notes from Empire of AI by Karen Hao, 8 years in the making, hundreds of interviews
AI
Generative AI models are deep learning models trained to generate reproductions of their data inputs. From old text, they learn to synthesize new text, from old images synthesize new images. Neural networks are statistic pattern matchers.
Deep learning affords the greatest competitive advantage to players with the most data.
AI is a tech that has many forms. Large language models have taken the limelight, but this is a narrow view of what AI could be, and what the world is. The culmination of thousands of subjective choices.
How does AI shift power? Does it consolidate or redistribute?
AI from Hito Steyerl’s Medium Hot
AI: computational power plus bureaucracy, performative stats, probability-based patterns
Largescale data kidnapping, polluting hardware based on menial and disenfranchised labour
Subprime visibility prone to speculation and crashes. Platform authoritarianism. Tech evangelists.
ChatGPT is a chatbot that will do almost anything you ask of it, in the manner of an utterly sycophantic slimeball servant. It will experience art like Kant, Benjamin or your mother. It will spit out recipes for bioweapons, margaritas or self-optimisation. It will pretend to be your romantic partner, or mentor, but in essence it is optimised to be your bottom.
Elite
The history of AI shows us that AI development has always been shaped by a powerful elite. It’s not a coincidence that AI today has become synonymous with colossal, resource-hungry models that only a tiny handful of companies are able to develop, and that desire to make their products into the foundations for everything.
Open AI leading acceleration towards new colonial world order. In pursuit of amorphous vision of progress, aggressive push to scale up has invaded privacy, committed theft on global scale, devastated environment, relied on traumatized and underpaid workers. They project racist, dehumanizing ideas of their own superiority and modernity to justify and even entice the conquered into accepting the theft and subjugation.
Boomers + Doomers
Boomers say: AI yes! Let’s race to create machines that think like humans. Doomers: safety first. AI can be used for war, target parts of populations for oppression. Can promote racism, sexism, able-ism, hate speech, propaganda. AI requires environmental catastrophe and exploited labour. A few white men shouldn’t be global data stewards making decisions that affect everyone. What about partners in knowledge production? Large AI companies built on theft and non-consent, tech requires capitulation of privacy, agency, worth, including value of labour and art, towards an imperial centralization project. Data is the last frontier of colonization.
Fixing AI
Karen Hao: We don’t need to accept logic of giant scale and production for progress. What society needs is better health care, education, clean air/water, faster transition from fossil fuels. Social cohesion and global co-operation.
Governments: strong data privacy and transparency rules around supply chains, environmental costs and whose data is scraped. Update intellectual property protections to return people’s agency over data and work. International labour laws – wage minimums and humane working conditions for data labelers. Hollywood strike showed importance of unions to resist devaluing labour. Diversity of approaches to what AI could be. Social and environmental costs are hidden behind an elusive vision of progress.
What exactly is in the training data they feed their models? Restricting model weights give smaller developers fewer pathways to create their own AI products, further entrench the dominance of the giants, and make AI models harder to scrutinize. Independent knowledge production: journalists, civil society groups. How to redistribute knowledge away from empire?
Joseph Weizenbaum: “Once a particular program is unmasked, once is inner workings are explained in language sufficiently plain to induce understanding, its magic crumbles away.”
Environment
Environmental catastrophe, needs water, power. Land, energy and water required to house and run massive data centers (warehouses filled with computers – “campuses” large tracts of land with massive buildings filled with densely packed rack on rack of computers.) These computers give off heat – buildings have massive cooling systems. Together the gear creates a cacophony of humming, whirring and crackling that can be heard for miles, 24 hours a day. Noise pollution.
Chile hosts 50 data centers – communities across the country fight against the dispossession of their land, water and other resources in service of global north visions that do not include or benefit them. These movements ask: how to develop AI without extraction. “If we are going to develop this tech in the same way we used to, we are going to devastate the earth.”
Each ChatGPT query needs 10X more electricity than a typical Google search.
