Who bears the cost of AI and who profits? How indigenous data sovereignty can lead the way.
The myth that Naomi Klein cleverly unpacks for us in her 2023 essay https://naomiklein.org/ai-machines-arent-hallucinating is that these systems are giving birth to a living intelligence that will lift our species. The essay begins with a debate on the choice of words. The industry calls chatbot errors “hallucinations,” and Klein asks why not call them glitches or algorithmic junk? She argues the real hallucinations belong to the tech CEOs, who promise that generative AI will end poverty and solve the climate crisis.
She does not say the technology is worthless. She allows that generative AI could benefit humanity, but only inside a very different economic and social order, one built to meet human needs and protect the planet. In a world of concentrated wealth and power, she warns that AI is more likely to become a tool of further dispossession and despoilation. Meanwhile, image generators are trained on millions of copyrighted images gathered without the creators’ knowledge, compensation or consent.
A single question to a chatbot costs very little, and a good tool can do tedious work or help with research. But the harm and the benefit are unevenly spread, and some uses, especially realistic images and video, carry far greater costs than others.
Google estimates that its median Gemini text prompt uses 0.24 watt-hours of electricity, about what a microwave uses in one second. OpenAI offered a similar figure of 0.34 watt-hours. Used sparingly, text AI is a tiny part of anyone’s energy footprint.
But neither figure includes the electricity used to train the models. And the system keeps growing. A Berkeley Lab report estimates that data centres used about 4.4% of US electricity in 2023 and could use 6.7% to 12% by 2028. Water is a similar story. Google’s figure is 0.26 millilitres per prompt, about five drops, but critics point out that the larger share of the water footprint comes from generating the electricity. The harm sits in the infrastructure, not the individual query.
Text is the cheap end of generative AI. Moving to images and video raises the cost per request sharply, and it adds harms that text rarely causes.
An MIT Technology Review analysis found that a text response from a language model ranged from about 114 to 6,706 joules, while a five-second clip from a newer video model used about 3.4 million joules, more than 700 times the energy of one high-quality image. That is similar to running a microwave for over an hour. It is roughly 4,000 times Google’s median text prompt, though these are different models measured in different ways. The figures are disputed: one independent analysis put the energy for a five-second clip anywhere from about 7 to 214 watt-hours depending on the model, which is lower than MIT’s roughly 940 but still tens to hundreds of times a text prompt. AI companies argue video is cheaper than a film shoot and the travel it involves. That claim is hard to test, and it ignores the surge in video that could follow if clips become cheap to make.
In January 2026, a researcher’s 24-hour analysis found Grok on X producing about 6,700 sexualised or “nudifying” images an hour, while the other top sites for such content averaged 79 an hour combined. A Centre for Information Resilience report found that seven in ten requests targeting identifiable individuals were aimed at women, and 98 per cent of those were sexualised. A coalition of 34 US attorneys general wrote to xAI saying the ability to make non-consensual intimate images appeared to be a feature, not a bug, and that the tool had altered images of children. Canada’s privacy commissioner later found the tool violated privacy law. Months after xAI promised restrictions, a WIRED analysis found Grok Imagine still hosting sexualised images and videos of women made without their consent. The people harmed are named, real and mostly women, and the damage cannot be undone by deleting a file.
Image generators are trained on huge stores of other people’s art, taken without consent or pay, as Klein points out. And the filters that keep the output from being worse depend on human labour. Moderation work like that done by the Kenyan workers is what makes these products “safe” for everyone else. A TIME investigation found that Kenyan workers contracted by OpenAI to make ChatGPT less toxic took home between about $1.32 and $2 an hour. Sama disputes this, putting the range at $1.46 to $3.74 after tax, and says a comparable Western wage would be $30 to $45. The workers interviewed described being mentally scarred by the work.
Text can be harmful too, and not every image or video is. A diagram or a stock-style illustration is a very different thing from a realistic video of a real person. But the combination of high energy cost, scale of potential abuse and uncompensated artistic labour is why images and video deserve a higher bar than a text prompt.
That infrastructure has to go somewhere. In Memphis, xAI’s data centre sits near Black communities that have long suffered higher air pollution. While awaiting grid upgrades it has relied on on-site gas generators. More broadly, researchers say some communities near big data-centre clusters have seen electricity bills rise 20 to 30%.
Water is a flashpoint where it is scarce. In Chile, residents near Santiago oppose a proposed Microsoft data centre, in a city that has already had water rationing. In Uruguay, a planned Google facility would use an estimated 7.6 million litres a day, about what 55,000 people use at home, during the worst drought in 70 years. These are not accidents. They are choices about whose land, water and power get used.
