AI systems trained on historical data reproduce and amplify historical inequities. Amazon's internal hiring AI downgraded resumes from women because it trained on hiring data from a male-dominated industry. AMZN Facial recognition systems have false positive rates 10-100x higher for Black and East Asian faces. NIST Healthcare algorithms systematically under-assigned Black patients to care programmes. The problem isn't that AI is racist — it's that the data it learned from reflects societies that were.

Clearview AI scraped 30 billion facial images from social media without consent and sold law enforcement access to a system for identifying anyone from a photo. CLAIR Employee monitoring tools using AI track keystrokes, webcam activity, and infer emotional states. AI-powered advertising systems build detailed behavioural profiles from activity across thousands of websites. People have largely not consented to this surveillance architecture — it emerged through incremental product decisions, not collective choices.

Generative AI dramatically reduces the cost of creating convincing false content. AI-generated audio of politicians saying things they never said has circulated before elections in multiple countries. Deepfake pornography of non-consenting individuals is a documented harm. AI-generated "news" articles with fabricated quotes have been published as legitimate journalism. The information verification problem that predated AI has been made structurally harder. STAN

Generative AI models are trained on copyrighted books, images, music, and code without permission or compensation to creators. Artists have found their distinctive styles reproduced by AI image generators. Authors including Sarah Silverman and George R.R. Martin have filed suit against AI companies over training data. Getty Images sued Stability AI over images. The legal question — does AI training on copyrighted content constitute infringement? — is actively being litigated and remains unresolved. GETTY

When an AI system makes a wrong decision that harms someone, who is responsible? The algorithm? The company that built it? The company that deployed it? The user who provided the input? Current legal frameworks weren't designed for distributed, opaque AI decision-making. A parole algorithm that contributed to excessive incarceration, a mortgage AI that discriminated by race, a hiring tool that systematically excluded qualified candidates — accountability is murky in each case.

The economic disruption from AI automation is real and uneven. Entry-level knowledge workers face the steepest task displacement. Junior writers, coders, customer service agents, data analysts — roles that provided entry points to careers are being automated first. The workers displaced tend to be younger, lower-paid, and less protected. The gains accrue to the companies deploying AI. WEF The net employment effect may be positive in aggregate — but aggregate statistics don't pay rent for individual people whose roles have been eliminated.

The most capable AI systems require hundreds of millions of dollars to train, enormous data centres, and proprietary datasets. This creates structural advantages for a small number of companies — primarily Google, Microsoft/OpenAI, Meta, Amazon, and Anthropic. The companies that control AI infrastructure have unprecedented leverage over knowledge production, communication, and economic activity. The open-source movement (Meta's LLaMA, Mistral) provides partial counterweight, but the gap between frontier closed models and open models remains substantial. STAN2

Training large AI models consumes significant computational resources. GPT-3's training run emitted roughly 552 tonnes of CO2 equivalent. ENV Data centre electricity consumption for AI inference at scale is growing rapidly — Goldman Sachs projected AI data centre power consumption would increase 160% by 2030. The environmental cost is real, concentrated, and largely borne by communities near data centres rather than by the companies or users benefiting.