Until recently largest data centers 150-megawatt facilities = 122,000 American households. New AI megacampuses will require 1-2000 megawatts of power. Each would use as much energy as 1.5-3.5 San Franciscos. “The kind of electricity growth that hasn’t been seen in a generation.” By 2023 data centers projected to use 8% of the country’s power, it was 3% in 2022. The data centers don’t provide many jobs after their initial construction.
Minerals needed: copper and lithium to build hardware.
Huge amounts of potable water – by 2027 half water used by UK – often in poor countries/communities.
Extraction
Chile is world’s largest copper producer. Copper mining reshaped the land and societies that rely on it. Mining has drained water to process copper. Arsenic in air and water has increased rates of cancer throughout the north of the country. Indigenous tribes can’t grow crops because soil poisoned. Towns plunged into deep poverty, crime rise along with depression, alcoholism. They don’t have enough food, running water, proper health care, education, having seen little benefit from the billions in profits their land generated for the north. Many forced to work for the industry that seized their territories.
Scaling
Scaling is equated with advancement. To get better you must get bigger.
AI development requires massive corporate/state investment.
Scale ups so large, requiring so much computing power, and billions of dollars, and billions of words/images. Only the rich can run these companies. Explosive global costs of its massive deep learning models, the perilous race it sparked to scale such models to planetary limits. Cost of a single competitive generative AI model $1billion in 2025. Could be $5-10 billion in next couple of years.
Surveillance
The aggressive push to collect more training data led to pervasive surveillance not just in digital world but physical as well. The gaze of physical surveillance fell on already vulnerable populations – children, historically marginalized groups, developing countries. Harvard start up sold AI-powered headband that said it could measure student’s brain wave activity to tell teachers whether the child was focused or not. Pilots in schools in Colombia and China. “Data colonialism” AI is the new colonialism.
South Africa filled with street cams gathering data for AI companies – restricting movements of black people already squeezed by racial legacies of apartheid. The revolution promising to bring everyone a better future but persecuting people on the margins.
Industrial capitalism derived value from producing material goods that people wanted to buy, Shoshana Zuboff argued that surveillance capitalism treated its users as the product. Tech giants sitting atop vast amounts of user data could easily pump those troves into neural networks to more precisely profile users and milk their engagement for ad revenue. They logged every user’s click, scroll, likes, encouraging them to supply increasingly personal digital artifacts, every email exchange.
Silicon Valley’s supercharging of deep learning in its quest to expand and entrench global-scale monopolies codified a culture among AI developers to view everything as data to be captured and consumed by their technologies.
Racism+Sexism
For large language models, companies scraped whatever they could find on the internet, capturing more toxic and abusive language and images. Amplify hate around gender, race and religion. Beautiful equals white, engineers are men. “Data swamps” (eg. Common crawl) Abeba Birhane “hate scaling laws” (because AI only learns from the examples you give it) the amount of hateful and abusve content scaled with the size of the dataset and exacerbated the discriminatory behaviours of the models trained on them.
People running AI/tech companies plus department heads are white, male and American.
Theft
Like other empires AI relies on theft of the commons – data commons. Eg. AI that can do computer coding stole from open-source sites turning community work into a commodity.
Ongoing lawsuits from artists, writers, coders, newspapers – AI profits without credit, consent or compensation, and is now used to automate jobs from creatives they’ve scraped.
Military
In 2018 thousands of Google employees protested a secret contract with pentagon for Project Maven to develop ai-powered surveillance drones, laying the groundwork for autonomous weapons.
Medium Cool by Hito Steyerl: War an optimal scenario for generating massive amounts of data
Gaza and Ukraine: labs for developing autonomous weapons- drones with explosives “fire and forget”
Military activities are a ready-made component of AI development touted by major data corporations
Data/images wrecking surroundings via extraction of labour, data and resources, rendering war, marketing and surveillance as variations on the same spectrum
Secrecy
It started as an academic research company – but inconvenient truths are blocked, plus restrict research publications, reveal progress selectively, keep it all a secret, don’t share. The end of accountability.
“Open AI” closed up when it was for profit – the material it is stealing/scraping, how testing is done, whose data is used.
