaker A
The company is called Mercor, and it's one of many contractors which connect companies like OpenAI and Google to a distributed workforce to train and improve their AI systems.
Speaker A
In the same way the AI boom has led to unprecedented growth in data center construction, it has also created an insatiable appetite for these workers.
Speaker A
AI companies kind of present these tools as digital technology. [Karen] Tim Newman is a labor researcher who studies the tech sector.
Speaker A
But this is really the result of an entire global supply chain. Everything from mineral extraction to manufacturing to data centers and the data workers.
Speaker A
For years, tech companies searched for data workers largely outside the U.S. in low-wage countries like Kenya and collapsing economies like Venezuela.
Speaker A
But that is shifting, as AI companies push to build more capable systems. With GPT-5, now it's like talking to an expert, a legitimate PhD-level expert that can help you with whatever your goals are.
Speaker A
So one of the big narratives that's developing in the AI space is that in order to make the next big leap forward, what they need to do is fit the use case for very specific industries.
Speaker A
We've gone from maybe the models being like, you know, a smart high school student to really starting to get at the PhD level.
Speaker A
[Tim] So, in order to make these kind of leaps forward, the companies are really seeking out people with expertise.
Speaker A
[Karen] AI firms are looking for expertise in pretty much everything because their models are only as good as what they're trained on.
Speaker A
[Demis Hassabis] We’re lacking consistency. So, you often hear some of our competitors talk about, you know, these modern systems that we have today are PhD intelligences.
Speaker A
I think that's nonsense. They have some capabilities that are PhD level, but they're not, in general, capable.
Speaker A
By catering to this demand, the four largest data work startups have each had gross revenue of roughly $1 billion a year.
Speaker A
Scale AI claims to have more than 700,000 graduates at its disposal, while Mercor says it has around 30,000 active professionals, like Jen.
Speaker A
[Jen] We all get a message in our group comms where it's like, “Actually, like, this contract is ending.” [Karen] Two weeks after Jen started her first project, right before her graduation, Mercor pulled the rug.
Speaker A
So I'm like, what do I do? You know, like, can I even afford this—should I not get this dinner to celebrate my graduation anymore?
Speaker A
The next month, Mercor offered her a second project, but the pay had dropped to $45 an hour.
Speaker A
When she pushed back, she was ghosted. A few months later, Mercor offered her $35 an hour.
Speaker A
I signed it, even though last year I would have found it so offensive. She had learned her lesson. With this kind of work, you have to take whatever you're offered. And if you're ever offered a high-paying task, it's a race against the clock.
Speaker A
My last contract for Mercor offered me $101 an hour. I told the teacher, you're not going to see me next week, because I'm going to try to max out all these hours before they end the role.
Speaker A
I did those 40 hours. Role ended the next day. I shouldn't have to do that.
Speaker A
No one should have to do that. Like, just disrupt everything. It's predatory! [Voiceover] Paying up to $400 per hour.
Speaker A
Because progress comes from you. This is an industry that really seems to rely on precarious workers, wherever they exist.
Speaker A
Last year, Tim and his colleagues conducted a study of data workers across the country.
Speaker A
They found that 86% struggled to meet their financial responsibilities. A quarter of them relied on public assistance programs like Medicaid and food stamps.
Speaker A
I can't afford to go to the doctor. It feels like I'm rationing my ADHD medication for when I'm able to get back to someone consistently.
Speaker A
[Karen] More than 1 in 5 had experienced homelessness. And overall, the workers reported median earnings of less than $23,000 a year.
Speaker A
In 2021, Alexandr Wang, the former CEO of Scale, became the world's youngest self-made billionaire.
Speaker A
Last year, he was unseated by the three 22-year-old founders of Mercor. [Moderator] Do you hire people that currently work for, like, investment banks or game designers or stuff like that?
Speaker A
Occasionally that would happen, but for the most part it would be people that previously worked at those companies.
Speaker A
Here I've got, like, the home page and then this is just like a big list of all the projects that I've been offered.
Speaker A
[Karen] Ozzy is a college graduate in Oregon who majored in philosophy and worked on Surge AI’s platform.
Speaker A
[Ozzy] The point of it was just that somebody should be able to ask absolutely anything about absolutely anything, and it should be able to understand and do that.
Speaker A
That’s crazy. So that was my job, is to be able to read absolutely anything and then answer absolutely anything.
Speaker A
I read the entirety of Dracula in three hours and then I had, like, 30 minutes to grade this picture book that was, like, absolute dog water.
Speaker A
[Karen] At first, he liked the wide-ranging nature of the work, but then he recei
Speaker A
I think I know what I'm getting into. And it just comes up with the most monstrous of things.
Speaker A
It required him to review violent AI-generated videos. [Ozzy] It was two dudes murdering a golden retriever with their bare hands.
Speaker A
There was, like, weird videos of people constructing, like, furniture out of humans. Like, known celebrities, like, would be, like, in a jail cell, bleeding out, like, with extreme gore.
Speaker A
And like, I would have nightmares, honestly, for a couple of weeks after that. I asked Surge AI about this project, and its spokeswoman claimed its platform does not involve any graphic content.
Speaker A
But that contradicts the documentation Ozzy shared with me. He also thought it was strange that he was often asked to perform tasks he was not qualified for.
Speaker A
They were just, like, so incredibly niche and specialized. Like, can you help us with this calculus homework, or like, this massive math problem, and then, like, straight to a biology thing.
