AI for humanity; but whose humanity?

Adriana Leos
Jul 21
12 min read

OpenAI states that its mission is to ensure that “artificial general intelligence will benefit all of humanity.”

That’s one of the most ambitious promises ever made by a technology company. And although I do sincerely hope they are right, history makes me skeptical.

Every technological revolution has come with a promise of greater prosperity, greater freedom, or greater equality. The printing press promised access to knowledge, the automobile promised mobility and independence, and the internet promised democratized information.

Each of those technologies has transformed society. However, none of them transformed it equally. That’s because technology has never existed outside of politics, economics, or power; instead, it amplifies the systems in which it was introduced.

Which brings us back to artificial intelligence (AI).

AI may produce enormous benefits, but those benefits are unlikely to be distributed equally because they are being built within existing systems of economic, political, and social inequality.

If AI truly intends to benefit “all of humanity,” then we need to ask: who exactly is “humanity?”

Who builds it?

One of the biggest misconceptions about AI is that it simply teaches itself.

When in reality, thousands and thousands of people have spent years performing the invisible labor that enables these systems to function. And yet, these people will never receive the recognition that they deserve, simply because most of the general population don’t even know that these data workers exist.

Yet, modern AI would not exist without them.

Big tech companies, including OpenAI, Meta, Google, and Anthropic, have relied on outsourcing data annotation tasks to workers in countries like Kenya, India, Venezuela, the Philippines, and many more in the Global South.

The tasks include things like labeling, “cleaning,” and evaluating massive amounts of data so they can turn around and use this data to train their AI models.

For example, content moderation is a technique where workers review massive amounts of material, which can include extremely graphic content like child sexual abuse, murder, torture, racism, and much more gruesome material.

Workers are responsible for reviewing this graphic material and then categorizing it into various genres like “threat,” “self-harm,” “hate speech,” etc.,. This work teaches models what not to generate and what to refuse when given certain prompts.

Another process is called Reinforcement Learning from Human Feedback (RLHF). This is a technique that uses human judgement to train models. For instance, workers are given two examples: “Answer A” and “Answer B.” They have to compare these two answers and then determine: which of the two are more truthful, useful, less biased, which sounds more natural, and which follows the instructions of the prompt better.

Workers are also responsible for tasks such as “cleaning” datasets by removing duplicate content, spam, and other irrelevant information. They also evaluate language for things like grammar, fluency, and translation quality.

Finally, workers are also responsible for “labeling” the data.

If you’re familiar with Google’s reCAPTCHA where it asks users to do things like “select all the images with crosswalks” or “identify the stop signs,” then you are familiar with what it means to label data.

All of these tasks help to improve a model’s behavior.

And as you can see, there is a lot of work that goes into having to train these models. And the irony is that AI tends to market itself as automation while simultaneously depending on massive amounts of hidden manual labor.

This work is also done for extremely low wages.

A labeling task that might cost $20–$40 per hour in the U.S. can often be contracted for a small fraction of that in lower-income countries (and big tech companies know this).

Therefore, the concern becomes not only the low pay, but also the prolonged exposure to extremely disturbing material without sufficient psychological support (as most of these companies do not provide health care for employees to seek therapy). Consequently, this work has been linked to worsened mental health.

You can see where this creates an imbalance: not only are these workers receiving extremely low pay, while also actively deteriorating their mental health, but simultaneously helping to improve systems that contribute to products worth billions of dollars.

Hidden labor is only one way existing inequalities become embedded into AI. As we will see, even after these systems are trained, they continue to inherit the biases of the societies that created them.

Who gets seen?

In her book, Race After Technology, author Ruha Benjamin, asks us to consider the fact that algorithms that are developed in places like Japan, China or South Korea will be able to recognize East Asian faces far more easily than they will Caucasians.

Whereas, algorithms that are developed in places like Germany or the United States will be significantly better at recognizing Caucasian faces and features.

This is due to the fact that the engineer that develops the algorithm will more than likely program it to focus on specific features. And this is more than likely done subconsciously, (because we are naturally influenced by our environment and those we surround ourselves with).

And we have seen this play out in real time when, as a graduate student at MIT, Joy Buolamwini encountered a problem when facial analysis software couldn’t detect her face, yet it detected the faces of people with lighter skin without a problem. And it wasn’t until Buolamwini put on a white mask that the computer detected her.

Joy Buolamwini 'the coded gaze'; white mask example — Joy Buolamwini demonstrates a facial recognition software that recognizes her face only when she wears a white mask.

This becomes dangerous when these software programs are implemented in places like police departments all over the country.

These departments use this software to then compare “real time individuals to “hot lists” that are disproportionately filled with Black people – and [Black people] also happen to be the least recognizable figures in the world of facial recognition software.” (Race After Technology, p 113).

