The International AI Safety Report 2025, written by more than 100 global experts, argues that algorithmic bias is no longer a theoretical concern, it is producing measurable social harm. The report identifies discrimination in general-purpose AI systems as a “well-evidenced” and unresolved problem, affecting areas such as hiring, lending, healthcare, and law enforcement. What makes this especially concerning is that AI systems can amplify existing inequalities at scale.

The UNESCO CULTAI report suggests that algorithmic bias is no longer just a technical problem, it is becoming a question of cultural power. Generative AI systems increasingly shape what people read, create, and recognize online, yet the cultures represented within these systems remain deeply unequal. The report highlights the underrepresentation of African communities, local languages, and non-Western cultural perspectives in major AI training datasets. As a result, AI system

The Brookings analysis introduces a critical shift in how algorithmic bias should be understood. Most discussions focus on discrimination, when AI systems make unfair decisions about people. But the report argues that another form of harm is becoming increasingly important: algorithmic exclusion, where AI systems fail to “see” certain populations at all because of missing or fragmented data. This reframes bias as more than a technical error. AI systems are built on data

This week’s Field Notes highlight a central shift: Big Tech is no longer just shaping technology, it is shaping the conditions under which human rights are experienced. Across multiple analyses, a consistent pattern emerges. Platforms control access to information, economic participation, and public discourse, yet accountability mechanisms remain weak. Data shows this imbalance clearly. Major tech companies continue to score below 50% in human rights accountability benc

The BSR report reframes the conversation around Big Tech and human rights by shifting focus from individual AI models to the entire generative AI value chain. Instead of asking whether a single platform is harmful, it shows how risks are embedded across interconnected stages, from data collection and model training to deployment and everyday use. What stands out is how harm is distributed. Human rights risks begin long before users interact with AI. Low-paid data worker