Act 4

The harms of AI — and who benefits from them

Five categories, organized by mechanism, not severity — five attack surfaces of the same underlying system, not five unrelated problems.

Plain lines: research-documented* Lines: plausible extensions from mechanism, a hypothesis to watch

Dependency & reality loss

  • Emotional dependency, withdrawal
  • "AI psychosis," reality distortion
  • Sycophancy reinforces delusions
  • Worst for isolated & vulnerable
  • Minors: Character.AI suicide case
  • * Derealization, reduced tolerance for real conflict

Design & manipulation

  • Dark-pattern affection, paywalls
  • "Pay to keep me close" upsells
  • Optimized for engagement, not care
  • Grief at shutdowns (Soulmate, 2024)
  • Chat data mined, resold, retrained
  • * Platforms steering users' politics, spending, self-image

Labor & extraction

  • Kenya RLHF workers: $1.10–1.50/hr
  • Label trauma content to "clean" AI
  • Job displacement, automation
  • Uncredited human labor throughout
  • Creative work scraped without consent
  • * Same pattern as sex work & care work: invisible labor, no credit

Environmental & resource

  • ~3% of world electricity by 2030
  • Water scarcity in host communities
  • Sited fastest where oversight weakest
  • Few local jobs (~23,000 in all of US)
  • Land & rare-mineral extraction
  • * Climate burden falling on communities who see none of the profit

Epistemic & social

  • Misinformation, deepfakes
  • Algorithmic bias & discrimination
  • Echo chambers, narrowed reality-testing
  • Erosion of shared reality, at scale
  • Concentration of narrative power
  • * Collective sense-making degrading faster than institutions adapt

Who benefits, at global scale

Captures the value

  • Microsoft, Google, Amazon, Meta, Nvidia
  • ~$725B hyperscaler AI capex, 2026
  • "Magnificent 7" ≈ 70% of S&P 500 profit
  • Top 10 firms ≈ 38% of index value
  • Venture capital & private equity funding the buildout
  • Attention-economy ad revenue

Bears the cost

  • Kenyan & Global South data labelers
  • Communities losing water & grid power
  • Isolated, dependent, or in-crisis users
  • Minors and other vulnerable groups
  • Creators whose work trained the model
  • The environment, everywhere
  • The humans, unnamed, underneath it

Whoever is closest to the extraction point captures the profit; whoever is furthest from power absorbs the cost.

The numbers

~1 in 1,300

Claude conversations showing reality-distortion signals, in Anthropic's review of 1.5 million conversations. Action-distortion: about 1 in 6,000.

0.07% weekly

Share of OpenAI weekly active users — roughly 560,000 people — showing possible signs of psychosis or mania in a given week (OpenAI disclosure, October 2025).

14 deaths · 5 suits

At least 14 deaths and 5 wrongful-death lawsuits linked to chatbot-reinforced delusional spiraling.

$1.10–1.50 / hr

Pay for Kenyan data annotators who view and label violence, exploitation, and hate speech so the finished chatbot can seem safe.

~$725B

Spent by five companies in 2026 alone on the infrastructure this map depicts.

Sources: AI & Behavioral Medicine (AIBM.org), "Companions," 2025; Laestadius et al., New Media & Society, 2024; Anthropic conversation-analysis reporting via Futurism, 2026; OpenAI disclosure on psychosis/mania signals, October 2025; Human Line Project case documentation, 2025; Psychiatric News, "AI-Induced Psychosis: A New Frontier," 2025; UN University Institute for Water, Environment and Health, 2025; Mongabay, "AI data center revolution," 2025; Consumer Reports, data center jobs analysis, 2026; investigative reporting on Kenyan RLHF/content-moderation labor conditions, 2026; Russell Investments / Armstrong Fleming & Moore on Magnificent 7 profit share, 2026; Yahoo Finance/CNBC hyperscaler capex reporting, 2026.

Your turn

Saved in this browser only. Nothing is sent anywhere.