I'm a contractor for one of these companies. It pays okay ($45+/hour) if you can pass qualifications for your area of expertise but the work isn't steady and communication is non-existent. The coding qualifications I did were difficult FAANG algorithm analysis questions. The work has definitely gotten harder over the last year and often says we need to come up with Masters/PhD level work or problems that someone with 5+ years of experience in a field would have difficulty solving. I wish I had a regular job but I live in rural North Carolina and remote work is hard to come by.
Look up Mercor, DataAnnotation.tech, and Outlier. You create a profile, upload a resume, and do some required tasks for each job posting they have. It may involve a combination of interviewing with an AI, doing a few trial tasks, and submitting a portfolio or Github profile.
I just want to note that asking this question implies an openness to one’s personal affairs that may not be appropriate in an anonymous, public setting. A person offering context and insight to a topic is not necessarily an invitation to an for more personal contexts and insights.
I understand it's personal, but I also recognize they went out of their way to bring it up. Some people, including me, are more willing to discuss things anonymously because it adds a layer of impersonality. This is just a discussion board. If OP doesn't answer, that's ok. I don't ever think I'm entitled a response.
I only started seriously looking for work again about a month ago. I'd like to stay in this area for a few reasons but I would relocate if necessary. I worked remotely from 2015 until a layoff in late 2023 and this was the first thing I came across after that. It was okay for awhile and actually pretty interesting at first but the hours aren't reliable and there doesn't seem to be much opportunity for getting promoted.
> “At first they told [me]: ‘Don’t worry about time – it’s quality versus quantity,’” she said.
> But before long, she was pulled up for taking too much time to complete her tasks. “I was trying to get things right and really understand and learn it, [but] was getting hounded by leaders [asking], ‘Why aren’t you getting this done? You’ve been working on this for an hour.’”
And:
> Dinika said he’s seen this pattern time and again where safety is only prioritized until it slows the race for market dominance. Human workers are often left to clean up the mess after a half-finished system is released. “Speed eclipses ethics,” he said. “The AI safety promise collapses the moment safety threatens profit.”
Finally:
> One work day, her task was to enter details on chemotherapy options for bladder cancer, which haunted her because she wasn’t an expert on the subject.
Something I'd be interested to understand is how widespread this practice is. Are all of the LLMs trained using human labor that is sometimes exposed to extreme content?
There are a whole lot of organizations training competent LLMs these days in addition to the big three (OpenAI, Google, Anthropic).
What about Mistral and Moonshot and Qwen and DeepSeek and Meta and Microsoft (Phi) and Hugging Face and Ai2 and MBZUAI? Do they all have their own (potentially outsourced) teams of human labelers?
I always look out for notes about this in model cards and papers but it's pretty rare to see any transparency about how this is done.
> Are all of the LLMs trained using human labor that is sometimes exposed to extreme content?
The business process outsourcing companies labelling things for AI training are often the same outsourcing companies providing moderation services to facebook and other social media companies.
I need 100k images labelled by the type of flower shown, for my flower-identifying AI, so I contract a business that does that sort of thing.
Facebook need 100k flagged images labelled by is-it-an-isis-beheading-video to keep on top of human reviews for their moderation queues. They contract with the same business.
The outsourcing company rotates workers between tasks, so nobody has to be on isis beheading videos for a whole shift.
It has been documented that human image moderators exist and that some have been deeply traumatized by their work. I have zero doubts that the datasets of content and metadata created by human image moderators are being bought and sold, literally trafficking in human suffering. Can you point to a comprehensive effort by the tech majors to create a freely-licensed dataset of violent content and metadata to prevent duplication of human suffering?
One of the key innovations behind the DNN/CNN models was Mechanical Turk. OpenAI used a similar system extensively to improve the early GPT models. I would not be surprised that the practice continues today; NN models needs a lot of quality ground truth training data.
Given the number of labs that are competing these days on "open weights" and "transparency" I'd be very interested to read details of how some of them are handling the human side of their model training.
I'm puzzled at how little information I've been able to find.
Beyond that, I think the reason you haven't heard more about it is that it happens in developing countries, so western media doesn't care much, and also because big AI companies work hard to distance themselves from it. They'll never be the ones directly employing these AI sweatshop works, it's all contracted out.
