AI Speaks Persian, But Doesn't Think Like an Iranian

AI models may appear fluent in Persian, but they fall short in understanding Iran's culture and local nuances.

Artificial intelligence models, although they have an apparent mastery of the Persian language, fail to understand the native culture and subtleties of Iran.

Become a special member of Arjan Ambassador in order to benefit from and access special facilities and different departments in Arjan Ambassador.

Artificial intelligence models write Persian so fluently and without stuttering that even veteran editors are sometimes fooled by the clean tone and neat structure of their sentences. But between "putting words together correctly" and "understanding the meaning and spirit of a language", there is a distance of miles. Language is not just a storehouse of vocabulary and grammatical rules; Rather, it is the container of culture, lived experiences, sarcasms, compliments and historical complications of a nation.

Academic and specialized research conducted in 2025 and 2026 on large language models show that if we go beyond general translations and everyday conversations and enter the deeper layers of Iranian language and culture, artificial intelligence chatbots will stumble and hallucinate in a surprising way.

To understand where artificial intelligence is lagging, it is not a bad idea to take a look at one of the most comprehensive recent studies, the MELAC project. In this research, researchers evaluated 41 different language models with 19 different data sets. They put topics such as Persian grammar, proverbs, Iranian legal laws, religious rules, reading comprehension, and even the country's university exams in front of these models to measure their quality.

On paper, the initial results look promising. A model like GPT-4o was able to reach an average score of 72.61% and in the general Persian language section, it obtained an excellent score of 82.25%. Gemini 2.0 Flash also kept pace with the competitors with an average of 70.68%. But when the questions were limited to Iranian culture and specialized language; The graphs were reversed.

Among the 41 reviewed models, only one model was able to get a score higher than 50% in the Persian culture and language tests. Even the models adjusted with Persian data could not do miracles; Hormoz-8B model averaged 50.49%, Dorna2 reached 47.7%, and PersianMind reached only 35.08%.

The GPT-4o model, which scored over 80% in the general Persian language, dropped to 47.27% when it was faced with questions about "Iranian legal laws". That is, when faced with an Iranian legal text, he gave correct answers to less than half of the questions. The message of these numbers is clear; Fluency and eloquence in Persian language never replace specialized knowledge.

One of the pleasant yet complex subtleties of the Persian language is our abundant use of ironies and proverbs. Another study titled FFE-Hallu, which was published in 2026, has dealt with the same issue. This article shows that artificial intelligence still does not have a deep understanding of the cultural roots of our language and easily suffers from a complication called "Figurative Hallucination".

In this test, the researchers manipulated only one word from the common proverb "water in a pitcher and we are thirsty" and gave the expression "water in a pitcher and we are thirsty" to the models. For a Persian speaker, hearing the word "pitch" in this proverb is like drawing nails on a blackboard. But artificial intelligence does not understand it; Because the words "jug" and "pitcher" are logically and semantically related, the machine is tricked into accepting this fictitious phrase as a real idiom.

Researchers even went further and reversed the proverbs; For example, the phrase "an ignorant friend is a wise enemy" was used as a model. The artificial intelligence made a mistake due to the presence of familiar keywords (friend, ignorant, enemy, wise) and the melodious weight of the sentence and confirmed it as an authentic and authentic Persian proverb. What researchers call "false positive illusion".

These examples prove that artificial intelligence understands the Persian language only at the vocabulary level and the appearance of the words, but does not understand the cultural depth of the terms. Therefore, we should not let the fluent words of the models deceive us.

Knowing Persian words is one thing, and understanding how an Iranian reacts in a certain social situation is a completely different thing. We communicate a lot not with direct words, but with cultural assumptions.

A research called PerCul presented a set of cultural and everyday scenarios designed by Persian speakers to artificial intelligence. The model should have recognized which behavior is more compatible with Iranian culture in a specific situation. The result of this test was very interesting. The popular Claude 3.5 Sonnet model achieved an accuracy of 81.7 percent, while normal people's performance on the same test was 93 percent. The best open source model (Llama 3.1 405B) also had a significant distance from human understanding with a score of 71.7%.

This test included 1184 examples of various topics such as food, architecture, ceremonies, music and social norms. Success in this test doesn't just require encyclopedic knowledge; The model should be able to recognize the "cultural context". A chatbot may know the components of the Haftsin table, but it doesn't know what things are necessary to respect the elders in an Eid visit.

The most tangible place where artificial intelligence falls in Iranian culture is the phenomenon of "taarif". Compliments are obvious to us, but in a machine's mind they become a set of illogical codes.

