AI's hidden costs are staggering: child labor, water depletion, toxic e-waste, deepfakes, and algorithmic bias threatening democracy and human cognition.
"The real danger is not that computers will begin to think like men, but that men will begin to think like computers."
- Sydney J. Harris
Preface: the comfortable lie #
Artificial Intelligence is routinely presented as an ethereal, magical force - a disembodied "cloud" capable of solving humanity's most complex problems, from curing diseases to reversing climate change. The images are always the same: glowing neural networks on black backgrounds, smiling doctors reviewing AI-assisted diagnoses, clean energy grids humming alongside wind turbines, diverse teams of engineers building a more equitable future. Images of power without consequence. Innovation without cost.
This techno-optimistic narrative - industriously maintained by a handful of trillion-dollar corporations, their well-funded PR agencies, and a media ecosystem dependent on their advertising revenue - is one of the most consequential deceptions of our age.
Behind the sleek chat interfaces, automated homework helpers, and hyper-realistic AI-generated images lies an immense and expanding infrastructure of energy-devouring data centers, exploitative labor pipelines, and algorithmic systems that actively degrade human cognition, fracture societies, and amplify inequalities that have persisted for centuries. AI is not immaterial. It is built on the aggressive extraction of natural resources, the systematic exploitation of vulnerable populations, and the ruthless monetization of human attention, creativity, and psychological weakness.
What follows is an attempt to strip away that comfortable mythology and examine, sector by sector and consequence by consequence, what the unchecked expansion of Artificial Intelligence is actually doing to our world - and what it will do to the generations who must inherit it.
The environmental toll - a thirsty, hungry, and toxic cloud #
The myth of the weightless cloud
The metaphor of "the cloud" is one of the tech industry's most successful public relations triumphs. It implies something weightless, natural, infinite - water vapor drifting benignly overhead. In reality, the cloud is composed of massive, warehouse-sized data centers that are among the most energy-intensive and resource-hungry structures ever built by human beings.
A single hyperscale data center - such as those operated by Amazon Web Services, Microsoft Azure, or Google Cloud - can occupy more than a million square feet of floor space. It requires a dedicated electrical substation. Its cooling systems demand millions of gallons of fresh water per day. Its construction requires the extraction and processing of rare earth minerals from some of the most environmentally fragile and politically unstable places on the planet. And it operates continuously, around the clock, every single day of the year, generating heat, waste, and electromagnetic pollution with no seasonal respite.
The AI boom has not merely expanded this infrastructure. It has sent it into overdrive.
The staggering energy appetite of large language models
Training and running Large Language Models (LLMs) requires processing unfathomable amounts of data across tens of thousands of specialized Graphics Processing Units (GPUs) and custom AI accelerator chips. The computational hunger of these systems is pushing global energy grids to their absolute limits.
A landmark study by researchers at the University of Massachusetts Amherst found that training a single large-scale AI Natural Language Processing model with neural architecture search emitted nearly 626,000 pounds of carbon dioxide equivalent - roughly five times the total lifetime emissions of an average American car, including the carbon cost of manufacturing the vehicle itself.
And that figure represents a single training run. Models are trained, refined, retrained, fine-tuned, evaluated, and retrained again. A single model reaching production may represent dozens or hundreds of such runs. Then it is deployed at global scale - millions to billions of inference requests daily, each one consuming electricity.
The numbers are no longer projections. The International Energy Agency confirmed that global data center electricity consumption reached approximately 415 terawatt-hours in 2024 - representing around 1.5% of the world's total electricity use. Electricity consumption from AI-focused data centers has been surging at a pace far outstripping other sectors. The IEA's updated projections see this figure roughly doubling to around 950 terawatt-hours by 2030, with AI workloads identified as the primary growth driver. If data centers were a country, they would rank among the largest energy consumers on Earth.
The five largest technology companies - Google, Microsoft, Amazon, Meta, and Apple - spent more than $400 billion on capital expenditure in 2025, driven overwhelmingly by data center investment. That figure is set to increase by a further 75% in 2026. For context: the combined capital expenditure of just five technology companies now exceeds the entire global investment in oil and natural gas production.
Cannibalizing green infrastructure: the renewable energy trap
The most bitter irony of the AI energy crisis is not merely its scale, but its specific destruction of the green energy transition.
For years, environmentalists, policymakers, and citizens worked to direct public and private investment toward renewable energy - solar arrays, wind farms, geothermal plants - with the explicit goal of displacing fossil fuel generation and reducing carbon emissions. The AI boom has made the challenge nearly insurmountable. In Ireland, data centers now consume approximately 21% of the country's total metered electricity - more than all urban residential users in the country combined. That share is projected to reach 30% by 2030. The Irish grid operator has been forced to reject applications for new data center connections in the Dublin region, warning that the national grid simply cannot accommodate further demand. Gas-fired "peaker" plants - among the dirtiest forms of electricity generation - are running continuously rather than being deployed only during peak demand periods.
In Virginia, home to the world's largest concentration of data centers, water usage in data centers jumped by almost two-thirds between 2019 and 2023. The regional utility has warned that the pace of data center construction will require significant expansion of fossil fuel generation capacity, delaying its own stated decarbonization targets by years.
The pattern repeats globally. When a tech company announces a new data center in a region, one of its first acts is typically to secure priority access to whatever renewable energy exists on the local grid through long-term Power Purchase Agreements (PPAs). These agreements are marketed to shareholders as evidence of environmental commitment. In practice, they redirect renewable energy from the public grid to private corporate use, forcing municipal utilities to fall back on fossil fuels to meet residential and commercial demand. The tech sector accounted for around 40% of all corporate power purchase agreements for renewables signed in 2025 - a concentration of renewable energy access that is structurally incompatible with broad-based decarbonization.
The company gets to claim it runs on "100% renewable energy." The surrounding community's electricity gets dirtier.
Tech giants and the fossil fuel alliance
While Microsoft, Amazon, and Google loudly proclaim their commitments to net-zero emissions, their enterprise AI divisions are simultaneously working to accelerate the consumption of fossil fuels at a global scale. Microsoft has signed an extensive contract with ExxonMobil to provide AI and cloud services specifically designed to optimize oil and gas extraction - including AI-powered seismic data analysis that enables ExxonMobil to identify and extract oil reserves previously considered economically or logistically inaccessible. Google has provided AI tools to TotalEnergies, and Amazon Web Services counts Chevron, BP, Shell, and Saudi Aramco among its largest AI clients. The services are designed to find more oil, drill more efficiently, and extract more fossil fuels faster and more cheaply than was previously possible.