I have been a generalist annotator for some of the others you mentioned, due to NDA will not specify which. I would venture to guess that basically all major models use some degree of human feedback if there is money coming in from somewhere.
it explores the world of outsourced labeling work. Unfortunately hard numbers on the number of people involved are hard to come by because as the article notes:
"This tangled supply chain is deliberately hard to map. According to people in the industry, the companies buying the data demand strict confidentiality. (This is the reason Scale cited to explain why Remotasks has a different name.) Annotation reveals too much about the systems being developed, and the huge number of workers required makes leaks difficult to prevent. Annotators are warned repeatedly not to tell anyone about their jobs, not even their friends and co-workers, but corporate aliases, project code names, and, crucially, the extreme division of labor ensure they don’t have enough information about them to talk even if they wanted to. (Most workers requested pseudonyms for fear of being booted from the platforms.) Consequently, there are no granular estimates of the number of people who work in annotation, but it is a lot, and it is growing. A recent Google Research paper gave an order-of-magnitude figure of “millions” with the potential to become “billions.” "
I too would love to know more about how much human effort is going into labeling and feedback for each of these models, it would be interesting to know.
That was indeed a great article, but it is a couple of years old now. A lot of of the labeling work described there relates to older forms of machine learning - moderation models, spam labelers, image segmentation etc.
Is it possible in 2025 to train a useful LLM without hiring thousands of labelers? Maybe through application of open datasets (themselves based on human labor) that did not exist two years ago?
Good question, I don't personally know. The linked article would suggest there are plenty of people working on human feedback for chatbots, but that still doesn't give us any hard numbers or any sense of how the number of people involved is changing over time. Perhaps the best datapoint I have is that revenue for SurgeAI (one of many companies that provides data labeling services to Google and OpenAI among others) has grown significantly in recent years, partly due to ScaleAI's acquisition by Meta, and is now at $1.2 billion without having raised any outside VC funding:
Their continued revenue growth is at least one datapoint to suggest that the number of people working in this field (or at least the amount of money spent on this field) is not decreasing.
Also see the really helpful comment above from cjbarber, there's quite a lot of companies providing these services to foundation model companies. Another datapoint to suggest the number of people working providing labeling / feedback is definitely not decreasing and is more likely increasing. Hard numbers / increased transparency would be nice but I suspect will be hard to find.
So why do you think asking this question here would yield a satisfying answer, especially how the HN community likes to dispute any vague conclusions for anything as hyped as AI training?
To counter your question, what makes you think that's not the case? Do you think Mistral/Moonshot/Qwen/etc. are all employing their own data labelers? Why would you expect this kind of transparency from for-profit bodies that are evaluated in the billions?
> If you don't ask the question you'll definitely not get an answer.
Sure, but the way you're formulating the question is already casting an opinion. Besides, no one could even attempt to answer your questions without falling into the trap of true diligence... one question just asks how all (with emphasis!) LLMs are trained:
> Are all of the LLMs trained using human labor that is sometimes exposed to extreme content?
Who in the world would even be in such a position?
That question could be answered by proving the opposite: if someone has trained a single competent LLM without any human labor that was exposed to extreme content then not all LLMs were trained that way.
"Google said in a statement: “Quality raters are employed by our suppliers and are temporarily assigned to provide external feedback on our products. Their ratings are one of many aggregated data points that help us measure how well our systems are working, but do not directly impact our algorithms or models.” GlobalLogic declined to comment for this story." (emphasis mine)
How is this not a straight up lie? For this to be true they would have to throw away labeled training data.
Is there a reason not to use validation data in your next round of training data? Or is it more efficient to reuse validation and instead get more training data?
More recent models actually use "reinforcement learning from AI feedback", where the task of assigning a reward is essentially fed back into the model itself. Human feedback is then only used to ground the training, on selected examples (potentially even entirely artificial ones) where the AI is most highly uncertain about what feedback should be given.
The title is biased, blaming Google for mistreating people and implying that Google's AI isn't smart, but the OP is worth reading, because it gives readers a sense of the labor and cost involved in providing AI models with human feedback, the HF in RLHF, to ensure they behave in ways acceptable to human beings, more aligned with human expectations, values, and preferences.
RLHF (and its evolution, RLAIF) is actually used for more than setting "values and preferences". It's what makes AI models engage in recognizable behavior, as opposed to simply continuing a given text. It's how the "Chat" part of "ChatGPT" can be made to work in the first place.
> Sawyer is one among the thousands of AI workers contracted for Google through Japanese conglomerate Hitachi’s GlobalLogic to rate and moderate the output of Google’s AI products...
Depends how you look at it. I think a brand like Google should vet a mere one level down the supply chain.
I'm a big fan of cyberpunk dystopian fiction, but I still can't quite understand what you're alluding to here. Can you give an example value that google align the AI with that you think isn't a positive human value?
Their entire business model? Making search results worse to juice page impressions? Every dark pattern they use to juice subscriptions like every other SaaS company? Brand lock-in for Android? Paying Apple for prominent placement of their search engine in iOS? Anti-competitive practices in the Play store? Taking a massive cut of Play Store revenue from people actually making software?
