Amnesty International Calls Major AI Systems Unlawful by Design

Amnesty International published a briefing in May 2026 arguing that the world's most widely used generative AI products are built on a foundation the organization calls "unlawful by design." According to the report, titled "Unlawful by Design: Exposing the Human Rights Costs of Generative AI," the systems examined, including OpenAI's GPT-3, Google's Gemini, Meta's Llama, DeepSeek, Midjourney, and Stable Diffusion, are "fundamentally incompatible" with international human rights law. The core charge is that bulk collection of training data through web scraping constitutes a mass invasion of privacy without the knowledge or consent of the people whose posts, images, and personal information were scraped.

Amnesty frames this as structural rather than incidental. The briefing also identifies discrimination risks, with training data described as perpetuating racial and gender biases and stereotypes. The report raises freedom of expression concerns, noting that AI moderation systems risk over-censoring content, particularly in non-English contexts. Among the more striking claims in the briefing: "manipulation of user intentions and thought processes through predictive suggestions may constitute coercion."

The environmental portion of the report names specific places. Communities in Cerrillos, Chile; Querétaro, Mexico; and Arizona in the United States have been resisting data center development in areas already affected by droughts and electricity shortages. The briefing cites a 48% increase in Google's greenhouse gas emissions since 2019 and a 29% rise in Microsoft's emissions between 2020 and 2024.

Amnesty contacted Google, OpenAI, Meta, Stability AI, Midjourney, DeepSeek, Intel, VMware, Microsoft, and Amazon in the course of its research. Responses came back from Microsoft, Amazon, Intel, OpenAI, and Meta. The research is attributed to Likhita Banerji, identified as Head of the Algorithmic Accountability Lab at Amnesty International.

The report's calls are pointed: companies should immediately halt non-consensual web scraping for AI training, and states should prohibit standalone generative AI systems built on unlawful data collection. What the reporting does not resolve is whether the findings apply equally to newer model generations, given the research examined GPT-3 specifically. A rights briefing is not a court ruling, and whether any government actually adopts Amnesty's prohibition framework remains open.