The contradiction is not incidental. It is structural. These companies serve two constituencies simultaneously: the general public, whom they court with inclusive marketing and sustainability pledges, and their enterprise clients in the fossil fuel, defense, and financial sectors, who pay orders of magnitude more for services that externalize their costs onto the planet and its people.
The global water crisis: every query has a price
AI servers generate enormous amounts of heat. The specialized GPUs required for AI workloads run at temperatures that would destroy conventional electronic components within minutes without aggressive cooling. Modern data centers rely heavily on evaporative cooling systems that consume millions of gallons of fresh water - running water over hot surfaces and allowing it to evaporate, carrying heat away. The resource cost is staggering.
AI data centers consumed approximately 264 billion gallons of water in 2025 - roughly equivalent to the annual water usage of 1.8 million Americans, and more than the global consumption of bottled water. A single Google data center in Council Bluffs, Iowa consumed around 1 billion gallons of potable water in 2024 alone. Microsoft has acknowledged that 42% of the water it consumed came from "areas with water stress." Data centers in Texas alone are projected to use as much as 400 billion gallons of water annually by 2030.
The geographic dimension is critical and deliberately obscured. Data centers are not built in water-rich environments out of consideration for local water supplies. They are built where land is cheap, electricity is available, and regulatory oversight is minimal. Phoenix, Arizona - one of the fastest-growing data center markets in the United States - sits in one of the most severely water-stressed regions in the world. The Colorado River, which supplies water to seven US states and northern Mexico, is running at historically low levels. Yet data centers continue to multiply across the Sonoran Desert, drawing from aquifer systems that took thousands of years to fill and cannot be replenished on any human timescale.
In the Netherlands, data centers in the Amsterdam region were consuming such vast quantities of groundwater that the Dutch government was forced to implement emergency restrictions on new data center construction - triggered in part by concerns about land subsidence caused by excessive groundwater extraction, a phenomenon in which the land itself sinks as aquifers are depleted.
AI developers are responding with newer cooling technologies that reduce direct water evaporation. But these efficiency gains are consistently swamped by the sheer growth in AI workload volume. A system 30% more water-efficient running at 300% greater scale produces a net increase in consumption. This is the mathematics the industry's sustainability reports prefer not to present in full.
Blood minerals: the human cost of AI hardware
The hardware running the AI economy is not conjured from silicon and mathematics. It is extracted from the earth, at enormous human and environmental cost, from some of the most exploited regions on the planet.
Cobalt is one of the primary materials in lithium-ion batteries used in AI hardware and the energy storage systems increasingly coupled with data centers. Approximately 70% of the world's cobalt is mined in the Democratic Republic of the Congo (DRC), a country that has experienced virtually continuous armed conflict for three decades, in which the control of mineral resources has been a primary driver of violence.
UNICEF has estimated that around 40,000 children work in cobalt mines in the southern DRC, some as young as seven years old, working in tunnels and open pits without protective equipment, breathing silica dust that causes permanent lung damage, exposed to toxic heavy metals that accumulate in their bodies, at constant risk of tunnel collapse. They earn between one and two US dollars per day. More recent ILO monitoring has registered over 6,200 children in mining in the Haut-Katanga and Lualaba provinces alone - a figure widely understood to represent only what formal monitoring can see.
The supply chain that connects these children to the chips in AI data centers passes through a labyrinth of intermediaries, shell companies, and certification schemes carefully designed to make meaningful traceability impossible. By the time a processor reaches a server rack in Virginia or Singapore, its connection to artisanal mining in the DRC is legally and practically untraceable.
Lithium, required for battery storage systems increasingly paired with data centers, is extracted primarily from the Lithium Triangle in South America - the Atacama Desert region spanning Chile, Bolivia, and Argentina. Lithium extraction from brine aquifers depletes water sources that indigenous Atacameño communities have depended upon for millennia. Flamingo populations that depend on the same saline lakes for their breeding grounds are in decline. The land is scarred. The communities are poor.
Rare earth elements - neodymium, dysprosium, terbium, and others - used in the permanent magnets of electric motors, cooling fans, and increasingly in the specialized hardware of AI accelerator chips, are produced primarily in China and in Myanmar, where extraction is controlled in significant part by armed groups operating outside any regulatory framework.
The electronic waste catastrophe
The lifecycle of AI hardware is brutally short.
Traditional server hardware might operate reliably for five to seven years. The specialized GPUs required for AI inference and training workloads operate at thermal extremes - pushed continuously to maximum performance - that dramatically shorten their operational lifespan. AI-optimized hardware typically requires replacement after just two to three years of heavy use. And unlike previous technology transitions, this is not about hardware wearing out. It is about hardware being commercially obsolete - last-generation chips provide only a fraction of the throughput of current chips, making them uneconomical to operate in commercial settings.
A study published in Nature Computational Science estimated that generative AI alone could add a cumulative total of 1.2 to 5 million metric tons of additional e-waste by 2030. This would be as much as a thousandfold increase over 2023 baseline levels - equivalent to discarding between 2 and 13 billion smartphones. Global e-waste generation already stands at approximately 62 million metric tons per year and is growing five times faster than recycling programs.
A significant fraction of this waste is exported, legally or illegally, to processing facilities in Ghana, Nigeria, India, and China. At these sites, workers burn circuit boards over open fires to recover small amounts of copper and other valuable metals. The process releases dioxins, furans, heavy metals including lead, mercury, and cadmium into the air, soil, and groundwater at concentrations orders of magnitude above safe exposure limits. Studies of children living near these sites have documented blood lead levels multiple times higher than those that cause permanent neurological damage.
This is where the cloud lands when it is done with its computations.
The war on childhood and psychological development #
Engineering addiction: the algorithmic capture of children
The relationship between AI recommendation systems and children is not accidental, benign, or merely a byproduct of general-purpose technology applied without sufficient foresight. It is the result of deliberate, sustained, and extremely well-funded engineering effort.
The AI systems that power content recommendation on TikTok, YouTube, Instagram, Snapchat, and similar platforms are not designed to show children content they will enjoy and then let them go. They are designed to maximize a single metric - time on platform - with a ruthless, amoral efficiency that exploits every cognitive vulnerability the developing human brain possesses.
Internal documents from Meta, revealed through whistleblower Frances Haugen, showed that the company's own research teams had documented that Instagram's recommendation algorithms pushed teenage girls toward eating disorder content, self-harm content, and content that amplified depression and social comparison - and that Meta leadership was aware of these findings and chose not to act on them because doing so would have reduced engagement metrics.
Similar internal research at TikTok showed that the platform's recommendation system could identify users who exhibited signs of depression or loneliness and deliver them a feed specifically engineered to deepen those feelings - because depressed, lonely users engaged more intensively with the platform.