Google Gemini 2.5 Pro actually has a quite nuanced reply when asked to consider this statement, including the following:
> "Massive privacy invasion: The core of modern adtech runs on tracking your behavior across different websites and apps. It collects vast amounts of personal data to build a detailed profile about your interests, habits, location, and more, often without your full understanding or consent."
You don't boil the frog instantly. You first lobotomize it, by gaining its trust. Then you turn up the heat. See how YouTube went from Ads are optional to Adblockers are immoral.
Google likes it when it can show you more ads, it is not positive human value.
It does not have to have anything ro do with cyberpunk. Corporations are not people, but if they were people, they would be powerful sociopaths. Their interests and anybody elses interests are not the same.
Well, it doesn’t say universal so it’s clearly going to be a specific set of human values and preferences. It’s obviously referring to the preferences of the humans who are footing the bill and who stand to profit from it. The extent to which those values happen to align with those of the eventual consumer of this product could potentially determine whether the aforementioned profits ever materialize.
The title seems kinda misleading, this is from the article (GlobalLogic is the company contracted by Google):
"AI raters at GlobalLogic are paid more than their data-labeling counterparts in Africa and South America, with wages starting at $16 an hour for generalist raters and $21 an hour for super raters, according to workers. Some are simply thankful to have a gig as the US job market sours, but others say that trying to make Google’s AI products better has come at a personal cost."
That argument is as old as any mistreated worker complaining about their situation and as old as any argument against workers rights in general. Anybody not liking their job could just leave right? Simple! No, the world just isn't that simple and it didn't become simpler just because it happens in an AI context that produces a tool you like.
There are lots of jobs out there that suck and people do them anyway. Because the freedom that they supposedly have is not as free as you imagine.
What explains not changing jobs because you find it distressing and claiming that you're being paid below what you're worth? It seems like if that were true, then you'd be motivated to find a job that pays market rate. And if you couldn't, then you could at least find another job that pays less than market rate, like your current job, but isn't so distressing.
Maybe these people are trying to keep their skills and degrees honed somehow in a bad market, rather than going straight for a less-distressing-but-also-lower-paying job that does nothing to their skillset.
So what specific rights do you think they should have that they don't right now?
They're making more money than minimum wage. They're free to leave. It's not violating any safety regulations. There aren't any complaints of harassment.
So what precisely is the complaint here around worker's rights?
When they switch to aligning with algorithms instead of humans we'll get another story about how terrible it was that they removed the jobs that were terrible when they existed.
This doesn't sound as bad to me as the Facebook moderator job or even a call center job, but it does sound pretty tedious.
Congratulations, you just described most jobs. And many backbreaking laborers make about the same or less, even in the U.S., not to mention the rest of the world.
Can you believe that companies would ask people to do things they normally wouldn't in exchange for money!?
These types of articles always have an elitist view of the workers hired. That's a big source of the right (in the US) despising the left. The left don't say it directly, but when they talk about how shitty their town is and how the job they have is exploitative, there's an implicit judgment on the persons who live/work there.
From my shallow understanding, it seems that human training is involved heavily in the post-training/fine-tuning stage, after the base model has been solidified already.
In that case, how is the notion of truthiness (what the model accepts as right or wrong) affected during this stage , that is affected by human beings vs. it being sealed into the basic model itself, that is truthiness being deduced by the method / part of its world model.
Their work doesn’t seem that bad. This article tries really hard to portray that a simple freelance desk job is somehow literally exploitation or something.
Lots of people would do anything to get such work.
Diminishing returns is an ugly business. And thats obviously where we are at. The end not the beginning of LLM "innovation".
Any technology that creates "sysiphian" tasks, is not worth anyones time. That includes LLMs, and "Big Data". The "herculean effort" that never ends is the proof in the pudding. The tech doesnt work.
Its like using machine learning for self driving instead of having an actual working algorythm. Your bust.
Scale AI’s entire business model was using people in developing countries to label data for training models. Once you look into it, it comes across as rather predatory.
Couple of months ago I received a job invite for Kotlin AI trainers from the team at Upwork. I asked what the job is about and she says something like "for the opportunity to review & evaluate content for generative AI." And I'm from a developed country too.
Karen Hao's recent book "Empire of AI" about the rise of OpenAI goes into detail how people in Africa and South America were hired (and arguably exploited) for their training efforts.
According to the book they kept dropping the rates paid per item forcing people to work ridiculous 12+ hours/day just to get enough to live on, even in the low cost of living places they were in. It was like something in a cyberpunk dystopia but real.