AI salaries so high universities lose profs, phd students.
Other kinds of research have no money and are left behind, even AI research. At scale, the practice began to erode the boundaries of truly independent research.
Science in captivity.
Hidden workers
AI hidden workers – content moderation – global pipeline of labour (book: ‘ghost work’ by gray and suri) repeatedly exposed workers to violent and graphic vids, suicides, beheadings, child abuse. This work broke many of its workers. Kenya content moderation workers: $1.50-3.75/hr
Kenya was among the top destination that Silicon Valley had been outsourcing its dirtiest work to for years: it is poor, in the global south, with a govt hungry for foreign investment – long legacy of colonialism so no institutions left to protect its citizens from exploitation.
2021 Open AI hires firm Sama to provide workers for content moderation for GPT-4. In 2022 Billy Perrigo (Time mag reporter) reveals Sama workers on Meta (Facebook) contract exposed to violent and graphic vids suicides and beheadings. 200 workers filed lawsuits.
New companies: Scale AI, Hive, Might AI, Appen – work is called tasking (workers paid by task).
Competitive pressures lead to race to bottom. Pay workers as little as possible.
After Venezuela’s economy collapsed in 2018 hundreds of thousands joined data-annotation industry. Should this item be listed under clothing or accessories? Does this vid contain crime or human rights violations? Each time task completed, the sum of money earned would increase by a few pennies. Need minimum of 10 us dollars to withdraw. Erratic nature of when work came and went control workers lives. Poverty wasn’t just lack of money, it was erratic sleep, poor health, no agency or control over your own life, lack of self-esteem. Workers want a trad’l employer: full time contract, manager they can talk to, security, consistent salary and health care.
After economic collapse Venezuela checked off perfect mix of conditions to find an inexhaustible supply of cheap labour: population had high level of education, good internet, zealous desire to work for whatever wages. Result? Firms could offer astonishingly good prices for their services.
One of the defining features that drives an empire’s rapid accumulation of wealth is its ability to pay very little or nothing at all to reap the economic benefits of a broad base of human labour.
Scale AI: specialized, quality services at low prices. Kenya, Philippines (English-speaking former colonies with long history of servicing US companies through call centers and digital work). High density of people with good education and internet yet were poor and willing to work hard for little money. Worker-facing platform Remo-tasks. Then shift to Venezuela (cheapest in the market). it promised workers opportunities to learn new skills, advance careers, increase earnings through consistent working hours and wages. But wages declined within weeks of the program’s launch, workers who started at $40/week soon made less than $6 or nothing at all. Anyone who complained was kicked off the platform. To support its expanding client needs Scale AI entered countries with large populations facing financial duress and who could speak most economically valuable languages. Africa, Southeast Asia were targeted.
In Kenya, Okinyi works for AI to train OpenAI’s content-moderation filter. Two streams: sexual content, violence, hate speech, self-harm. Read and sort texts per AI instructions. Sexual content: 15,000 pieces of content/month. Child sexual abuse, incest, bestiality, rape, sex trafficking, sexual slavery. Okinyi mental degeneration after reading perverse sexual acts all day every day.
In 2024 Scale blocked Kenya as a country from Remotasks, just like it did with Venezuela. New priority follows demands of AI: highly educated workers perform RLHF (reinforcement learning from humans): doctors, coders, physicists, paid $40/hr. All workers in Kenya lose work without notice. AI logic: promises of economic freedom, creating new jobs while companies pad bottom lines, the poor lose more and more and highly educated become ventriloquists for chatbots. Empire devalues human labour.
The perpetuation of the empire rests as much on rewarding those with power and privilege as it does on exploiting and depriving those, often far away and hidden from view, without them.
Scaling 2
Like Edward Snowden told us: scaling up as rapidly as possible means grab everything, everything on the internet, everything possible.
