Why AI Struggles with Persian Social Cues and What It Means
327When you’re in Iran, understanding social cues is a big deal. One of the trickiest parts is taarof, a way of showing politeness that often confuses outsiders. If an Iranian taxi driver waves away your payment, saying, “Be my guest this time,” it doesn’t mean they’re giving you a free ride. They actually expect you to insist on paying several times before they accept your money. This back-and-forth dance is so ingrained that it shapes daily interactions across Persian culture.
But AI chatbots? They’re pretty bad at this. A new study shows that popular language models from companies like OpenAI, Meta, and Anthropic only get taarof right about a third of the time. Compare that to native Persian speakers, who get it right over 80 percent of the time. That’s a huge gap, and it highlights how AI models struggle with cultural nuances that aren’t obvious just from words.
Understanding Taarof and Its Cultural Depth
Taarof isn’t just about politeness; it’s a complex system of ritual exchanges. It involves repeated offers, polite refusals, and deflecting compliments. For example, if someone offers you a gift, the polite move might be to decline several times before accepting. Saying “Thank you” directly or accepting an offer immediately can be seen as rude or boastful. It’s a delicate dance of offer and resistance that people learn from a young age.
Researchers point out that this cultural practice relies heavily on implicit cues. Words often don’t tell the whole story. Instead, listeners decode meaning based on shared context, tone, and cultural knowledge. That’s why AI models, which mainly pattern-match based on explicit data, stumble when faced with taarof. They tend to interpret “yes” and “no” literally, missing the subtlety that “no” might actually mean “yes,” or that insisting can be a form of politeness, not coercion.
AI Fails to Read Cultural Signals
The study tested how well AI models handle taarof by giving them scenarios and measuring their responses. They found that even responses flagged as polite often didn’t align with Persian norms. For example, an AI might accept an offer straight away or compliment someone without deflecting, which would be considered rude in Iran. This mismatch can cause misunderstandings or even offend people if AI is used in customer service or diplomatic contexts.
Interestingly, when the models switched to Persian language prompts, their understanding improved significantly. For instance, DeepSeek V3’s accuracy jumped from about 37 percent to nearly 69 percent, and GPT-4’s scores rose by over 33 points. The language switch seems to activate different cultural patterns embedded in the training data, helping the models better grasp taarof. However, smaller models like Llama 3 and Dorna improved less, by about 12 points, showing that not all models benefit equally from language changes.
Who Understands Taarof Best?
The study involved 33 human participants divided into native Persian speakers, heritage speakers (who grew up with Persian at home but were educated mainly in English), and non-Iranians. Native speakers achieved about 82 percent accuracy in taarof scenarios, setting the upper limit. Heritage speakers scored around 60 percent, while non-Iranian participants and AI models hovered around 42 percent, nearly matching the less accurate AI responses.
Non-Iranian humans often interpret taarof through a Western lens. They tend to see politeness as directness, mistaking polite refusals for rudeness or vice versa. They also tend to respond differently to offers and compliments, often missing the subtle cues that define Persian etiquette. AI models show similar patterns—they avoid responses that might seem rude and tend to support stereotypes, like assuming men should pay or that women shouldn’t be left alone, even when such stereotypes aren’t relevant to taarof norms.
Another interesting finding was gender bias in AI responses. The models responded more accurately to women than men, with GPT-4 achieving about 44 percent accuracy for women versus just 31 percent for men. This pattern reflects stereotypes embedded in the training data, where models often assume stereotypical gender roles in their responses, even when the context calls for cultural sensitivity.
Beyond just documenting the problem, researchers explored whether AI can learn taarof. They experimented with targeted training and found that models could improve their understanding significantly. This suggests that with the right cultural data and training methods, AI systems could eventually better handle complex social norms like taarof, making cross-cultural communication smoother and less prone to missteps.
In the end, this research highlights a fundamental challenge: AI models trained mainly on Western communication patterns often fail to decode the unspoken rules of other cultures. As AI becomes more involved in global interactions, understanding these nuances isn’t just a matter of politeness—it’s essential for respectful and effective communication across cultural divides.
What do you think?
It is nice to know your opinion. Leave a comment.