There's nontrivial historical precedent for this exact playbook: when a new paradigm (Lisp machines and GOFAI search, GPU backprop, softmax self-attention) is scaling fast, a lot of promises get made, a lot of national security money gets involved, and AI Summer is just balmy.
But the next paradigm breakthrough is hard to forecast, and the current paradigm's asymptote is just as hard to predict, so it's +EV to say "tomorrow" and "forever".
When the second becomes clear before the first, you turk and expert label like it's 1988 and pray that the next paradigm breakthrough is soon, you bridge the gap with expert labeling and compute until it works or you run out of money and the DoD guy stops taking your calls. AI Winter is cold.
And just like Game of Thrones, no I mean no one, not Altman, not Amodei, not Allah Most Blessed knows when the seasons in A Song of Math and Grift will change.
There's a YouTube video titled "AI is a hype-fueled dumpster fire" [0] that mentions OpenAI's shenanigans. I haven't fact checked that but I've heard enough stories to believe it.
It's strange that the Guardian mentions OpenAI's "O3" model and not GPT-5.
Maybe they think o3 is SOTA still, but they should at least name it correctly, in lowercase as OpenAI does.
"Google" posted a job opening. They applied for and took the job, agreeing to posted pay and conditions. End of the story. It's not up to the Guardian to decide
The way you defend against an article citing "thousands of workers" by using a nitpicky criticism about grammar style makes me suspect that it raises a cognitive dissonance in your head that you are not ready to address yet.
Yeah they should simply buy widgets from the abundance of other widget sellers since this is a perfectly competitive market with no transaction costs and perfectly symmetric information
I'm a contractor for one of these companies. It pays okay ($45+/hour) if you can pass qualifications for your area of expertise but the work isn't steady and communication is non-existent. The coding qualifications I did were difficult FAANG algorithm analysis questions. The work has definitely gotten harder over the last year and often says we need to come up with Masters/PhD level work or problems that someone with 5+ years of experience in a field would have difficulty solving. I wish I had a regular job but I live in rural North Carolina and remote work is hard to come by.
I wouldn't mind this work at that pay, being particularly strong in leetcode and in CS itself.
How do I join?
Look up Mercor, DataAnnotation.tech, and Outlier. You create a profile, upload a resume, and do some required tasks for each job posting they have. It may involve a combination of interviewing with an AI, doing a few trial tasks, and submitting a portfolio or Github profile.
hmm, this feels like ScaleAI
Is something stronger than your wish to get a regular job tying you to where you currently live?
I just want to note that asking this question implies an openness to one’s personal affairs that may not be appropriate in an anonymous, public setting. A person offering context and insight to a topic is not necessarily an invitation to an for more personal contexts and insights.
I understand it's personal, but I also recognize they went out of their way to bring it up. Some people, including me, are more willing to discuss things anonymously because it adds a layer of impersonality. This is just a discussion board. If OP doesn't answer, that's ok. I don't ever think I'm entitled a response.
It is a reasonable question that also emphasizes the composite cost of decisions.
Personally I would love to live in a more rural place, but until I am self sufficient enough, this is not an opportunity I am willing to take.
I only started seriously looking for work again about a month ago. I'd like to stay in this area for a few reasons but I would relocate if necessary. I worked remotely from 2015 until a layoff in late 2023 and this was the first thing I came across after that. It was okay for awhile and actually pretty interesting at first but the hours aren't reliable and there doesn't seem to be much opportunity for getting promoted.
At least a few of these anecdoates are worrying:
> “At first they told [me]: ‘Don’t worry about time – it’s quality versus quantity,’” she said.
> But before long, she was pulled up for taking too much time to complete her tasks. “I was trying to get things right and really understand and learn it, [but] was getting hounded by leaders [asking], ‘Why aren’t you getting this done? You’ve been working on this for an hour.’”
And:
> Dinika said he’s seen this pattern time and again where safety is only prioritized until it slows the race for market dominance. Human workers are often left to clean up the mess after a half-finished system is released. “Speed eclipses ethics,” he said. “The AI safety promise collapses the moment safety threatens profit.”
Finally:
> One work day, her task was to enter details on chemotherapy options for bladder cancer, which haunted her because she wasn’t an expert on the subject.
Something I'd be interested to understand is how widespread this practice is. Are all of the LLMs trained using human labor that is sometimes exposed to extreme content?
There are a whole lot of organizations training competent LLMs these days in addition to the big three (OpenAI, Google, Anthropic).
What about Mistral and Moonshot and Qwen and DeepSeek and Meta and Microsoft (Phi) and Hugging Face and Ai2 and MBZUAI? Do they all have their own (potentially outsourced) teams of human labelers?
I always look out for notes about this in model cards and papers but it's pretty rare to see any transparency about how this is done.