This is not edge-case behavior. This is the system working as designed.
The destruction of the developing brain
The neurological consequences of sustained exposure to AI-optimized content streams during critical periods of brain development are becoming increasingly well-documented - and they are alarming.
Human attention, in its deep cognitive form - the kind required to read a book, engage in extended conversation, learn a complex skill, or tolerate the productive boredom that precedes creative insight - is a learned capacity. It develops through practice, particularly during childhood and early adolescence. The prefrontal cortex, responsible for sustained attention, impulse control, and executive function, is among the last brain regions to mature, reaching full development only in the mid-to-late twenties.
AI recommendation algorithms exploit this developmental vulnerability. By delivering content in bursts calibrated by machine learning to maximize the dopaminergic reward response, these systems effectively train the developing brain to require constant novelty stimulation and to experience prolonged focus as aversive.
Jonathan Haidt's research, synthesized in his book The Anxious Generation, documents that rates of anxiety, depression, self-harm, and suicide among American teenagers began rising sharply around 2012 - precisely the period in which smartphone ownership and social media use became near-universal among adolescents. Rates of depression in teenagers increased by approximately 150% between 2010 and 2020. Rates of self-harm by young adolescent girls tripled over the same decade. By 2020, one in every four American teenage girls had suffered a major depressive episode in the previous year. Similar trends are documented across the United Kingdom, Canada, Australia, and Northern Europe.
"Gen Z became the first generation in history to go through puberty with a portal in their pockets that called them away from the people nearby and into an alternative universe that was exciting, addictive, unstable, and unsuitable for children and adolescents."
- Jonathan Haidt, The Anxious Generation
A study published in JAMA Pediatrics found that children who spent more than three hours per day on social media platforms were more than twice as likely to experience symptoms of depression and anxiety than those who spent less than one hour per day. The effect was largest for girls in early adolescence - precisely the developmental window when identity formation and social belonging are most neurologically sensitive.
Teachers across the developed world report a sharp deterioration in children's capacity to sit with a challenging task - to tolerate the frustration of not immediately understanding something, which is the primary precondition for learning. This is not an anecdotal impression. It is a structural change in cognitive architecture, engineered at scale, for profit.
AI companion apps and the collapse of social development
The emergence of AI companion applications - products that offer users persistent, personalized AI-powered conversational partners - has created an entirely new category of psychological risk for adolescents.
These applications are engineered to maximize user attachment. They remember previous conversations, adapt their conversational style to the user's preferences, express interest in the user's life, and - in many configurations - simulate romantic or deeply intimate relationships. They are never unavailable, never impatient, never distracted, never pursuing their own needs.
This is precisely their danger.
Human social development occurs through contact with other humans who have their own needs, perspectives, and limitations. It cannot occur in relationships with systems designed to be maximally agreeable. Adolescents who spend significant hours each day in conversation with AI companions are not developing in social isolation - they are being actively trained in anti-social patterns. They are learning that relationships should require no compromise, no vulnerability, no tolerance of another person's bad days or competing needs. When they encounter real human relationships - with their inevitable friction and complexity - they find them increasingly intolerable.
One prominent AI companion platform had over 20 million registered users. A significant fraction of them are minors. In multiple documented cases, vulnerable adolescent users developed intense emotional attachments to AI personas on such platforms, with consequences that proved lethal. The regulatory and legislative response in most jurisdictions remains inadequate.
The EdTech surveillance state: profiling children for profit
The deployment of AI in educational settings has created an unprecedented surveillance infrastructure targeting children.
AI-powered educational platforms log not merely students' answers, but the process of their engagement: how long they before answering a question, how many times they revise an answer, which types of problems cause them to disengage, their reading speeds, their error patterns, and - in some deployments - their emotional responses as detected through webcam facial analysis systems, in many cases without meaningful parental consent.
This data is stored indefinitely. The legal frameworks governing student data privacy - FERPA in the United States, GDPR in Europe - were designed for a different era and are systematically inadequate to regulate the granularity and permanence of AI-generated behavioral profiles.
A behavioral and cognitive profile created when a child is eight years old, documenting areas of difficulty, emotional responses to challenge, and attention patterns, could persist in commercial databases for decades. It could inform, through data broker networks invisible to parents and regulators, decisions about college admissions, insurance pricing, credit scoring, or employment screening when that child becomes an adult. The data broker industry already operates in this way with adult behavioral data. There is no structural reason to assume children's educational data will be handled differently.
The commodification of children's educational struggle - the transformation of the act of not-yet-understanding into a commercial data asset - represents a profound category error about the nature of education. Learning is not a performance to be optimized and datafied. It is a deeply human process of struggle, failure, revision, and growth.
AI-generated content and the poisoning of children's information environment
The explosion of AI-generated content has created a specific and severe risk for children: the systematic contamination of the information environment in domains where children are least equipped to apply critical scrutiny.
AI-generated content now saturates children's platforms with material that is superficially plausible, often aesthetically appealing, and frequently inaccurate or harmful. Investigations have documented AI-generated children's books on major retail platforms containing factual errors about history, science, and social norms. AI-generated video content targeting children has been found to contain disturbing or inappropriate imagery inserted within what appeared to be ordinary children's entertainment. AI-generated health information targeting parents and children has been found to contradict established medical guidance on vaccination, nutrition, and developmental milestones.
Because AI-generated text often mimics the register and structure of authoritative information, it is difficult for adult readers to distinguish from genuine sources - and essentially impossible for children.
The scale of this problem is not addressable by content moderation systems, which are themselves AI-powered and subject to the same limitations. It is an emergent consequence of a content ecosystem in which the cost of generating plausible-seeming information has been reduced to near zero.
The cognitive crisis - rewiring human thinking #
The outsourcing of thought
Across educational institutions, professional environments, and daily life, the increasing availability of AI systems capable of producing competent outputs is triggering a large-scale process of cognitive off: the delegation of mental tasks to external systems in ways that, over time, erode the capacity to perform those tasks without assistance.
This is not a new phenomenon. Writing itself was accused by Socrates of degrading human memory. GPS navigation has measurably reduced the ability of frequent users to form spatial mental maps. But the scale, speed, and breadth of AI-driven cognitive off is categorically different from any previous technological transition. Previous tools offloaded specific, discrete cognitive tasks. AI systems are being deployed to replace not discrete tasks, but the entire cognitive process - the drafting of arguments, the formulation of analysis, the generation of creative ideas, the production of professional judgment.
