AI-Written Emails Are Getting Caught as Spam More Often Than Human Ones
New research from Folderly finds AI-written cold emails hit spam filters more often and get lower open and reply rates than human copy.
GiveDirectly's Rwanda Pilot Shows Frontier AI Still Fails the World's Poorest
By Prompt AI News2 min read
#global-ai#llm#givedirectly#africa
As The Verge first reported on findings circulating from GiveDirectly's field research, a trial deploying frontier AI chatbots alongside cash transfers in rural Rwanda found that recipients wanted the technology — and that it kept letting them down. Users asked for business advice, second opinions on financial decisions, and practical guidance they had no other way to access. The AI consistently failed to deliver: language gaps, culturally irrelevant suggestions, and outright wrong answers surfaced across the pilot.
The findings matter beyond Rwanda. The AI industry's standard benchmarks are rich in data about wealthy, English-speaking users in connected cities. Deploying those same models into Kinyarwanda-speaking communities with different economic contexts and limited internet access exposes brittleness that leaderboard scores don't capture.
GiveDirectly's experiment is one of the few systematic attempts to test whether today's most capable models can serve the roughly 4 billion people outside the core training distribution. The honest answer from this pilot: they can't — not yet, and not without significant localization work that most AI labs have not prioritized.
The recipients who were most enthusiastic about the idea were also the ones most disappointed by the reality. That gap is a design failure, not a market one.
Leave a Comment
All comments are reviewed before appearing. Keep it respectful.
More in Commentary
View all →Conservative Coalition Targets AI Data Centers in Nationwide Protest Campaign
Right-leaning rural groups are organizing multi-state protests against AI data center expansion, citing power draw, water use, and local grid strain.
Open-Source Tool Queries 11 AI Models at Once to Filter Hallucinations
A developer's open-source Quorum system cross-checks 11 LLMs simultaneously, surfacing answers only when a supermajority agrees — trading cost for reliability.