> Are all of the LLMs trained using human labor that is sometimes exposed to extreme content?
The business process outsourcing companies labelling things for AI training are often the same outsourcing companies providing moderation services to facebook and other social media companies.
I need 100k images labelled by the type of flower shown, for my flower-identifying AI, so I contract a business that does that sort of thing.
Facebook need 100k flagged images labelled by is-it-an-isis-beheading-video to keep on top of human reviews for their moderation queues. They contract with the same business.
The outsourcing company rotates workers between tasks, so nobody has to be on isis beheading videos for a whole shift.
> The outsourcing company rotates workers between tasks, so nobody has to be on isis beheading videos for a whole shift.
Is that an assumption on your side, a claim made by the business, a documented process or something entirely different?
It has been documented that human image moderators exist and that some have been deeply traumatized by their work. I have zero doubts that the datasets of content and metadata created by human image moderators are being bought and sold, literally trafficking in human suffering. Can you point to a comprehensive effort by the tech majors to create a freely-licensed dataset of violent content and metadata to prevent duplication of human suffering?
One of the key innovations behind the DNN/CNN models was Mechanical Turk. OpenAI used a similar system extensively to improve the early GPT models. I would not be surprised that the practice continues today; NN models needs a lot of quality ground truth training data.
Right, but where are the details?
Given the number of labs that are competing these days on "open weights" and "transparency" I'd be very interested to read details of how some of them are handling the human side of their model training.
I'm puzzled at how little information I've been able to find.
This is not going to be as deep/specific as you want but a starting point from one of the companies that handles this sort of work is here: https://humandata.mercor.com/mercors-approach/black-box-vs-o...
I read this a few years ago.
Time Exclusive: OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic
https://time.com/6247678/openai-chatgpt-kenya-workers/
Beyond that, I think the reason you haven't heard more about it is that it happens in developing countries, so western media doesn't care much, and also because big AI companies work hard to distance themselves from it. They'll never be the ones directly employing these AI sweatshop works, it's all contracted out.
Good article from 2023, not much data though if that's what you're looking for:
https://nymag.com/intelligencer/article/ai-artificial-intell...
unwalled: https://archive.ph/Z6t35
Generally seems similar today just on a bigger Scale. And much more focus on coding
Here in the US DataAnnotation seems to be the most marketed company offering these jobs
I have been a generalist annotator for some of the others you mentioned, due to NDA will not specify which. I would venture to guess that basically all major models use some degree of human feedback if there is money coming in from somewhere.
I've shared this once on HN before, but it's very relevant to this question and just a really great article so I'll reshare it here:
https://www.theverge.com/features/23764584/ai-artificial-int...
it explores the world of outsourced labeling work. Unfortunately hard numbers on the number of people involved are hard to come by because as the article notes:
"This tangled supply chain is deliberately hard to map. According to people in the industry, the companies buying the data demand strict confidentiality. (This is the reason Scale cited to explain why Remotasks has a different name.) Annotation reveals too much about the systems being developed, and the huge number of workers required makes leaks difficult to prevent. Annotators are warned repeatedly not to tell anyone about their jobs, not even their friends and co-workers, but corporate aliases, project code names, and, crucially, the extreme division of labor ensure they don’t have enough information about them to talk even if they wanted to. (Most workers requested pseudonyms for fear of being booted from the platforms.) Consequently, there are no granular estimates of the number of people who work in annotation, but it is a lot, and it is growing. A recent Google Research paper gave an order-of-magnitude figure of “millions” with the potential to become “billions.” "
I too would love to know more about how much human effort is going into labeling and feedback for each of these models, it would be interesting to know.
That was indeed a great article, but it is a couple of years old now. A lot of of the labeling work described there relates to older forms of machine learning - moderation models, spam labelers, image segmentation etc.
Is it possible in 2025 to train a useful LLM without hiring thousands of labelers? Maybe through application of open datasets (themselves based on human labor) that did not exist two years ago?
Good question, I don't personally know. The linked article would suggest there are plenty of people working on human feedback for chatbots, but that still doesn't give us any hard numbers or any sense of how the number of people involved is changing over time. Perhaps the best datapoint I have is that revenue for SurgeAI (one of many companies that provides data labeling services to Google and OpenAI among others) has grown significantly in recent years, partly due to ScaleAI's acquisition by Meta, and is now at $1.2 billion without having raised any outside VC funding:
https://finance.yahoo.com/news/surge-ai-quietly-hit-1b-15005...
Their continued revenue growth is at least one datapoint to suggest that the number of people working in this field (or at least the amount of money spent on this field) is not decreasing.