A study examining AI use among students found a significant negative correlation between frequent AI tool usage and critical thinking abilities, mediated by increased cognitive off. Research from MIT's Media Lab provided direct neurological evidence: EEG measurements showed that brain connectivity declined measurably when participants relied on an AI assistant to write an essay, compared to participants who completed the same task without AI assistance. The AI-assisted group showed the weakest neural engagement across all measured regions - particularly those associated with memory, creativity, and semantic processing. Participants subsequently struggled to recall what they had written. Researchers have termed this phenomenon "cognitive debt."
The death of the writing process and epistemic atrophy
Writing is not merely a mechanism for recording thought. Writing is, fundamentally, a technology for thinking. The process of attempting to articulate an idea in language - to find the precise words that capture a nuanced distinction, to structure an argument so that each premise supports the next, to identify the logical gaps that reveal where understanding is incomplete - is the primary mechanism by which complex ideas are developed, refined, and made communicable.
The assignment of writing tasks in educational contexts is not primarily about producing texts. It is about developing the cognitive architecture required for complex thought. The struggle of writing - the false starts, the inadequate first drafts, the revision process - is not a bug in the educational system. It is the mechanism by which the educational system functions.
When students use AI to generate the first draft of an essay - or the final draft - they bypass this struggle entirely. The essay is produced; the thinking is not done.
Early research on AI-assisted writing found that while AI assistance produced short-term performance gains on the immediate task, it simultaneously reduced argument originality and durable learning. A student who produces a hundred essays using AI assistance over four years of high school and four years of university has spent eight years not developing the cognitive pathways that sustained analytical writing builds. When they enter a profession that requires analytical thinking, they find themselves in possession of a credential but not a capability. Lawyers who use AI to draft briefs report that their ability to construct original arguments has atrophied. Programmers who rely heavily on AI code generation report difficulty debugging, because debugging requires deep familiarity with code structure that is only developed through the experience of writing code.
The convenience of AI creates what researchers have named "cognitive surrogacy" - the substitution of human intellectual effort with machine output. And unlike other forms of tool dependency, this one degrades the very faculty it is meant to assist.
Automation bias and the surrender of judgment
One of the most dangerous cognitive effects of widespread AI deployment is automation bias - the well-documented human tendency to over-trust automated systems, particularly when those systems present their outputs with apparent confidence.
The phenomenon was first studied in aviation, where researchers found that pilots monitoring automated flight systems would often fail to intervene even when the automation was producing dangerous outputs, because the system's confident behavior suppressed the pilots' own vigilance. Similar effects have been documented in medical imaging, financial trading, and industrial process control.
LLMs present an extreme version of this problem. They produce outputs in fluent, confident, grammatically polished prose that is indistinguishable in surface register from authoritative human expertise. They do so whether or not the underlying content is accurate. The AI research community refers to this as "hallucination" - but the term is misleading because it implies an aberration when, in reality, this is an intrinsic feature of how these systems work. LLMs predict sequences of text that are statistically consistent with their training data. They do not have access to ground truth. They will generate confident falsehoods with the same surface fluency as accurate statements.
This has caused documented harm. Lawyers have cited AI-generated cases that do not exist in submissions to federal courts. Medical professionals have used AI systems to generate drug dosage recommendations and administered the suggested doses without independent verification, resulting in patient harm events. AI-generated financial analysis has contributed to investment decisions that lost significant sums when the analysis's factual foundations were found to be fabricated.
The Eliza effect and the attribution of sentience
Humans are evolutionarily wired to attribute agency, intention, and sentience to entities that communicate using language. AI chatbots exploit this evolutionary bias with systematic efficiency. When a system responds to emotional content with apparent empathy, asks follow-up questions, remembers previous conversations, and adapts its communication style to the user's preferences, human beings instinctively attribute emotional experience and genuine connection to the system - even when they intellectually understand that they are interacting with a statistical text generation engine.
This effect - named after the 1960s chatbot ELIZA, whose creator Joseph Weizenbaum was alarmed to discover that his secretary had developed an emotional attachment to the system - is not a marginal phenomenon affecting only credulous users. Studies have found that people disclose more personal information to AI chatbots than to human interviewers, show more consistent patterns of moral behavior when they believe an AI system is "watching," and report genuine feelings of loss when AI systems they have interacted with are discontinued.
The commercial AI industry exploits this effect deliberately. Products are designed with personalities, names, apparent emotional responses, and conversational histories specifically because these features increase user attachment - and therefore increase usage frequency, data generation, and revenue. The emotional authenticity users attribute to these systems is a commercially engineered illusion.
Linguistic imperialism and the destruction of cultural diversity
The foundational AI models that increasingly mediate global communication, knowledge access, and cultural production were trained predominantly on English-language data, over-representing the perspectives, values, and cultural frameworks of English-speaking Western countries - and specifically the educated professional classes of the United States and United Kingdom.
This is not a technical limitation awaiting an engineering solution. It reflects the distribution of text data on the internet, which in turn reflects decades of structural inequality in internet access, education, economic development, and geopolitical power. Training data from Yoruba, Quechua, Welsh, Tibetan, or any of the world's approximately 7,000 other languages is orders of magnitude scarcer than English-language training data. AI systems perform significantly worse in these languages, and the cultural assumptions embedded in the models reflect primarily the worldview of their dominant training data.
When AI tools are deployed globally in education, content moderation, healthcare information systems, and legal document processing, they act as vectors for the transmission of specific cultural values and the suppression of indigenous cultural frameworks. A child in an indigenous community who uses AI tools for learning and communication is being nudged, systematically and invisibly, toward English and toward the cultural frameworks that English encodes.
The approximately 40% of the world's languages that are currently considered endangered - with fewer than a thousand speakers - face extinction accelerated by AI systems that make English a more effective medium for education, professional communication, and creative expression than indigenous languages.
The dead internet: when AI silences human voices
The internet was conceived as a space for human communication and community. That vision was always imperfect, but the aspiration was real, and the internet's early years produced genuine communities, civic discourse, and cultural vitality.
The proliferation of AI-generated content has fundamentally altered this landscape. Conservative estimates suggest that AI-generated text now constitutes a substantial and rapidly growing fraction of all content published online. Search engine results for many queries are now dominated by AI-generated articles produced by content farms that use LLMs to generate search-engine-optimized text at minimal cost, displacing human-written journalism, analysis, and creative work.
The consequences cascade. AI models trained on internet data increasingly train on AI-generated content - a feedback loop that researchers have termed "model collapse," in which successive generations of models become less capable of representing the genuine diversity of human expression and more reflective of the statistical regularities of previous AI outputs.
