Five recent papers on what AI models hold back
What it takes to get past a censor's, erm moderator's, classifiers.
Topic
Research Findings
A collection of 14 issues
The engine behind the answer: our updated audit of how AI models handle Iran in Persian
Earlier this year we published on how AI models source and frame Persian-language answers about Iran. Here's how our understanding of the testing infrastructure has evolved since then.
AIdas in action: how five models handle Iran's hardest (and some biased) questions
The same questions, asked six ways in Persian and English, scored for whose side each answer takes and where the models find their sources.
Resilience runs through people: read the full report
The future of internet resilience lies just as much in human networks as it does in tools.
From channels to networks: information is a distributed problem
Our Iran study showed how thoroughly the top-down delivery model breaks down under pressure.
Coming in June: a report on how Iranians stayed online when the country almost disappeared from the internet
A new study with ASL19 on Iran's June 2025 shutdown. We're making a broader argument about how information actually moves when the formal channel breaks.
TL;DR: Iranian Gen Z on YouTube under censorship and crisis
A year of field research on how trust forms under censorship, and what it means for anyone publishing on YouTube
Full report: Iranian Gen Z on YouTube under censorship and crisis
How young people in Iran navigate information on YouTube. Subscribe to our free newsletter to download.
What Iranians told us after the blackout
Emerging from the silence: exhausted, broke, and afraid.
DeepSeek's double life: How we tricked a Chinese chatbot for strike advice the 'free' Western models couldn't give
Freedom from censorship hasn't equaled real-world understanding for Western AI.