Overview
- Amnesty published a briefing Thursday that says many major standalone generative AI systems were trained on large‑scale, non‑consensual web scraping that violates users’ right to privacy.
- The research examined models behind GPT‑3, Gemini, Llama, Midjourney, Stable Diffusion and others and found that scraped datasets carry racial, gender and cultural biases that show up in model outputs.
- Amnesty links the growth of these systems to higher energy, water and material demands, citing company sustainability figures that record rising greenhouse gas emissions and local pushback against new data centres.
- The organization contacted major firms and infrastructure providers during its research and received limited responses from Microsoft, Amazon, Intel, OpenAI and Meta, highlighting transparency gaps.
- Amnesty calls on states to ban standalone generative AI built on unlawful scraping and urges companies to stop non‑consensual mass data collection, a move that could increase legal and regulatory pressure on the industry.