Also see the really helpful comment above from cjbarber, there's quite a lot of companies providing these services to foundation model companies. Another datapoint to suggest the number of people working providing labeling / feedback is definitely not decreasing and is more likely increasing. Hard numbers / increased transparency would be nice but I suspect will be hard to find.
So why do you think asking this question here would yield a satisfying answer, especially how the HN community likes to dispute any vague conclusions for anything as hyped as AI training?
To counter your question, what makes you think that's not the case? Do you think Mistral/Moonshot/Qwen/etc. are all employing their own data labelers? Why would you expect this kind of transparency from for-profit bodies that are evaluated in the billions?
If you don't ask the question you'll definitely not get an answer. Given how many AI labs follow Hacker News it's not a bad place to pose this.
"what makes you think that's not the case?"
I genuinely do not have enough information to form an opinion one way or the other.
> If you don't ask the question you'll definitely not get an answer.
Sure, but the way you're formulating the question is already casting an opinion. Besides, no one could even attempt to answer your questions without falling into the trap of true diligence... one question just asks how all (with emphasis!) LLMs are trained:
> Are all of the LLMs trained using human labor that is sometimes exposed to extreme content?
Who in the world would even be in such a position?
That question could be answered by proving the opposite: if someone has trained a single competent LLM without any human labor that was exposed to extreme content then not all LLMs were trained that way.
I previously made a list on twitter of some data labeling startups that work with foundation model companies.[1] Here's the RLHF provider section:
RLHF providers:
1. Surge. $1b+ revenue bootstrapped. DataAnnotation is the worker-side (you might've seen their ads), also TaskUp and Gethybrid.
2. Scale. The most well known. Remotasks and Outlier are the worker-side
3. Invisible. Started as a kind of managed VA service.
4. Mercor. Started mostly as a way to hire remote devs I think.
5. Handshake AI. Handshake is a college hiring network. This is a spinout
6. Pareto
7. Prolific
8. Toloka
9. Turing
10. Sepal AI. The team is ex-Turing
11. Datacurve. Coding data.
12. Snorkel. Started as a software platform for data labeling. Offers some data as a service now.
13. Micro1. Also started as a way to hire remote contractor devs
[1]: https://x.com/chrisbarber/status/1965096585555272072
This is great!
Are there companies that focus on labeling of inputs rather than RLHF of outputs?
Yes, there are quite a few that do that. Appen, iMerit, TELUS, etc. Also Scale AI started focused on input annotation I think for self driving.
"Google said in a statement: “Quality raters are employed by our suppliers and are temporarily assigned to provide external feedback on our products. Their ratings are one of many aggregated data points that help us measure how well our systems are working, but do not directly impact our algorithms or models.” GlobalLogic declined to comment for this story." (emphasis mine)
How is this not a straight up lie? For this to be true they would have to throw away labeled training data.
Because they are doing it to compute quality metrics not to implement RLHF. It’s not training data.
Every decision they take based on evals influences the model.
> For this to be true they would have to throw away labeled training data.
That's how validation works.
Is there a reason not to use validation data in your next round of training data? Or is it more efficient to reuse validation and instead get more training data?
You'd have to recreate your validation if you trained your model on it every iteration and then they wouldn't be consistent enough to show any trends
Key word: "directly"
It does so indirectly, so it's a true albeit misleading statement.
They probably don’t do it at a scale large enough to do RLHF with it, but it’s still useful feedback the people working on the projects / products.
More recent models actually use "reinforcement learning from AI feedback", where the task of assigning a reward is essentially fed back into the model itself. Human feedback is then only used to ground the training, on selected examples (potentially even entirely artificial ones) where the AI is most highly uncertain about what feedback should be given.
The title is biased, blaming Google for mistreating people and implying that Google's AI isn't smart, but the OP is worth reading, because it gives readers a sense of the labor and cost involved in providing AI models with human feedback, the HF in RLHF, to ensure they behave in ways acceptable to human beings, more aligned with human expectations, values, and preferences.
RLHF (and its evolution, RLAIF) is actually used for more than setting "values and preferences". It's what makes AI models engage in recognizable behavior, as opposed to simply continuing a given text. It's how the "Chat" part of "ChatGPT" can be made to work in the first place.
Yes. I updated my comment to reflect as much. Thank you.
> Sawyer is one among the thousands of AI workers contracted for Google through Japanese conglomerate Hitachi’s GlobalLogic to rate and moderate the output of Google’s AI products...
Depends how you look at it. I think a brand like Google should vet a mere one level down the supply chain.
I had no idea Hitachi was also running software sweatshops.
What is a "human value" and whose preferences?
> to ensure the AI models are more aligned with human values and preferences.
to ensure the AI models are more aligned with Google's values and preferences.
