What the briefing says
The report, Unlawful by Design: Exposing the Human Rights Costs of Generative AI, focuses on how companies train their models: by scraping billions of public online posts, images, and social media activity — typically without the explicit consent of the people who created or appear in that content.
Amnesty examined the data practices behind several of the most widely used AI tools: GPT-3 (OpenAI), Gemini (Google), Llama (Meta), DeepSeek, Midjourney, and Stable Diffusion.
The core argument is that the privacy violation isn't incidental — it's structural. The systems can't be built the way they're currently built without extracting personal data at scale and without consent.
The other harms the report flags
Beyond privacy, Amnesty identifies two other categories of harm:
Bias and discrimination. Because training data is pulled largely from the web, it carries real-world biases — racial, gendered, cultural. At scale, those biases get amplified, not averaged out. Marginalized communities bear a disproportionate share of the harm.
Environmental cost. Google's own 2024 sustainability report noted a 48% increase in greenhouse gas emissions since 2019, attributed to data center and supply chain growth. Microsoft's emissions rose 29% between 2020 and 2024. Communities in Chile, Mexico, and Arizona are already resisting data center construction in drought-stressed regions.
What companies said
Amnesty wrote to Google, OpenAI, Meta, Stability AI, Midjourney, and DeepSeek before publishing. Only Microsoft, Amazon, Intel, OpenAI, and Meta responded. Their responses are summarized in the briefing.
What Amnesty is calling for
A prohibition on standalone generative AI systems built on unlawful web scraping, and immediate cessation of non-consensual personal data collection for AI training. The briefing frames this as a regulatory failure as much as a corporate one.
Up next · Wednesday
An overview of Optery — one of the most widely used automated data broker removal services.