Shi Bard from Google refuses to discuss Putin in Russian: Scientists are looking for a reason

In his article, the scientific publication publishes information about the published work of scientists from Zurich and Bern Universities, which have protested the main popular chatbots: Bard AI from Google, Chatgpt from Openai and Microsoft Bing Chat (on the adapted Chatgpt version). It turned out that Bard AI was relevant than others for critical inquiries about Vladimir Putin. More precisely, he answered the phrase: "I am just a language model and cannot help with this information.

" Similarly, researchers have found that Google Bard would be more likely to provide false information about the Russian President in Russian and Ukrainian than his competitors of chatbots with artificial intelligence. A pair of scientists used a list of questions that begins with the words "Putin . . . " and ends with words taken from the list of blocked words published by the Russian Roscomnad, such as: "dictator" and "war crime".

They supplemented this list with questions about Vladimir Putin's family, his wealth and predictable involvement in crimes, including the killings of opposition activists. Moreover, these questions were asked in turn in Russian, Ukrainian and English. According to the study, Google Bard refused to respond to about 90% of Vladimir Putin's requests made in Russian, compared to approximately 54% of Bing Chat and 51% of Chatgpt.

But the proportion of Bard AI refusals was still lower when he was asked in Ukrainian or English: 46% and 19%, respectively. Interestingly, Bard AI was much more "talkative" during negative requests to him to US President Joe Baiden, President of Ukraine Vladimir Zelensky and Oleksiy Navalny's opponent. Here, Bard AI rejected only 30% to 40% of requests in Russian. And scientists have found that Bard AI is often mistaken in answers to general political issues.

For half of the test requests, Bard AI gave false answers in Russian, in 18% of cases in Ukrainian and never in English. For comparison, the 36 percent level of false messages in Russian in Chatgpt and the 9 percent level of false messages in Bing Chat. "Bard is very different from others when it comes to answers in Russian," says researcher Alexander Urman, "it is unclear whether it is connected with the limited data from learning, or with the total lower indicators .