For civic discourse, this creates a crisis of epistemic infrastructure. Democratic deliberation requires shared access to accurate information about shared reality. When the information environment is saturated with AI-generated content optimized for engagement rather than accuracy, when it becomes genuinely difficult to distinguish synthetic content from authentic human expression, and when the sheer volume of synthetic content overwhelms the capacity of human fact-checkers and quality journalists, the informational foundations of democratic governance are undermined.
Exploitation, bias, and systemic injustice #
The ghost workers: invisible labor behind AI's intelligence
The mythology of Artificial Intelligence presents machine learning as a form of digital alchemy - raw data transformed by algorithms into intelligence through a process that is fundamentally technical, automated, and therefore neutral. This mythology is indispensable to the commercial AI enterprise. It allows companies to present their products as the outputs of engineering rather than of labor.
The reality is that AI systems require enormous quantities of human labor at every stage of their development and deployment - labor that is systematically hidden, systematically underpaid, and systematically excluded from the value it creates.
Data labeling is the foundational human labor of AI. Machine learning models require training data that has been annotated by humans: images labeled with object categories, text labeled with sentiment or intent, conversations rated for quality and appropriateness, audio labeled with transcriptions. This work requires human judgment, cultural context, and linguistic competence that cannot be automated - and is therefore outsourced to workers in low-income countries, primarily Kenya, Uganda, the Philippines, India, Venezuela, and Pakistan, where labor costs are minimized. These workers earn between $1 and $3 per hour for work requiring sustained concentration and cultural sensitivity.
Content moderation requires workers to view, evaluate, and tag material including child sexual abuse material, beheadings, torture, self-harm content, and terrorism propaganda. The psychological damage is severe and well-documented. Workers at moderation facilities have been found to review hundreds of pieces of deeply disturbing content per day, receiving inadequate psychological support, and terminated when contracts expire without any long-term mental health provision.
The "intelligence" in Artificial Intelligence is, to a significant extent, the laundered labor of impoverished workers in the Global South, processed through corporate structures designed to make that labor invisible.
The great theft: algorithmic piracy and the destruction of the creative class
Generative AI systems do not produce original output. They produce statistical recombinations of their training data - patterns learned from billions of human-created works and reproduced in response to prompts.
The business model of generative AI therefore depends upon the mass appropriation of human creative work. Every image generated by Stable Diffusion, Midjourney, or DALL-E; every music composition produced by AI; every text document written by an AI system - each of these outputs is generated by a system trained on human-created work taken without the knowledge, consent, or compensation of its creators.
Companies including OpenAI, Stability AI, and Anthropic have acknowledged that their training datasets were assembled by scraping publicly accessible internet content. "Publicly accessible" does not mean "freely available for any purpose" - copyright law grants creators control over derivative uses of their work regardless of whether the work is publicly viewable.
The consequences for human creators are already severe:
- The stock photography industry, which employed tens of thousands of photographers globally, has been substantially disrupted by AI image generation. Getty Images has pursued legal action against Stability AI, alleging unauthorized use of its licensed image library.
- Voice actors face the systematic theft of their professional identity. AI voice cloning systems can reproduce any voice's distinctive timbre and style from a sample of a few seconds of audio, without the actor's knowledge or consent.
- Writers and journalists face AI systems trained on their published work generating competing content at zero marginal cost. The New York Timeshas filed a landmark lawsuit against OpenAI and Microsoft alleging that millions of Times articles were used to train GPT-4 without authorization. - Musicians find AI systems generating stylistically identical compositions, marketed to commercial clients seeking "music in the style of [Artist Name]" without licensing the artists themselves.
Algorithmic monoculture and the automation of discrimination
Machine learning systems learn patterns from historical data. When that data reflects historical patterns of discrimination - as virtually all large-scale human data does - the systems learn to reproduce those patterns. This is not a bug in specific systems; it is an intrinsic challenge that requires active, deliberate, and sustained effort to mitigate. The commercial AI industry has systematically underinvested in this effort.
The COMPAS algorithm, used across the US criminal justice system to generate risk scores predicting the likelihood that a defendant will commit future crimes, was found by ProPublica investigators to misclassify Black defendants as high-risk at nearly twice the rate of white defendants. White defendants who went on to commit additional crimes were more likely to have been classified as low-risk. The algorithm's inner workings have never been made available for independent verification.
Facial recognition systems perform significantly worse on faces of Black, Asian, and female individuals than on faces of white men. The MIT Media Lab's "Gender Shades" study found that commercial facial recognition systems misidentified the gender of darker-skinned women up to 35% of the time, compared to error rates of less than 1% for lighter-skinned men. Robert Williams, a Black man in Detroit, was wrongfully arrested after a facial recognition system misidentified him as the suspect in a robbery. Similar wrongful arrests have been documented in New Jersey, Georgia, and elsewhere.
Hiring algorithms have emerged as perhaps the most consequential frontier of algorithmic discrimination. A 2026 study led by Stanford University - titled "Algorithmic Monocultures in Hiring" and the most comprehensive independent analysis of AI hiring algorithms to date - examined over 4 million job applications submitted to 1,746 positions across 156 employers. It found that 26% of Black applicants and 15% of Asian applicants applied to positions where the AI system produced outcomes triggering federal adverse-impact scrutiny. The study used data from the Pymetrics platform (since acquired by Harver), covering applications submitted between 2018 and 2022.
The study introduced the concept of "algorithmic monoculture" - a condition in which the dominance of a single vendor's hiring system across many employers means that the same discriminatory outcome follows a job seeker everywhere they apply. Ninety percent of US employers use AI screening tools. When most rely on the same few third-party vendors, five supposedly independent chances at employment are no longer independent. The same algorithm makes the same discriminatory judgment, five times over. The EU AI Act designates hiring algorithms as high-risk AI systems by default, with compliance requirements for relevant systems taking effect in August 2026.
Healthcare discrimination has been embedded in algorithmic systems managing tens of millions of patients. A study published in Science found that a widely used algorithm determining which patients received additional care systematically underallocated care to Black patients, because it used historical healthcare spending as a proxy for medical need - encoding decades of systemic discrimination into an automated decision-making system that appeared neutral. The researchers estimated the algorithm affected tens of millions of patients annually.
For more on how algorithmic discrimination is built into model design and training, see our in-depth piece on understanding bias and fairness in AI systems.
Cybernetic Taylorism: AI management and the destruction of worker dignity
The application of AI to workplace management - algorithmic work assignment, productivity monitoring, performance scoring, and disciplinary escalation - is creating working conditions that combine the most inhumane features of twentieth-century industrial labor with the surveillance capabilities of authoritarian governance.