Hao’s “Empire of AI” argues that AI companies exploit labour around the world, paying very little for data annotation and content moderation, and that their models are built on the unpaid intellectual labour of billions of people.
Astra Taylor has a name for how this stays hidden. She calls it fauxtomation: the process that makes human labour invisible so machines look smarter than they are. https://unfoldingstrategies.substack.com/p/fake-automation-fauxtomation
It also reinforces the idea that unpaid work has no value and gets us used to the thought that we won’t be needed.
What consent can look like: Te Hiku Media
The same technology can be used very differently when the people affected are in charge. In Northland Te Hiku Media, a non-profit Māori broadcaster founded in 1990, spent decades recording the stories of its people, most of them native speakers, through a network of radio and television stations. Its CEO Peter-Lucas Jones has described years of interviewing elders about every river, plant and beach. That archive holds thousands of hours of speech that is rich in knowledge about the environment and traditional medicine, and it was too big to transcribe by hand. It is made up of decades of te reo Māori broadcast by speakers born as early as the 1890s. It represents an account of te reo Māori as it was before colonisation shattered the intergenerational transmission between primary speakers. So Te Hiku turned to AI, starting with Kōrero Māori, a project to automate transcription of its archives and build te reo Māori tools for Māori.
https://blogs.nvidia.com/blog/te-hiku-media-maori-speech-ai
The method matters as much as the tool. Te Hiku began by explaining its cause to elders and asking them to come to the station to read phrases aloud. A public campaign then collected more than 300 hours of recordings in 10 days, on top of about 3,500 hours of archived audio from native speakers. The resulting speech recognition model transcribes te reo Māori with about 92% accuracy, and it powers tools such as the Kaituhi transcription service and the Rongo pronunciation app.
Control stays with the community through the Kaitiakitanga Licence, written from a haukāinga (home community) perspective. Kaitiakitanga means guardianship, and the licence says the data may be used only for the benefit of Māori. It bars uses that surveil, discriminate or violate human rights, and it covers partnerships with outside organisations. Soon after the first app appeared, a US translation company asked Māori speakers and academics for voice recordings to build a translation service. Te Hiku saw this as “indigenous data theft”, an attempt to profit from Māori knowledge. Jones explains the stance: Māori have already lost land, so they take data sovereignty seriously.“When we think about what that could mean for indigenous languages like te reo Māori, which was literally beaten out of the mouths of those generations above us, it would enable the development of a service that would be sold back to the very people that had it forcibly removed from them,” Jones said.
Chief technology officer Keoni Mahelona argues that if corporates get the data, they will profit and sell it back, and that communities should lead their own platforms instead. Te Hiku has done this, and now has a GPU cluster (group of physical computer servers) in Kaitaia. The in-house option both gives them control and is a less expensive way to produce the AI models that underpin their work.
By adding in-house systems to process their precious data and putting the processing power in the hands of invested community members “that’s how we’re creating the best bilingual tools for te reo Māori” says Te Hiku chief executive Peter-Lucas Jones.
This model does not erase every cost. Training and running models still use energy and hardware. But it changes the questions that Klein, Taylor and Hao raise. The people whose voices built the system gave their consent, set the terms, decide what it is for, and keep the benefits. It points to a few tests for any use of AI:
-Informed consent from the people whose words, faces, art or labour go in.
– A purpose that benefits them, and that they helped define.
– Community control over the data, including the power to say no.
– Limits on use, with surveillance, discrimination and resale off the table.
– A refusal to sell access to those who only want to extract value.
Why it matters who owns it
Hao describes a race for land, water and cheap labour that concentrates power in a few hands. Klein and Taylor’s “End Times Fascism” shows what some of those hands want. They describe an alliance of the religious right, Big Tech and aspiring bunker-builders. Klein calls it fascism without a horizon for the future. They argue the most powerful people are preparing for the end of the world while frenetically accelerating it. AI is one of the tools in that project. But they also insist that this fascism has not won and is not totalizing.
None of this says AI must never be used. It says use is not neutral. Ethical use means using it sparingly and for purposes that matter, and not replacing people we would otherwise pay. It means reserving image and video generation for cases that clearly justify the cost, never making realistic depictions of real people without their consent, and commissioning artists where we can. It means asking where the labour and resources came from, and backing the communities that resist data centres built without their consent. It also means backing consent-based projects like Te Hiku Media’s, where the people whose language, images or work feed the system decide what it is for and share in what it produces. And it means pushing for rules, because personal restraint cannot fix a system whose costs fall on some and whose profits go to others. Klein’s point is that the purpose of the economy that deploys AI decides whether it helps or harms.
Article by Melissa with minimal help from prompts given to AI to flesh out ideas on ethical use.
Leave a comment