FTFY
I'm a big fan of cyberpunk dystopian fiction, but I still can't quite understand what you're alluding to here. Can you give an example value that google align the AI with that you think isn't a positive human value?
Their entire business model? Making search results worse to juice page impressions? Every dark pattern they use to juice subscriptions like every other SaaS company? Brand lock-in for Android? Paying Apple for prominent placement of their search engine in iOS? Anti-competitive practices in the Play store? Taking a massive cut of Play Store revenue from people actually making software?
How does all of that affect the desired outputs for their LLMs?
You'll see once they figure it out.
Or, if they really figure it out, you’ll only feel it.
"Adtech is good. Adblockers are unnatural"
Google Gemini 2.5 Pro actually has a quite nuanced reply when asked to consider this statement, including the following:
> "Massive privacy invasion: The core of modern adtech runs on tracking your behavior across different websites and apps. It collects vast amounts of personal data to build a detailed profile about your interests, habits, location, and more, often without your full understanding or consent."
You don't boil the frog instantly. You first lobotomize it, by gaining its trust. Then you turn up the heat. See how YouTube went from Ads are optional to Adblockers are immoral.
Google likes it when it can show you more ads, it is not positive human value.
It does not have to have anything ro do with cyberpunk. Corporations are not people, but if they were people, they would be powerful sociopaths. Their interests and anybody elses interests are not the same.
Yes, and one more tweak: the values of Google or anyone paying Google to deliver their marketing or political messaging.
> to ensure the AI models are more aligned with human values and preferences.
And which are these universal human values and preferences ? Or are we talking about silicon valley's executives values ?
Well, it doesn’t say universal so it’s clearly going to be a specific set of human values and preferences. It’s obviously referring to the preferences of the humans who are footing the bill and who stand to profit from it. The extent to which those values happen to align with those of the eventual consumer of this product could potentially determine whether the aforementioned profits ever materialize.
The title seems kinda misleading, this is from the article (GlobalLogic is the company contracted by Google):
"AI raters at GlobalLogic are paid more than their data-labeling counterparts in Africa and South America, with wages starting at $16 an hour for generalist raters and $21 an hour for super raters, according to workers. Some are simply thankful to have a gig as the US job market sours, but others say that trying to make Google’s AI products better has come at a personal cost."
It's employment at will. They are free to go work somewhere else if they don't like it...
That argument is as old as any mistreated worker complaining about their situation and as old as any argument against workers rights in general. Anybody not liking their job could just leave right? Simple! No, the world just isn't that simple and it didn't become simpler just because it happens in an AI context that produces a tool you like.
There are lots of jobs out there that suck and people do them anyway. Because the freedom that they supposedly have is not as free as you imagine.
What explains not changing jobs because you find it distressing and claiming that you're being paid below what you're worth? It seems like if that were true, then you'd be motivated to find a job that pays market rate. And if you couldn't, then you could at least find another job that pays less than market rate, like your current job, but isn't so distressing.
This definitely comes from someone who never had trouble looking for a job and cannot possibly understand how hard real life is for other people.
Maybe these people are trying to keep their skills and degrees honed somehow in a bad market, rather than going straight for a less-distressing-but-also-lower-paying job that does nothing to their skillset.
So what specific rights do you think they should have that they don't right now?
They're making more money than minimum wage. They're free to leave. It's not violating any safety regulations. There aren't any complaints of harassment.
So what precisely is the complaint here around worker's rights?
I think the complaint is that they should have safety regulations. the impact of these gigs on mental health is well documented
When they switch to aligning with algorithms instead of humans we'll get another story about how terrible it was that they removed the jobs that were terrible when they existed.
This doesn't sound as bad to me as the Facebook moderator job or even a call center job, but it does sound pretty tedious.
> [job] … has come at a personal cost.
Congratulations, you just described most jobs. And many backbreaking laborers make about the same or less, even in the U.S., not to mention the rest of the world.
Can you believe that companies would ask people to do things they normally wouldn't in exchange for money!?
These types of articles always have an elitist view of the workers hired. That's a big source of the right (in the US) despising the left. The left don't say it directly, but when they talk about how shitty their town is and how the job they have is exploitative, there's an implicit judgment on the persons who live/work there.
From my shallow understanding, it seems that human training is involved heavily in the post-training/fine-tuning stage, after the base model has been solidified already.
In that case, how is the notion of truthiness (what the model accepts as right or wrong) affected during this stage , that is affected by human beings vs. it being sealed into the basic model itself, that is truthiness being deduced by the method / part of its world model.
Their work doesn’t seem that bad. This article tries really hard to portray that a simple freelance desk job is somehow literally exploitation or something.
