Three AI agents, two countries, and one uneven world wide web

A researcher tested three AI agents (GPT, Claude, and Muse) on their ability to research and update World Bank data for the U.S. and Iran, examining how they handle multilingual tasks, web access, and human oversight. The study found significant differences in how each agent manages human-in-the-loop permissions, source access, and transparency, with GPT requiring minimal user intervention while Claude requested permission repeatedly and Muse presented observability challenges. The research highlights how language, context, and country representation create an uneven experience across AI agents, with implications for equitable access to AI-powered research tools globally.