Let’s also consider that there are now new AI systems for vetting job applicants.

For instance, a company called HireVue states on their website that they were founded based on “the idea that humans are more than just bullet points on a resume.”

And yet, they offer an AI interviewer service that you can implement into your business that will “combine science and voice… to reveal who meets the demands for every role earlier in the hiring journey.”

As Benjamin states, this might be “magical for employers, perhaps, looking to streamline the grueling work of recruitment, but a curse for many job seekers” (p 142).

We have also seen this play out when, in 2014, Amazon had a team of engineers that were working on a project that would help to automate hiring for the company. But they soon discovered that the tool systematically discriminated against women that were applying for technical jobs.

It’s now becoming increasingly more difficult for job seekers to even get past the point of automation in order to speak to a real human. This, in my opinion, is so unfortunate because this means that companies are more than likely missing out on thousands of incredibly talented and capable applicants.

Thankfully, Amazon stopped using the software program, but this is just another clear example of how we have already seen and experienced algorithmic bias.

AI is a symbiotic process in that it learns from us just as much as we learn from it. Meaning, it simply reproduces what it’s being trained on. If algorithms are trained on existing data, then that data is a reflection of existing bias.

But as we will see, even perfectly fair algorithms would not simply solve the problem because they still exist inside an economy where not everyone has the same ability to adapt.

The single parent

I’ve listened to several podcasts episodes from The Diary of a CEO with Steven Bartlett. The narrative you often hear on the episodes covering AI is that ‘everyone will simply be able to seamlessly adapt and use AI to become an entrepreneur, build a business, and thrive in this new world.’

But in my opinion, this is a limited perspective.

Growing up, my mother was a single mom of me and my four siblings. Being the youngest, I got to see how much limited time and energy my mother had to expend on not only her three different jobs, but also her five children.

These conversations tend to assume that everyone has free evenings, childcare, reliable access to the internet, disposable income, education, and the emotional bandwidth.

Many people do not.

I say that these conversations stem from a limited perspective because it completely disregards the single parents, the caregivers, disabled people, people working two (or more) jobs, refugees, or people living paycheck to paycheck simply trying to survive.

Many people do not have the luxury to reinvent themselves every time technology changes so I genuinely want to know how do we expect a tech-illiterate, single mom of five kids, to stop what she’s doing and sit on a computer (given she has access to one) to learn AI in order to all of a sudden become a savvy business strategist and build a business from scratch?

Now I’m not saying this is impossible – women are incredible and do hard things every single day – but the chances of this becoming a reality for everyone are probably pretty low. Because history has shown us that technological revolutions often reward those who already possess time, education, capital, and networks.

And adaption also assumes that there will still be meaningful work available to adapt into.

Who gets left behind?

I have two brothers. One of which is 12 years older than me. He was 12 years old when he was riding his bike down the street and was struck by a car. This accident caused him to suffer a life-long brain injury.

Because of his disability, he is only capable of performing efficiently at so many roles. And while he’s always had a job, most of them are the type of jobs that most people tend to overlook; such as a grocery bagger, a movie ticket seller, or a cashier.

Automation rarely eliminates all jobs. Instead, it tends to eliminate particular jobs first.

And historically, these jobs tend to be those that are repetitive, low wage, based in customer service, administrative, or entry level.

Those jobs also tend to be where people like teenagers, disabled workers, immigrants, older adults, and/or people without a college degree often enter the workforce.

We’ve already seen this play out in real time as jobs such as cashiers have been replaced for self-checkout, box office ticket sellers have been replaced by kiosks and online sales, customer service representatives have been replaced by automated chatbots.

Not only is it unfair to the those who are being directly impacted and have little to nowhere else to turn to, but this also begs the question: if the first rung of the economic ladder disappears, how does anyone climb to the second?

Finally, let’s discuss how entry-level workers aren’t the only one who are bearing the cost of AI.

Who pays the environmental cost?

One of the greatest marketing successes has been convincing us that “the cloud” exists somewhere above us or this idea that it’s weightless, intangible, and invisible.

When it is anything but.

For instance, the moment you type in a prompt into a chatbot like ChatGPT or Claude, it travels through miles of underground fiber optic cables until it reaches a data center. A data center is a massive industrial warehouse filled with thousands upon thousands of computers stacked inside metal cabinets or server racks. Those computers perform billions of calculations in just a matter of seconds.

Every single one of those calculations generates a ton of heat.

Imagine leaving your laptop running all day with 40 different tabs open. Now imagine thousands of high-powered computers running billions of calculations 24-hours a day, seven days a week.

The heat they produce becomes so intense that cooling systems have to work constantly just to prevent them from overheating.

Those cooling systems require a significant amount of electricity.

In fact, the International Energy Agency (IEA) estimates that data centers accounted for roughly 1.5% of global electricity consumption in 2024, and projects that demand could double by 2030 – with AI being the main driver behind that increase in consumption.