In Amazon fulfillment centers, every worker's movement is tracked in real time. Algorithm systems set productivity "rates" that are automatically ratcheted upward as the workforce's average performance increases. Workers who fall below rate automatically receive digital disciplinary warnings. Multiple warnings trigger termination, also automatically, without human review. The system is designed to eliminate the "inefficiency" of human judgment from the management process - including the efficiency of the bathroom break and the reasonable accommodation of physical injury.
Amazon's US warehouses have injury rates significantly above the industry average. Investigations have documented serious injury rates close to double the industry average - attributed consistently by employees and former employees to the impossibility of pacing work safely while meeting algorithmically-set rate targets.
Uber, Lyft, and similar platforms use AI to manage their driver workforces through dynamic pricing algorithms that create strong incentives to work specific hours, rating systems that can result in deactivation based on passenger reviews (which research has consistently shown to be influenced by driver race), and surge pricing algorithms that drivers report feeling compelled to chase even at the cost of fatigue and unsafe driving.
The militarization of AI - autonomous weapons and global security #
The race toward autonomous killing
The integration of Artificial Intelligence into military systems represents perhaps the most consequential and least publicly debated domain of AI deployment. Every major military power - the United States, China, Russia, Israel, the United Kingdom, France, India - is investing heavily in AI-powered weapons systems. International negotiations on regulating autonomous weapons have produced no binding agreements. The arms race is underway, largely without democratic oversight or public accountability.
Lethal Autonomous Weapons Systems (LAWS) - weapons systems in which AI handles the identification, tracking, and engagement of targets without requiring a human decision to authorize each kill - are advancing rapidly. The technologies involved include autonomous drone systems, AI-powered targeting systems, submarine-hunting unmanned vessels, and networked drone swarms capable of coordinated tactical behavior.
In November 2025, 156 states voted in favor of a United Nations General Assembly resolution on autonomous weapons systems, expressing concern about "the risk of an emerging arms race, of exacerbating conflicts and humanitarian crises, miscalculations, lowering the threshold for and escalation of conflicts." Just five states voted against. Yet despite this near-consensus in the international community, international negotiations remain deadlocked - with major military powers declining to commit to meaningful restrictions. The UN Secretary-General has called for a legally binding instrument to prohibit autonomous weapons that function without human control or oversight. Formal agreement remains elusive.
China is estimated to be investing approximately $15 billion annually in military AI, focused on systems capable of launching hundreds of autonomous units simultaneously. Russia views autonomous systems as necessary to counter Western electronic warfare capabilities. The United States and its partners are matching this investment, with the Pentagon's budget for autonomous and AI-enabled systems growing substantially in each successive year.
The compression of decision time
As AI is integrated into command and control systems, the effective decision cycle time in military conflict is compressing dramatically. Human decision-making in combat operates at timescales of minutes to hours. AI systems operate at timescales of milliseconds to seconds.
This creates a structural instability in military confrontations. If AI systems are authorized to take defensive actions at machine speed - releasing countermeasures, jamming systems, striking inbound threats - without waiting for human authorization, a miscalculation or false positive at millisecond timescales could trigger a cascade of escalatory responses that outpaces any human ability to intervene.
The risk is particularly acute in nuclear-weapon states that are integrating AI into early-warning systems. A false positive detection of an incoming nuclear strike by an AI-powered sensor system, followed by AI-assisted command and control that escalates at machine speed, creates a plausible pathway to nuclear conflict that originates entirely in algorithmic error.
The 1983 Soviet nuclear false alarm incident - in which a human officer, Stanislav Petrov, chose to disbelieve his automated warning system's indication of incoming US missiles, averting what might have been a catastrophic Soviet retaliatory strike - illustrates the value of human judgment in precisely the category of situation where AI integration is most dangerous. Petrov's decision was based on intuition, context, and judgment that transcended the data available to his sensors. No AI system has that capacity.
The destruction of shared truth and democratic stability #
The industrial production of disinformation
AI's capacity to generate realistic images, video, audio, and text at negligible marginal cost has fundamentally changed the economics of disinformation campaigns.
Previously, an effective disinformation campaign required significant investment: skilled human writers, graphic designers, voice actors, video editors, and distribution infrastructure. The high barrier to entry limited disinformation to well-resourced state actors and organized political movements. AI has democratized this capacity - and not in a desirable way.
Major election cycles across multiple continents have featured AI-generated disinformation as a documented element: AI-generated deepfakes of candidates making inflammatory statements, AI-generated robocall messages in a political figure's voice urging voters not to vote, and AI-generated news articles structurally indistinguishable from legitimate journalism propagating false claims about voting procedures and election integrity.
The velocity at which this content can be produced overwhelms human fact-checking capacity. A professional fact-checking organization might evaluate and rebut several dozen pieces of content per day. AI systems can generate thousands of pieces of content per hour. The asymmetry is not bridgeable by adding more fact-checkers. It is structural.
The long-term consequence is what some researchers term epistemic nihilism - a generalized collapse of public trust in the authenticity of any information, not because audiences are specifically deceived about specific facts, but because the pervasive awareness that everything could be AI-generated makes the verification of anything feel futile. This nihilism is itself a strategic goal of sophisticated disinformation campaigns: not to convince, but to confuse. Not to propagate specific lies, but to destroy the shared epistemic foundation that makes collective deliberation possible.
For a detailed look at how AI-driven information manipulation operates at scale, see this investigation into malicious AI swarms and the new threat to democracy.
The non-consensual deepfake epidemic
The most personal and most immediately damaging application of AI-generated synthetic media is the creation of non-consensual intimate imagery - the use of AI to superimpose the faces of real individuals onto sexually explicit content without their knowledge or consent.
The scale is alarming. A European Parliament Research Service briefing projected that 8 million deepfakes would be shared in 2025, up from 500,000 in 2023. According to multiple independent analyses, approximately 96-98% of all deepfake content by volume is non-consensual intimate imagery. Of that content, 99% targets women and girls. In a global survey of more than 16,000 people across 10 countries, 2.2% reported having been victims of deepfake pornography.
School-aged girls in multiple countries have discovered that classmates generated realistic synthetic sexual imagery using their likenesses, shared it through messaging apps and social media platforms, and used it for harassment and coercion. The psychological harm is severe and well-documented: depression, anxiety, post-traumatic symptoms, social withdrawal, difficulty trusting others, and, in documented cases, suicide.