Lots of people would do anything to get such work.
Diminishing returns is an ugly business. And thats obviously where we are at. The end not the beginning of LLM "innovation".
Any technology that creates "sysiphian" tasks, is not worth anyones time. That includes LLMs, and "Big Data". The "herculean effort" that never ends is the proof in the pudding. The tech doesnt work.
Its like using machine learning for self driving instead of having an actual working algorythm. Your bust.
Are other AI companies doing the same thing? Would like to see more articles about this...
Scale AI’s entire business model was using people in developing countries to label data for training models. Once you look into it, it comes across as rather predatory.
This was one of the first links I found re: Scale’s labor practices https://techcrunch.com/2025/01/22/scale-ai-is-facing-a-third...
Here’s another: https://relationaldemocracy.medium.com/an-authoritarian-work...
Couple of months ago I received a job invite for Kotlin AI trainers from the team at Upwork. I asked what the job is about and she says something like "for the opportunity to review & evaluate content for generative AI." And I'm from a developed country too.
Karen Hao's recent book "Empire of AI" about the rise of OpenAI goes into detail how people in Africa and South America were hired (and arguably exploited) for their training efforts.
Can you explain the exploited part?
My understanding is they performed work and were paid for it at market rate. So just regular capitalism. Or was there more to it?
[delayed]
According to the book they kept dropping the rates paid per item forcing people to work ridiculous 12+ hours/day just to get enough to live on, even in the low cost of living places they were in. It was like something in a cyberpunk dystopia but real.
There's nontrivial historical precedent for this exact playbook: when a new paradigm (Lisp machines and GOFAI search, GPU backprop, softmax self-attention) is scaling fast, a lot of promises get made, a lot of national security money gets involved, and AI Summer is just balmy.
But the next paradigm breakthrough is hard to forecast, and the current paradigm's asymptote is just as hard to predict, so it's +EV to say "tomorrow" and "forever".
When the second becomes clear before the first, you turk and expert label like it's 1988 and pray that the next paradigm breakthrough is soon, you bridge the gap with expert labeling and compute until it works or you run out of money and the DoD guy stops taking your calls. AI Winter is cold.
And just like Game of Thrones, no I mean no one, not Altman, not Amodei, not Allah Most Blessed knows when the seasons in A Song of Math and Grift will change.
There's a YouTube video titled "AI is a hype-fueled dumpster fire" [0] that mentions OpenAI's shenanigans. I haven't fact checked that but I've heard enough stories to believe it.
[0] https://youtu.be/0bF_AQvHs1M?si=rpMG2CY3TxnG3EYQ
It's strange that the Guardian mentions OpenAI's "O3" model and not GPT-5. Maybe they think o3 is SOTA still, but they should at least name it correctly, in lowercase as OpenAI does.
"Google" posted a job opening. They applied for and took the job, agreeing to posted pay and conditions. End of the story. It's not up to the Guardian to decide
I agree, article is pretty low quality ragebait. Not good journalism at all.
It is amazing how much their quality levels have fallen during the past two decades.
I used to point to their reporting as something that my nation’s newspapers should seek to emulate.
(My nation’s newspapers have since fallen even lower.)
Not so easy. What if you get hired as a physiotherapist somewhere but on your first day you find out you will work in a brothel?
Or join an hospital as nurse, but then you are asked to perform surgery as you were a doctor?
There are serious issues outlined in the article.
This is not what the article is outlining.
If they're underpaid and overworked, by definition words that are relative to other options, they should go to one of the better options.
The way you defend against an article citing "thousands of workers" by using a nitpicky criticism about grammar style makes me suspect that it raises a cognitive dissonance in your head that you are not ready to address yet.
Glad to learn from your post that the labor market has recently become perfectly competitive and efficient.
Yeah they should simply buy widgets from the abundance of other widget sellers since this is a perfectly competitive market with no transaction costs and perfectly symmetric information
It seems a deja vu of previous Amazon's Mechanical Turk[1] discussions[2] but with AI.
[1] https://www.mturk.com/
[2] https://tinyurl.com/4r2p39v3
This definitely explains why Google’s AI Search Results is so bad at what it purports to do.
with wages starting at $16 an hour for generalist raters and $21 an hour for super raters, according to workers
That’s sort of what I expect the Guardian’s UK online non-sub readers to make.
Perhaps GlobalLogic should open a subsidiary in the UK?
In many things "AI" is just another form exploiting the poor to make the rich even wealthier. A form of digital colonialism.
AI means actual indians, did we not learn that from the initial OpenAI GPT 3.0 training? It made it to HN.
Gemini is faked.
How this industry managed to not grasp that meaning exists entirely separate from words is altogether bizarre.