Speaker A
And I was like, I'm not an expert in any of this. Like, people go to grad school for counseling.
Speaker A
I thought about it for, like, 20 minutes. And then I came up with your answer about whether you should reconnect with, like, your abusive father.
Speaker A
[Karen] What I heard from Ozzy and Jen is distinctly similar to what data workers outside the US have told me.
Speaker A
There's a reason for that. [Tim] There are hundreds of contract companies and they're all working for a small number of top clients at the top of the AI supply chain.
Speaker A
It starts at the top, with AI firms seeking to get the best bang for their buck for this kind of labor.
Speaker A
To compete for contracts from those AI giants, the data work companies are incentivized to keep their costs and commitments low.
Speaker A
When you're platform-jumping all over the place, you feel like you don't have any power or room to stand up and say, “Hey, this isn't right.” I'm working 13 hours a day and I might be making $40.
Speaker A
[Karen] Krystal Kauffman is a longtime data worker who now researches labor conditions in the industry.
Speaker A
She joined her first platform after illness pulled her out of traditional work. [Krystal] I was working for one platform at the time.
Speaker A
And then before I knew it, to actually make money to pay bills and things, it became a case of jumping from platform to platform as all of these smaller platforms appeared.
Speaker A
It's not a coincidence that data workers often come from vulnerable backgrounds, whether those on food stamps, without housing, or with disability.
Speaker A
When you're at that point, and I remember being at that point, you take what you can get.
Speaker A
We're describing a world of work that is really about this new mechanism. It's dismantling full-time employment.
Speaker A
[Karen] In 2019, a researcher named Mary Gray gave me a haunting warning: data work could represent the beginning of the Uber-ization of all knowledge work.
Speaker A
On a relentless quest to consume the world's expertise and become its sole provider, Silicon Valley would seek to create a gig version of every job, to acquire bits of knowledge as cheaply as possible from a growing underclass that could eventually include most of us.
Speaker A
We see a future where intelligence is a utility, like electricity or water. And people buy it from us on a meter.
Speaker A
[Daron Acemoglu] The whole ecosystem around Silicon Valley is very much based on this idea that computers are superior to humans.
Speaker A
Most humans are unnecessary. [Karen] Daron Acemoglu is an economist who has extensively researched the effect of AI on the labor market.
Speaker A
He told me that the AI industry's drive for automation is fueled not only by profit motives, but also by an ideology.
Speaker A
There is a sort of elitist attitude towards most workers. People think that the best way you can use AI is more automation and more automation.
Speaker A
And just the less human input there is, the better. No human CEO can talk to every employee at a company, every customer, be in every meeting, be an expert in every field.
Speaker A
And so, more and more, I think, of these jobs will be supervising a bunch of AI.
Speaker A
Right now, we are choosing to use AI as an automation technology. It creates a vicious cycle whereby more and more of the tasks start being taken by AI.
Speaker A
And on the way to that, we need human workers to train AI models. [Karen] It looks like this: employers cite AI as a reason to lay off workers.
Speaker A
Tech firms hire those increasingly desperate workers for cheap — to train AI. Now, the kind of inequality that we’re talking about here could be something we’ve never experienced.
Speaker A
A handful of corporations controlling most of work, and a large fraction of workers essentially completely sidelined from meaningful work.
Speaker A
That is a choice. That's not the nature of the AI technology itself. We can use AI for other things.
Speaker A
Rather than automate teaching, we can give teachers tools so that they can provide individualized education in a much cheaper and effective form.
Speaker A
We can have pro-nurse AI, meaning AI tools that increase the capabilities of nurses for diagnosis, cure, and treatment.
Speaker A
But those options are not being exploited. But in its quest to ingest the world's knowledge, Silicon Valley has a problem.
Speaker A
The AI industry is dependent on the very workers it exploits. Data workers have leverage, and some of them are already using it.
Speaker A
Krystal was the lead organizer of Turkopticon, one of the first efforts to build power to improve labor conditions on Amazon's platform.
Speaker A
On some issues, they won. We fought very hard for a change to the rejections policy.
Speaker A
To get a corporation like that to actually listen to workers was almost surreal. We absolutely have to have some type of global coalition collectively demanding better conditions.
Speaker A
For hundreds of thousands of far-flung data workers to organize will be an enormous challenge.
Speaker A
But advocates like Tim believe there are models in recent history that workers can look to.
Speaker A
[Tim] Once workers started organizing in the garment sector, we've seen them creating powerful coalitions globally, including with consumers, with investors, with students.
Speaker A
A new bill introduced in California this year borrows from that kind of strategy. [Assemblymember Lee] This is AB 2653, The California Sweatshop-Free AI Procurement Act.
Speaker A
It would ensure that when the state of California is procuring an AI tool, the data work that created those systems actually complied with certain labor standards.
Speaker A
This is simply about how we are spending taxpayer dollars, and it should not go to AI companies who exploit their workers, whether they're here in California or anywhere else in the world.
Speaker A
There is a narrative about how the AI companies operate, that this technology is inevitable.
Speaker A
That is not the case. All of us shape the future of technology. And what's most important is that we start working together to be sure that the future of work and the future of technology at work is one that, you know, benefits all of us
Speaker A
and makes our jobs better, not leading to a race to the bottom purely to enrich a handful of very wealthy billionaires in Silicon Valley.
Topics:ChatGPTAI workforcedata annotationAI training jobsprecarious laborMercorScale AIAI job crisisSilicon ValleyAI inequality