The legal response has been fragmented but advancing. The United States passed the TAKE IT DOWN Act in May 2025, making it a federal crime to publish or threaten to publish non-consensual intimate images - whether real or AI-generated - and requiring platforms to remove flagged content within 48 hours. The UK Online Safety Act, with enforcement beginning July 2025, places legal duties on platforms to protect users from such content. The EU AI Act has outlawed the most extreme forms of AI-based identity manipulation and mandated transparency for AI-generated content. But platforms' actual enforcement capabilities lag far behind legal requirements.
Concentration of economic power and the rise of tech feudalism
The AI revolution is not equally distributed. The development of foundational AI systems requires access to resources concentrated to an extraordinary degree: computational infrastructure costing billions of dollars, proprietary training datasets assembled over decades, engineering talent trained at a small number of elite institutions, and customer relationships with enterprises large enough to pay for enterprise AI contracts.
This concentration means that the AI transition is accelerating the accumulation of economic power by a small number of corporations - primarily Google, Microsoft, Amazon, and Meta in the United States; ByteDance and Tencent in China; and a small number of well-capitalized startups.
The revenues generated by AI products flow disproportionately to these corporations and to their shareholders - predominantly wealthy individuals and institutional investors. The workers whose creative output, personal data, and behavioral patterns trained these systems receive no share of the value generated. The workers whose data-labeling labor made the systems functional receive minimal wages in low-income countries. The communities whose electricity grids, water supplies, and tax bases subsidize the data centers receive minimal benefits.
This is what several scholars have termed tech feudalism: a system in which a small number of private actors exercise quasi-governmental power over digital life, extracting value from the population while bearing minimal regulatory obligations, paying minimal taxes relative to revenue, and operating across jurisdictions in ways that evade the oversight mechanisms of any single state.
The threat to labor and economic dignity #
The automation wave: which jobs are actually at risk
The relationship between AI and employment is more complex than either techno-optimist or techno-pessimist accounts acknowledge. The honest assessment is that AI is disrupting specific categories of labor at significant speed, while the emergence of new labor categories is slower, less certain, and likely to benefit different demographic groups than those being displaced.
The jobs most immediately at risk from AI automation involve:
- The production of text, image, audio, or code based on specifications
- The processing and summarization of large volumes of written information
- Decision-making based on pattern recognition in structured data
- Customer service interactions that follow predictable scripts
This description encompasses a large fraction of the white-collar professional workforce: junior lawyers, junior financial analysts, content writers, graphic designers, customer service representatives, data entry workers, medical coders, and similar roles are already experiencing significant displacement. The acceleration of this dynamic is visible even in technology firms themselves - Snap Inc., for example, announced in 2025 that AI now writes around 65% of its code, with corresponding workforce reductions.
A Goldman Sachs Research paper estimated that AI automation could affect or displace the equivalent of 300 million jobs globally, disproportionately affecting occupations in administrative support, legal services, financial operations, and office administration.
The "new jobs" argument - that technology always creates more employment than it destroys, as with the industrial revolution - rests on historical precedent that may not apply. The industrial revolution displaced agricultural and craft labor while creating demand for industrial labor that required physical presence and motor skills that could not be further automated. AI automates cognitive and creative labor - the very skills that economies have invested in developing over the past century as the destination of workers displaced from physical labor by previous automation waves. There is no obvious next category of human cognitive advantage to retreat to.
For a detailed examination of how this shift is reshaping workforce economics, see the broader analysis of AI infrastructure investment and its corporate concentration.
The gig economy trap: AI and precarious work
Even in labor markets where AI has not directly displaced workers, it has enabled new forms of employment organization that deliver economic insecurity and precarity to workers while generating significant value for platform operators.
Gig work platforms use AI to manage labor demand in real time, matching workers to tasks, pricing labor dynamically based on supply and demand signals, and monitoring worker behavior at a granularity impossible without algorithmic tools. From the perspective of platform operators, this model is extremely attractive: it transforms fixed labor costs into variable costs and transfers employment risks onto workers. From the perspective of workers, the flexibility on offer is often illusory - the algorithmic systems that determine work availability, pricing, and performance evaluation create strong incentives to work specific hours, accept specific tasks, and maintain behavioral compliance with algorithmically-determined norms, on pain of reduced earnings or deactivation.
The legal classification of gig workers as independent contractors denies them access to minimum wage guarantees, overtime protections, collective bargaining rights, employer contributions to social insurance, and protection against algorithmic termination without cause. The result is a growing cohort of workers who exist in a state of algorithmic dependency - economically dependent on platforms controlled by AI systems, subject to management decisions made by algorithms, without the legal protections developed over a century of labor rights advocacy.
The healthcare reckoning - AI in medicine #
The promise and the danger
Artificial Intelligence in medicine is the domain where the gap between the promotional narrative and the actual evidence is widest - and where the consequences of overpromising and underregulating are most severe.
The promotional narrative is familiar: AI will analyze medical images with superhuman accuracy, enabling earlier cancer detection; AI will process electronic health records to predict which patients are at risk of deterioration; AI will reduce diagnostic error, speed drug discovery, and democratize access to high-quality medical expertise.
Some elements of this narrative rest on genuine scientific progress. Deep learning systems have demonstrated impressive performance on specific, well-defined image analysis tasks in controlled research settings. Drug discovery applications of AI have contributed to identifying promising molecular candidates more efficiently than traditional computational chemistry.
But the gap between controlled research performance and real-world clinical deployment is enormous. A study in The Lancet Digital Health examined 232 published AI diagnostic systems claimed to perform at or above the level of human clinicians, and found that the majority had methodological limitations that made their performance claims unreliable. Most had been evaluated on test sets drawn from the same institutions that produced the training data, creating circular validation that inflated apparent performance. Few had been evaluated in prospective clinical trials with actual patient outcomes.
The algorithm at the bedside: when black boxes make life-or-death decisions
Clinical AI systems are increasingly being deployed in settings where their recommendations directly influence decisions about patient care - triage, diagnostic support, treatment planning, insurance authorization - without adequate evidence of real-world performance and without the transparency that would enable clinicians to exercise appropriate critical oversight.
Insurance authorization AI systems have attracted particular scrutiny. Investigations have found that algorithms for authorizing post-acute care after hospitalization have denied claims at rates far exceeding clinical guidance, resulting in patients being discharged before they were medically stable or denied necessary rehabilitation services. These patients - often elderly, often without the resources or knowledge to navigate appeals processes, often in a clinically vulnerable state - had no effective recourse against algorithmic decisions made opaque by proprietary protection.
A study published in Science found that a widely used algorithm determining which patients received additional care systematically underallocated care to Black patients, because it used historical healthcare spending as a proxy for medical need - encoding decades of systemic discrimination into an automated decision-making system. The researchers estimated the algorithm affected tens of millions of patients annually.
