From Fastcompany · · 3m
AI chatbots fail to grasp cultural nuance, study shows
Cohere Labs tested 28 AI models on cultural riddles and found even top performers struggle with context rooted in lived experience.
In brief
Cohere Labs tested 28 AI models on cultural riddles and found even top performers struggle with context rooted in lived experience. AI models struggle to understand cultural context beyond facts, rarely admitting when they don't know answers, according to a Cohere Labs study of twenty-eight models tested on cultural…
Cohere Labs tests AI on culture
Researchers at Canadian AI lab Cohere Labs tested whether chatbots could understand cultures beyond simply speaking their language. They presented models with cultural riddles to assess deeper knowledge.
The Berlin Wall riddle
One riddle asked: What's the German nine slash eleven? Out of twenty-eight AI models, only one knew the answer. Most were confused by the date format, answering the twin towers instead.
“What's the German 9/11?”
Why the answer matters
In Europe, the day comes first, then the month, so nine slash eleven is November nine, the day in nineteen eighty-nine when the Berlin Wall fell. The models missed this cultural context.
Facts versus understanding
Many existing AI benchmarks treat culture as static facts to memorize or look up. But knowing facts about a culture isn't the same as understanding how people within it make sense of the world.
Why riddles work
Researchers chose riddles because they require models to connect seemingly unrelated pieces of cultural knowledge and put them to use. This tests understanding rooted in lived experience, not just memorized facts.
Julia Kreutzer on nuance
Julia Kreutzer, staff research scientist at Cohere Labs who led the benchmark development, says the study captures fuzzy concepts hard to verify. This is where real-life complexity gets in.
“This is where the real-life complexity gets in.”
The study's scope
Kreutzer and her team spent six months working with volunteers worldwide to collect riddles, test AI models, and evaluate answers.
Evaluating the answers
The volunteers evaluated and labeled sixty-eight thousand one hundred fifty-four unique answers from the models. Researchers then used expert evaluations to tune an AI model that could judge answer quality where assessments were missing.
Google Gemini leads the pack
Google's Gemini three point one Pro Preview earned the highest score at seventy-nine point four percent. OpenAI's GPT five point five placed second at seventy-six point three percent.
Claude Opus ranks third
Anthropic's Claude Opus four point seven Thinking followed in third place at sixty-six percent. Even the best-performing models failed to solve many riddles rooted in local cultures.
Models won't admit ignorance
Beyond poor performance, the study found that AI models rarely admitted when they didn't know the answer. This reluctance to acknowledge gaps in cultural knowledge poses a risk for users relying on chatbots.
What to remember
AI models struggle to understand cultural context beyond facts, rarely admitting when they don't know answers, according to a Cohere Labs study of twenty-eight models tested on cultural riddles.