Altman: tech needs to get 10X better with each generation. The speed to realize new gen is key. the ‘iteration’ cycle
Monopoly
Peter Thiel started PayPal, data-mining firm Palantir, early FB investor. “Increasing sales fixes all problems. Either you’re growing or you’re dying.” All company founders should “aim for monopoly” to create a successful business. Competition is for losers. Monopolies are good because more stable, longer-term business, more capital. It relies on proprietary technology (can’t be superseded by anyone else), network effects (building relationships), economies of scale, good branding. Companies needed “huge breakthrough” at beginning to establish dominance and the ”last breakthrough” to maintain dominance (improving at rapid enough pace so no one catches up). If you have structure where lots of innovation by many people for your product that’s great for society, but not good for business. No open source.
Language
Large language models accelerate language loss – consolidates major languages, only ones that provide enough data for AI. 7000 languages left today, half endangered, 2% supported by Google Translate, .2% supported by GPT-4.
Data is the last frontier of colonization.
Sam Altman
CEO of OpenAI – has sister Annie with mental health problems, housing problems. He did nothing to help her. She turned to sex work.
Profit
Research – what is being developed, what AI is ‘trained’ to do – is based on what makes money.
Diversity in AI research collapses under the logic of scaling. But scale is not the only pathway to improved performance. AI remade the logic of research.
The job security of deep learning and diminishing viable career paths in other methods.
Heat Hito Steyerl’s Medium Cool
Integration of AI into search engines increases energy demand 6 to 10 times
Fire may also have contributed to the rise of language, the division of labour and development of social forms. It changes the night. Control surroundings. Shape landscape.
Nicole Starosielski: The most powerful media orgs of the 21st century will be thermal. The circulation of images, sounds, videos, and texts will depend on a massive regime of heating and cooling. Data and networks, like the people they connect, will be ever more fragile. Too hot or too cold, and the platforms will collapse. Digital infrastructures – data centers, network exchanges, and fiberoptic cables – will drain the planet’s energy in order to create a stable thermal environment – not for people but for information.
A Prometheus attack takes a server hostage by encrypting all the files on it and demanding a random (a “timed ticket”) for people to access their own data. If the victims (called customers) don’t pay, info is auctioned.
Paying to use cloud services or subscription apps is a legal form of Prometheus attack
McLuhan: hot media like print, photo, radio and movies more immersive, rich in data
*profits privatized, risk collectivized
Flow Hito Steyerl’s Medium Cool
People treated like isolated suspended particles, violently tossed about by some kind of financial, ecological and social tsunami
What does a human body look like when generated by statistics?
Data needs pipelines to flow and massive material infrastructures and technology stacks to enable their circulation.
For stagnant cultures to change, everything else must remain the same.
Control over the means of circulation and extraction has supplemented control over the means of production.
Has art become a gateway for the automation of common sense?
We know art can be made by machines (cinema, ready-made, photography), can art be made for machines?
Waste Hito Steyerl’s Medium Cool
Human data waste constitutes a large part of internet traffic, toxified physical but also intellectual and political climates, leads to exhaustion, depletion and widespread burnout. The situation resembles a virtual version of the sanitation crises of the 19th century in large European cities. Disease spread as the water supply was infested by refuse. Huge public sanitation projects and investments were necessary to remedy the situation, clean the waste water to get it back into circulation. These projects were in many cases massively obstructed by private interests and companies who profited from inadequate water management in cities such as London and Berlin.
Future Hito Steyerl’s Medium Cool
AI will replace photography, translation, web design, game design, 3D work, advertising, digital post production, stage design, PR, programming.
In terms of image making, markets are becoming ‘cameras’ – or, more generally, data processors – environmental sensors that render reality as a pricing system. Statistical realism is based not on observation but on quantity of data
Sentences
Persistent claims AI will ‘solve climate change’ and ‘solve the housing crisis’
2021 marks major shift in Sam Altman’s personal investment strategy away from taking a large number of small bets toward taking a small number of really large ones.
During development of DALL-E 3: The amount of porn on internet so large that removing it shrank the training dataset and degraded the model’s performance.
Language model begins with: predict the next word.
We now have machines that can mindlessly generate words, but we haven’t learned how to stop imagining a mind behind them.” Emily Bender