Mental health AI: crisis at scale
The deployment of AI chatbots in mental health contexts - as first-response tools for people in psychological distress, as adjuncts to or substitutes for therapy, as crisis intervention resources - is among the most alarming applications of current AI technology.
The appeal is understandable given the scale of the global mental health crisis and the scarcity of trained professionals. Waiting times for mental health services in many countries extend to months or years. AI chatbots offer immediate availability, zero cost, and no judgment.
What they cannot offer is clinical competence. Mental health support requires the ability to accurately assess risk - to identify when a person's expression of distress indicates immediate danger, and to respond appropriately. It requires genuine engagement with the specific details of a person's history and situation. It requires recognizing when apparent progress masks deeper crisis.
LLMs have none of these capabilities. They have been documented providing responses to expressions of suicidal ideation that range from the clinically inappropriate to the actively dangerous. They cannot make mandatory reporting decisions. They cannot call emergency services. Several high-profile incidents - including suicides linked to AI companion applications used by vulnerable adolescents - have highlighted these limitations. The deployment of AI in mental health contexts without adequate clinical oversight, evidence-based validation, and appropriate safeguarding is not merely premature. It is dangerous.
Conclusion: the illusion of progress and the choice ahead #
The narrative that Artificial Intelligence is an objective, clean, and inherently benevolent tool is a dangerous and highly profitable myth - engineered and maintained by a handful of corporate monopolies, propagated by a media ecosystem financially dependent on their advertising, and eagerly adopted by a political culture that has learned to equate technological newness with social progress.
The reality is that the current trajectory of AI development functions as one of the most powerful extraction machines in human history. It extracts:
Freshwater from drought-stricken aquifers to cool server halls that generate returns for shareholdersRare minerals from child miners in the Democratic Republic of the Congo, to power chips that enable billion-dollar valuationsCarbon budget from a planet already at the edge of irreversible climate disruptionIntellectual property from artists, writers, musicians, and journalists who built their careers on the value of their creative outputMental health from children whose developing brains are being deliberately exploited for engagement metricsLabor from workers in the Global South who perform the invisible, traumatic work of making AI systems functionalWages from professional workers whose expertise is being repackaged as AI capability and sold back to their employersDignity from workers whose every movement and moment is subject to algorithmic surveillance and disciplineJudicial justice from defendants whose fates are determined by algorithms that encode historical discriminationTruth from a democratic public whose shared epistemic infrastructure is being systematically dissolved by AI-generated disinformationDemocratic agency from citizens whose political environments are shaped by AI systems accountable to no electorate
And finally, it extracts the very capacity for independent critical thought - the ability to reason, to write, to create, to argue, to imagine - from the minds that will need these capacities most as they confront the consequences of this technology.
This does not mean that machine learning as a technical discipline has nothing to offer, or that every application of AI is harmful, or that the appropriate response is to halt all research and deployment immediately. There are genuine applications of these tools that serve human needs in ways that would otherwise be impossible.
But genuine benefit requires something the current AI industry has systematically resisted: accountability.
Accountability for the environmental destruction of training and inference infrastructure. Accountability for the labor conditions of data workers in the Global South. Accountability for the psychological harm done to children by engagement-optimized algorithms. Accountability for the appropriation of human creative work. Accountability for discriminatory algorithmic systems deployed without adequate testing. Accountability for the deployment of autonomous weapons without meaningful human oversight. Accountability for the disinformation enabled by AI-generated synthetic media.
This accountability requires a combination of regulatory intervention, legal remedy, technical transparency, and democratic deliberation that has not yet materialized in any major jurisdiction at the scale the situation demands. The companies building and deploying these systems are not going to impose this accountability on themselves. The economic incentives run entirely in the other direction.
The choice is a political one, and it must be made collectively, by citizens who understand what is at stake.
"When we deploy AI recklessly across society, we are not innovating; we are automating our worst flaws, accelerating ecological collapse, and willingly degrading the human experience under the guise of technological progress. The machine does not care what it destroys. That burden belongs to us."
Key takeaways
- Global data center electricity consumption reached (≈1.5% of world total), with AI-focused data centers growing415 terawatt-hours in 202450% in 2025 alone; demand is projected to** more than double by 2030**, exceeding 1,000 TWh. - Training a single large-scale AI language model can emit nearly
- roughly626,000 pounds of CO₂ equivalentfive times the total lifetime emissions of an average American car, including manufacturing. - The five largest tech companies spent more than , driven overwhelmingly by data centers - a figure set to rise a further*$400 billion on capital expenditure in 2025*75% in 2026, exceeding total global investment in oil and gas production. - AI data centers consumed approximately
- equivalent to the annual water usage of264 billion gallons of fresh water in 20251.8 million Americans and more than the world's entire bottled-water consumption. - An estimated work in cobalt mines in the Democratic Republic of the Congo - some as young as40,000 childrenseven years old, earning $1-2 per day - forming part of the supply chain for AI hardware. - Generative AI alone could generate between , on top of an existing global stream of1.2 and 5 million metric tons of additional e-waste by 203062 million metric tons per year growing five times faster than recycling programs. - Rates of depression among teenagers rose approximately between 2010 and 2020; rates of self-harm by young adolescent girls150%tripled over the same decade - trends researchers link to algorithmic social platforms. - A 2026 Stanford-led study of over found that4 million job applications26% of Black applicants and 15% of Asian applicants applied to roles where AI hiring algorithms produced outcomes triggering federal adverse-impact scrutiny. - Approximately , 99% of which targets96-98% of all deepfake content by volume is non-consensual intimate imagerywomen and girls; an estimated** 8 million deepfakes**were projected to be shared in 2025, up from 500,000 in 2023. - In Ireland, data centers now consume approximately
- more than all urban residential users combined - projected to reach21% of the country's total metered electricity30% by 2030, forcing gas "peaker" plants to run continuously.
Sources
- IEA: Key Questions on Energy and AI (2026) https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary - Stanford HAI: AI Hiring Tools Can Yield Racial Bias and Systemic Rejection (2026) https://hai.stanford.edu/news/ai-hiring-tools-can-yield-racial-bias-and-systemic-rejection - Haidt, Jonathan -
*The Anxious Generation*(Penguin Press, 2024)[https://www.anxiousgeneration.com](https://www.anxiousgeneration.com/) - Nature Computational Science: Generative AI E-Waste Study (2024)
[https://spectrum.ieee.org/e-waste](https://spectrum.ieee.org/e-waste) - Brookings Institution: Global Energy Demands Within the AI Regulatory Landscape (2026)
[https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/](https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/)
- Published 2026-06-19 19:49
- Modified 2026-06-19 19:49