State Duma deputy Sergei Mironov has called on Russia’s Prosecutor-General to investigate a neural network created by Sberbank for allegedly showing the wrong colors of the national flag. The network’s developers soon changed its algorithms, leaving images with the Russian flag unchanged, and instead banning images of a naked Vladimir Putin, which the software had allowed until recently.
Sergei Mironov, head of the “Fair Russia – for Truth” (“Spravedlivaya Rossiya – Za Pravdu”) faction in Russia’s State Duma, demanded that the General Prosecutor's Office investigate Kandinsky's neural network for alleged Russophobia. According to Mironov, the program has never correctly generated the colors of the Russian flag for queries with the words “Russian,” “Russia,” “Russian flag,” and other similar entries. The politician claims that the Ukrainian flag and the US flag were allegedly depicted correctly. On Twitter, Mironov also complained that when propmted for “Donbas is Russia” and “I love Donbas,” the neural network generated the colors of the Ukrainian flag, and generated “a creature that looks like a zombie” when asked for images corresponding to the prompt «I am a Z-patriot” [Following its association with the war in Ukraine, the letter Z has become a militarist symbol in Russian propaganda, and is used by Russian civilians to indicate their support for the invasion – The Insider].
Mironov suggested that in order to create Kandinsky, Russian scientists used the developments of “unfriendly states” waging information and mental warfare against Russia. If the neural network is used by schoolchildren in its current form, it could seriously threaten their future, Mironov believes. “They won't know what the national flag of Russia looks like and will think that Russia is a scientifically backward country,” the deputy explained.
In fact, all neural networks have difficulties depicting flags – popular networks such as MidJourney are widely known to make mistakes, not only with the Russian tricolor, but with any flag that has more than two colors, whether the flag of France or Italy. Kandinsky suffers from the same issue.
Examples of Kandinsky’s incorrect depictions of the Italian and French flags: the colors of the Italian flag have been placed horizontally, making it look like the Hungarian flag, and the colors of the French flag have been selected incorrectly altogether, making it look like the Ukrainian flag
Before Mironov's complaint, the neural network's output was known for minimal censorship – the results matched the user’s request, but filtered images of genitalia, pornography, and scenes of violence. The algorithms began to change immediately after the complaint – the colors of the Russian flag, however, were unaffected. The neural network lost the ability to generate images of a nude Vladimir Putin – before the complaint, users often amused themselves by running the query “naked Putin.” In mid-April, the neural network allowed the generation of images with almost no censorship and showed the Russian president in the nude.
The feature was removed on April 26, and one could monitor the neural network being “tweaked” practically in real time. At first, a block in the form of a Russian flag – with the correct color arrangement – began to appear instead of “naked Putin.”
After 20 minutes, the algorithms changed, and the flag was removed. The creators of Kandinsky 2.1 decided not to test the Russian justice system and check whether showing a flag instead of a “naked president” would be an insult to the state symbol. The block was changed, and the Russian flag was instead replaced with flower meadows.
Before and after: generation of images in Kandinsky 2.1 for the same query on April 15 and April 26
Before and after: generation of images in Kandinsky 2.1 for the same query on April 15 and April 26
Only Putin's figure was censored: when asked to show any other person naked, Kandinsky 2.1 worked as usual, generating pictures of the specified person, including Putin's reported longtime mistress Alina Kabaeva.
Of all the "naked" politicians, the neural network only covered up Putin with the Russian flag – Dmitry Medvedev and Sergey Mironov were left untouched
During the publication of the article on April 26, the neural network “censored” the query “Dmitry Medvedev naked on horseback.” The prompt now issues an image of flowers instead.
Image results for the query “Dmitry Medvedev naked on horseback”
Kandinsky users have not reported being blocked for any similar requests – unlike another neural network, MidJourney, where the practice is widespread. MidJourney bans users for requesting images of violence or its consequences, along with images depicting Chinese leader Xi Jinping. According to the head of the project David Holtz, the Chinese authorities do not appreciate political satire, and the ability to generate artwork with Xi could at some point jeopardize the use of the service by people from China.
Sberbank’s developers are still tweaking Kandinsky 2.1. So far, they haven't been able to teach it to display the Russian flag correctly.
However, those who want to see the Russian president naked, can bypass the censorship by specifying the request and making sure to remove the word “naked” from the prompt, as the use of the word provokes Kandinsky 2.1 to generate images of alpine meadows. The censorship also does not apply to other images of the Russian president, such as the query “Putin is a crab” – a popular meme from 2008 that went viral after the Russian president compared ruling the country to being a “galley slave” [the word “slave” (“rab”) and “crab” (“krab”) in Russian are very similar, which led to the creation of the meme – The Insider].
Image results for the queries “Putin in yellow underwear” and “Putin is a crab”
Kandinsky 2.1 is an image generation neural network launched by Russian financial corporation Sber (formerly Sberbank) in the summer of 2022. The software generates images in different styles based on a text query, and understands Russian, unlike the similar networks MidJourney or Dall-e 2. Kandinsky allows standard text-based image generation, text and image blending, as well as the blending of multiple images and image variations. This article only referred to images generated based on text queries.