AI - Part 3: The Ugly — Existential and Systemic Harms

This is the third part of a 4 part post, the previous post is here.

Beyond the immediate impacts on workers and labour markets lies a deeper set of harms. These are not marginal side effects but structural consequences of AI's material footprint and its integration into the systems that shape daily life. They affect not only workers but entire communities, and they raise questions about the long-term compatibility of the current AI trajectory with human and ecological wellbeing.

The Environmental Toll: Water, Carbon, and Unequal Burdens

AI, like most software needs to run on hardware. It is a physical infrastructure of server farms, cooling systems, and electrical grids, and its environmental demands are escalating rapidly. A 2025 study published in Nature Sustainability estimated that AI servers in the United States alone would generate annual water footprints of 731 to 1,125 million cubic metres and additional carbon emissions of 24 to 44 million tonnes of CO₂-equivalent between 2024 and 2030, depending on deployment scale. The study concluded that the AI server industry is unlikely to meet its net-zero aspirations by 2030 without substantial reliance on "highly uncertain carbon offset and water restoration mechanisms".₁

A separate peer-reviewed paper projected that global data centre electricity demand may nearly double, from approximately 415 TWh in 2024 to 945 TWh by 2030, with AI workloads accounting for a disproportionate share of growth. In the United States, AI servers alone are expected to drive annual increases in water consumption of 200–300 billion gallons. Goldman Sachs Research has forecast that approximately 60% of the increased energy demand will be met by burning fossil fuels, adding roughly 220 million tonnes of carbon emissions that would not otherwise have occurred.₂

The water figures are particularly striking. One study estimated that AI data centres could consume 312–765 billion litres of water per year—a quantity exceeding the entire global bottled water industry's annual output. Most of this consumption is driven not by model training but by inference: the millions of daily queries, image generations, and always-on assistants that keep servers running continuously. Despite efficiency improvements in hardware, the sheer growth in demand has overwhelmed any per-unit gains, meaning better technology is leading to greater absolute resource consumption, not less.₃

These environmental burdens are not distributed equally. The communities hosting data centres—often rural or lower-income areas—bear the costs of strained water resources, grid instability, and local pollution while receiving few of the economic benefits generated by the facilities. In drought-prone regions, the water demands of data centres compete directly with agricultural and residential needs, creating conflicts that technology companies have thus far been able to externalise.

The Carbon Accounting Gap

A major obstacle to addressing these harms is transparency. Companies generally do not distinguish between AI and non-AI workloads in their environmental reporting, making it difficult for researchers and policymakers to assess the true impact of AI expansion₃. The Nature Sustainability authors noted that the holistic energy-water-climate implications of AI computing remain "largely unknown, constrained by untransparent industry reports and limited data". This opacity is not accidental; it shields the industry from accountability while growth continues unchecked.₁

Algorithmic Discrimination: Locking in Historical Inequality

When AI systems are deployed in housing, hiring, and credit decisions, they do not simply reflect existing biases—they automate and scale them. A large-scale audit of GPT-4's housing recommendations examined 168,000 prompts across ten U.S. cities, varying demographic characteristics protected under fair housing laws. The researchers found "evidence of racial steering, default whiteness, and steering of minority homeseekers toward neighbourhoods with lower opportunity indices"—all of which "could have the effect of exacerbating segregation in already segregated cities"₄

This is not a glitch. These systems are trained on data that encodes centuries of structural inequality, and without explicit intervention, they reproduce those patterns. In tenant screening, algorithmic tools have been shown to produce starkly different outcomes by race: Black and Latino renters were significantly less likely to have applications accepted when AI screening was used. In mortgage lending, research has found that otherwise equal Black and Latino borrowers were charged more, costing them an estimated $765 million annually in excess payments.

The mechanism is insidious precisely because it presents discrimination as neutral computation. When a human landlord or loan officer exhibits bias, there is at least the theoretical possibility of accountability. When an algorithm does the same thing, the decision is cloaked in the language of objectivity, making it harder to challenge and easier to defend.

Mental Health: The Algorithmic Shaping of Consciousness

The integration of AI into social media platforms has created a vast, unregulated experiment in behavioural modification, with adolescents as the primary subjects. AI-driven recommendation algorithms curate highly personalised content designed to maximise engagement. Research published in the Asian Journal of Psychiatry documented that this personalisation, while enhancing user experience, raises "significant concerns regarding adolescent mental health, including increased anxiety, depression, self-esteem issues, and body dissatisfaction"₅.

The mechanisms are well understood. Recommendation systems amplify and normalise harmful content, increasing young users' exposure to radical material and intensifying social comparison. A study of the AI-Social Comparison-Well-Being Framework found that algorithmic exposure and social comparison tendencies combined to produce negative psychological outcomes, with younger participants consistently reporting higher levels of anxiety, depression, and algorithmic reinforcement than older groups. The systems are designed to hold attention, and what holds attention most effectively is often what is most psychologically damaging.

The mental health costs extend to workers. Content moderators—the people who filter the worst material so that platforms remain usable—experience secondary trauma on a scale that the clinical literature is only beginning to document. A case report published in European Psychiatry described a 35-year-old woman with no prior psychiatric history who developed severe PTSD after five years of content moderation, including daily panic attacks, intrusive images, insomnia, and avoidance behaviours that interfered with her capacity to work and parent₆. A larger replication study found that over a quarter of commercial content moderators demonstrated moderate to severe psychological distress, and a quarter were experiencing low wellbeing.₇

This psychological harm is a direct transfer of cost from technology companies to the workers they employ. The platforms profit from user engagement while the moderators—often employed through subcontractors in countries with weak labour protections—absorb the trauma that makes that engagement possible. It is a form of moral externalisation: the harm is someone else's problem, even as it is produced by the company's core business model.

Systemic Fragility and Automation Cascades

A final category of harm is less visible but potentially more consequential: the introduction of systemic fragility through cascading automation. When organisations remove human decision-makers from critical workflows, they create single points of failure that can propagate errors at machine speed. The "human-in-the-loop" safeguard that many organisations rely on is already failing due to approval fatigue, overtrust in automation, and the mismatch between human intent and machine logic. Critics argue that keeping humans in the loop creates a bottleneck, but removing them entirely creates the risk of catastrophic failure without warning—failures that, in sectors like healthcare, energy, and finance, could affect millions of people simultaneously.

Next part is here.

Brainmade

Sources

Errata and Modifications

fixed a typo: AI is run on hardware, not software