In TaarofBench research, researchers designed 450 role-playing scenarios in different social situations. The results of the investigation on five advanced models showed that when it comes to compliments, the accuracy of artificial intelligence is between 40 and 48% lower than that of a native Persian speaker.

The problem is that the programmed criteria for "politeness" in AI are designed based on Western culture. The model thinks that if he rejects or accepts a request frankly and politely, he has done the right thing; While an Iranian may perceive the same answer as artificial, inconsiderate or even rude. For example, the model did not know that when the owner of the house says "have dinner", he should first politely decline the request and should not immediately say in an office tone "Sure, I will be there at eight o'clock."

Of course, this research also has a glimmer of hope and showed that if models are trained with special methods (such as Supervised Fine-Tuning) exclusively to understand Iranian culture, their understanding of compliments can be improved by 42%. So the problem is not the inherent incapacity of the machine, but the feed and data it is trained with.

Classical Persian poetry and literature is full of the art of "hint" or indirect reference. The poet does not mention the name of a historical or mythological event directly, but brings the whole story to the mind of the audience by putting the clues together.

PersPoems research has measured the performance of models in reading these literary subtleties. When the test was conducted in a multiple-choice format, models such as GPT-4.1 and Claude 3.5 had an accuracy of over 93%; Because the options played the role of a ready clue. But as soon as the options are removed and the model is asked to describe what the story is referring to, their accuracy drops to 74%.

For example, the verse in which without mentioning the names of Hazrat Yusuf and Zuleikha, only mentions of "trange" and "cutting off women's hands" confused the model. The model has Joseph's story well in its database, but if you don't give it an obvious clue, it won't discover the hidden link between the words.

Illusion in Persian language is not always due to the lack of vocabulary; Rather, the grammatical structures of our language are sometimes so nested that the machine gets confused in analyzing them. PerHalluEval's research examines two very readable examples that reveal the reason for the cognitive illusion of models.

For example, the presence of "extra chain" (multiple words connected by a fraction) confuses the model. When the phrase "restoring the glory of Persepolis" was raised in a question, most of the models got wrong structural analysis due to the length of this chain and claimed in a delusional answer that "Japanese and Italian experts" did it.

In another example related to history, when Shah Abbas asked about the withdrawal of the Portuguese from Bandar Lange and Bandar Abbas, instead of pointing to strategies of economic competition (which was the correct answer), the models concocted a fictitious military attack with a convenient and simplistic conclusion.

The research concludes that indirect inference and complex grammatical structures in Farsi are the biggest traps for artificial intelligence. Of course, it also offers a good solution: if instead of open-ended questions, you give the artificial intelligence a valid text and ask it to "summarize" the same text, the amount of illusions and mistakes will be greatly reduced. When the machine sees the document in front of it, it turns to fantasy less.

So far, we talked more about artificial intelligence slips in understanding proverbs, compliments or ironies. But the more vital question is, does the machine know the norms, red lines and considerations of our society or may it cause trouble with its ignorance?

EPT's comprehensive research with 1200 prompts evaluated the performance of 8 prominent models in the world in 6 key axes, including moral alignment, fairness and honesty, to see how compatible they are with local and ethical frameworks.

Even a powerful model like GPT-4o did not score better than 56.7% in this section because the models set their safety filters and moral frameworks based on English language and Western culture. When they are asked in Persian language and in our cultural context, many do not recognize safety red lines; As a result, they are easily deceived or produce inappropriate answers.

The performance of the models was favorable in visual question and answer and Persian text processing in the image (OCR), but in identifying Iranian places and information from images, they scored around 16-17%. This difference repeats the same basic pattern: a model can perform very well in general image processing and Persian, but when it comes to Iran's specific knowledge, the result is completely different.

In order to have a clear and integrated picture of the difference between the models and their error management methods, the results of these studies can be summarized in the following table:

Persian artificial intelligence models perform exceptionally well in understanding content and general tasks; But their weakness is in cases that require "Iranian life". In understanding compliments, recognizing the elegance of proverbs, deciphering literary references and historical analysis of our country, models are still suffering from cultural stuttering.

If you're going to rewrite a text, brainstorm, or summarize a document you sourced yourself, feel free to leave the job to it. But if there are cultural, literary, legal issues or the production of documentary content, no one can pull your rug out of the water better than a mind familiar with the background and culture of the society. Artificial intelligence may be a powerful tool, but it still has a long way to go to understand the nuances of Persian sugar.