Researchers have even found that some of these facilities create what they call “heat islands.” Which essentially means that these facilities raise the surrounding land temperatures from anywhere around 3.6 degrees to 16 degrees Fahrenheit.

But the question isn’t simply how much electricity AI uses – it’s where the electricity comes from and who bears the environmental consequence of generating it.

One study examining more than 2,100 data centers in the U.S. estimated that more than half of the electricity supplying them came from fossil-fuels, resulting in more than 105 million tons of carbon dioxide emissions.

Once again, we see a pattern emerging – the benefits of AI are distributed globally; meanwhile, the environmental costs are not.

Now let’s consider water.

AI uses water in a few different ways to power its operations. First, water is used at the data center itself to cool down servers and other facility operations. Secondly, power plants also consume water to generate electricity used by the facility. And third, water is also used to manufacture the materials needed to build the data center itself.

Now consider that data centers are built in communities that are already facing declining groundwater, drought conditions, threatened rivers and other bodies of water, and things like agricultural dependence.

This can create true imbalance because here we have a technology company that profits billions and serves customers around the world, but one local community has to bear the pressure on its natural resources, aquifer, pipes, wastewater system, and its electrical grid.

For instance, in 2022, Arizona faced a severe water crisis. The drought drained the Colorado River to dangerously low levels. Meanwhile, Microsoft laid the groundwork for a $100 billion facility.

Next, let’s look at natural resources.

Before a server ever reaches a data center, its materials must first be extracted from the Earth.

Chile, for example, is the world’s largest producer of copper and accounts for a quarter of the global supply. Chile also produces roughly a third of the world’s lithium.

These conditions therefore make it an ideal place for tech companies to come in and extract these resources as local and Indigenous communities are forced to watch their land be ripped apart. In 2022, the European Union set new policies around energy, and as a result, the demand for lithium skyrocketed.

Consequently, politicians in Chile said that the country’s mining industry would help propel the country into a brighter future. Meanwhile Indigenous communities were left asking: “a better future for whom?”

“Local people never have the ability to think about their own destiny outside the forces of economics and international politics,” (Empire of AI, p 283).

Finally, building the data centers themselves have an environmental cost.

These facilities require enormous amounts of land, concrete foundations, backup generators, sewer connections, among many other types of infrastructure. Additionally, these facilities consume so much land that, depending on where they’re located, development can also replace things like farmland, forest, grassland, or wildlife habitats.

And because data centers require stability in order to avoid disrupting their operations, they typically install diesel-based generators for backup power. However, the diesel combustion can release things like carbon monoxide, carbon dioxide, nitrogen oxides, (as well as other harmful pollutants), which cause significant respiratory issues, as well as exacerbate the risk of heart and lung diseases, and certain cancers.

Overall, discussions about AI tend to focus on what happens after someone types in a prompt into ChatGPT.

But the environmental cost begins long before that prompt is ever even written.

It begins in the mine — and it continues through factories, construction sites, electrical grids, and water systems — before ever arriving inside a data center.

Once again we can see how the benefits are distributed globally when the environmental costs are concentrated locally.

If AI is meant to benefit all of humanity, then we cannot ignore the communities that are asked to sacrifice their land, water, air, and natural resources to make that future possible.

Conclusion

Economists define an “externality” as a cost imposed on someone who didn’t choose it, and who doesn’t receive the benefit.

And that’s what appears to be happening here.

As I stated earlier, AI may produce enormous benefits, but those benefits are unlikely to be distributed equally because they are being built within existing systems of economic, political, and social inequality.

Therefore, we must ask ourselves: who owns AI? And if AI is supposed to dramatically increase productivity – who benefits from that productivity?

Is it the workers, government, communities, tech leaders, shareholders?

History has shown us time and time again that ownership matters. For instance, the internet created billions, yet, most of that wealth (and power) is accumulated to relatively few companies. And at this rate, AI may be following the same pattern.

If AI becomes everything from information, education, medicine, law, coding, government, science – then whoever controls AI controls knowledge infrastructure.

I don’t doubt that AI will change the world, it probably will, and in many ways, already has.

What I doubt is the assumption that technological progress automatically produces human progress. Technology isn’t neutral because power is never neutral. History has never worked that way.

Every major technological innovation has forced us to ask questions about labor, ownership, socio-economic status, power, and justice.

This era of AI is no different.

If we genuinely want AI to “benefit all of humanity,” then we better ask these important questions now, and not after the technology has already been woven into every fabric of our lives and society.

Who owns it? Who builds it? Who benefits? Who profits? Who is displaced by it? What do we do about the environmental impact that will eventually impact each and every single one of us?

And most importantly, when we say “all of humanity,” who exactly are we referring to?

Sources: