Only 20 percent of the world’s population speaks at home in English, but in this language almost half of the teaching data for the main models of AI is presented. Follow the tongue: AI and the battle for language diversity Fabrice Robin Culture and education over the past two years, acting under the auspices of the UN, the international organization of the Francophony is conducting a tireless campaign in the corridors of global digital diplomacy. Its main goal is to bring linguistic diversity to the sphere of artificial intelligence, where English is strongly dominated by. Having gone on the scene of the summit in the field of artificial intelligence in Paris, the executive director of & nbsp; Google Sundar Pichai from the rostrum of the Grand Palace announced the beginning of the “new golden age of innovation.” “using artificial intelligence methods, we last year, last year we are Added to & nbsp; Google & nbsp; Translate more than 110 new languages, which say half a billion people around the world, ”he said. – Thus, now this program supports a total of 249 languages, including 60 African ones, and this is only the beginning. ” ~ ~ > this statement did not make a large splash among the participants of the summit – world leaders, researchers, representatives of non -governmental organizations and managers of technological companies, & nbsp; but for Supporters of linguistic diversity in artificial intelligence, the words of Sundar Pichai marked the victory achieved as a result of two years of intense backstage negotiations. 62 ~~ 60 > 62 >“This shows that our efforts are not in vain, and technological companies listen to our ideas,” said Joseph Nalvo Ngula, adviser to the digital policy of the international organization of Francophony. 62 ~ linguistic gap A few years ago, generative AI-a branch of artificial intelligence, capable of creating original content, from text to images, music and animation-made many mistakes in his work. & Amp; nbsp; when the Openai developer launched & NBSP; Chatgpt, people who did not speak English quickly discovered the boundaries of this Programs. & nbsp; request in English led to a detailed, informative answer. If the same was asked, say, in French, & nbsp; Chatgpt issued no more than two paragraphs, followed by shy apologies: “Sorry, I am not trained” or “My model is updated.” 60 > 62 > 62 > 62 > 62 > 62 > 62 > 62 > 62 > 62 > 62 > 62 > 62Such a gap was caused by a complex structure of artificial intelligence tools, which rely on the so-called large language models (LLM), such as GPT-4, & nbsp; Llama from & NBSP; Meta or & AMP; NBSP; Gemini from & gogle. They process huge massifs of Internet data, allowing you to understand and generate the text. 60 > modern Internet space mostly English-speaking. Only 20 percent of the world’s population speaks at home in English, but almost half of the teaching data for the main models of AI is presented in this language. & Amp; NBSP; Today, ChatGPT answers in French, Portuguese or Spanish are more convincing than before, but still remain less informative than than English. 60 ~ h2 > “English dominance” ~ 60 > “volume of available information in English is much wider than other languages, and this information is more relevant,” says Nalvo Ngula. AI models are developed, trained and deployed in English by default, as a result of which other languages are forced to keep up with them. This is not only a quantitative gap. AI, devoid of reliable support in any language, begins to “hallucinate”-without doubt a drop of doubt the incorrect or absurd answers. & Amp; nbsp; So, by & nbsp; request of the UN news service, not presented in English, & chatgpt issued the following response to the question of who Victor Gyugo is: “The French writer & nbsp; the 19th century, which was also a passionate astronaut and made a significant contribution to the early design of the international space station.” black box ~ 60 > “this is a kind of black box that absorbs data,” explains the Nalvo Ngula. – The results can be formally consistent and logically structured, but virtually extremely inaccurate. ” 60 > besides actual errors, AI tends to smooth out linguistic wealth. Chat bots fight with regional accents and linguistic variations, such as Quebec French or Creole languages, which speak Haiti and French Caribbean. & Amp; nbsp; French texts created by artificial intelligence often seems living native speakers of the language emasculated and devoid of stylistic nuances. 60 > “Moliere, Leopold Sedar Sydar, Ema Sezer, Mongo Beti-all of them would have turned up in the coffin if they saw what and how II writes in French,” Nalvo jokes. Ngula. This problem is even more relevant for multilingual countries. For example, in their native Cameroon, youth usually speaks Camprangle – a hybrid of French, English and local languages. & Amp; nbsp; “I doubt that young people will be able to ask artificial intelligence a question on Camprangle and get a giving answer,” says he says He. diplomatic success The international organization of the Francophonia, where NILVO NGULA operates, unites 93 states that use the French to one degree or another, and represents the interests of more than 320 million people around the world. Overcoming the linguistic gap-the central element of its digital strategy. & Amp; nbsp; the culmination of the group’s efforts was the participation of the UN global digital agreement-a framework for managing AI, adopted by member states last year ~ 60 > 2023, the international organization of Francophonia uses its international organization of Francophony The diplomatic network, including the influential group of ambassadors of the French -speaking countries in the UN, so that the language diversity has become a key principle in developing a policy in the field of artificial intelligence. & NBSP; on this way, unexpected allies appeared: Portugal -speaking and Latin -American human rights groups and even official Washington joined the organization. “The United States has been defending the inclusion of new languages in the development of AI,” said Nalvo Ngula. ~ 60 > in the final version of the global digital agreement, cultural and linguistic diversity is directly recognized as a question that initially retreated to the background in discussions about expansion about expansion of expansion accessibility. 60 ~ h2 > problems remain Despite these achievements, problems remain. Franco -speaking content is often ignored by platform algorithms, says Nalvo NGUL. & Amp; nbsp; streaming broadcast giants such as & nbsp; netflix, & nbsp; youtube and & nbsp; spotify, give the priority of the priority The popularity of the product, which means that English -language content dominates the search results. ~ 60 > “If the language diversity was really taken into account, the franco -speaking user would see films in French in the upper part of the list, which is recommended to him,” says he says Ngula. & Amp; nbsp;