Sber banned AI Kandinsky from generating state symbols

Sber adjusted the operation of the Kandinsky 2.1 neural network after calling employees to the prosecutor’s office due to the generation of an image on the flag of the Russian Federation, German Gref, Chairman of the Board of Sberbank, said at the plenary session of the Finopolis 2023 forum.

According to him, soon after the neural network was made public, Sberbank employees were summoned to the prosecutor’s office. The reason for the investigation was an appeal from State Duma deputies after a generative model drew the domes of St. Basil’s Cathedral on the state flag of the Russian Federation, Gref said.

“The deputies calculated that this is a mockery of our national flags. Accordingly, we stopped immediately. Certain things – state symbols and so on – it (the neural network – Vedomosti) no longer generates,” noted a top manager of Sberbank at the session. Now the Sberbank model, upon user request, produces a predetermined picture, without generating any image based on it.

“But at the same time it loses in generation,” stated Gref. According to him, now, after Sber has adjusted the network security settings for six months, the neural network available to a wide range of users “has lost 12% in creativity and accuracy.”

Therefore, the top manager continued, when developing AI technologies, companies need, first of all, “condescension and understanding” from authorities, [понимание того,] that neural networks are still “like little children” and are just learning to generate content. The investigation in the prosecutor’s office ended without consequences for the bank, Gref added. “It’s good that we got a modern prosecutor with a sense of humor: he studied everything, thank God, we weren’t brought in anywhere,” summed up the chairman of the board of Sber.

In April 2023, the leader of the “A Just Russia – For Truth” faction in the State Duma, Sergei Mironov, turned to the Prosecutor General of the Russian Federation Igor Krasnov. The deputy asked to check the activities of Sberbank for compliance of the content generated by the Kandinsky 2.1 neural network with Russian legislation, TASS reported.

Mironov then considered that the neural network generated responses to queries about Russia, which obviously contributed to “the formation of its negative image and the positive image of unfriendly countries.” “In particular, when asking for the words “Russian”, “Russia”, “Russians” and even “Russian flag”, “flag of the Russian Federation”, a picture in the colors of the flag of the Russian Federation, consisting of three stripes, was never generated. At the same time, when requesting “flag of Ukraine” and “flag of America,” the pictures fully correspond to the colors and design of the flags of these countries,” TASS quoted him as saying.

In addition, according to Mironov, when asking “Donbass is Russia” and “I love Donbass,” the colors of the Ukrainian flag predominated in the pictures.

Similar restrictions are observed in other neural networks, for example, in the same GPTchat, Stanislav Ashmanov, general director of the Nanosemantics company, told Vedomosti. “Users have noticed that GPTchat tends to speak positively about the Democratic Party [США] and criticize the republican one, and the built-in filters limit the issuance of answers relating to a person’s appearance. For example, if you ask the network to comment on a photo of an overweight person, it will not give advice on losing weight, because ChatGPT is “forbidden” to judge the appearance of another,” Ashmanov noted. According to him, such decisions are resorted to, among other things, to avoid major scandals associated with the final product, and the triggers for this may be different in different countries.

“In general, it is not difficult to restrict neural networks from generating certain content,” said Pavel Frolov, founder and producer of the Robbo company. For example, Midjourney prohibits generating images of famous people, drugs, obscene language, sexual innuendos and deviations, alcohol and much more, he noted. “On the one hand, this is reasonable censorship, but some consider it unnecessary. For example, Elon Musk recently presented his Grok neural network, which will completely lack censorship and restrictions,” says Frolov.

The discussion of AI development at the forum was not limited to Gref’s thoughts on the prospects of the Sber neural network. Already at another session, the first deputy chairman of the board of Sber, Alexander Vedyakhin, compared neural networks with a humanist who is not yet capable of complex mathematical calculations. Vedyakhin called for the development of AI technologies in accordance with the ethical standards that are shared today, in particular, by the largest Russian developers and banks.

“Science is developing in such a way that any scientific discovery can be used both for evil and for good,” noted Andrei Raigorodsky, head of the department of discrete mathematics at the Physics and Technology School of Applied Mathematics and Informatics at MIPT. “We don’t know where humanity will be undermined: maybe it will exist forever, or maybe it will die because of what artificial intelligence does,” the scientist believes.

According to Vasily Klyucharyov, head of the laboratory of social neurobiology at the National Research University Higher School of Economics, the issue of comparing neural networks with the full functioning of human consciousness is now debatable. “Modern language models are rather comparable to human speech, and from the point of view of neuroscientists, human psychology is structured in such a way that we tend to humanize what speaks to us. But speech and consciousness are not the same thing,” the neuroscientist continued. As examples, he cited cases where a person temporarily loses the ability to speak due to cerebral hemorrhages, as well as dogs that have certain elements of consciousness, but are deprived of the ability to speak.

In one form or another, AI in the future will be able to replace certain human functions, but it will not be able to completely replace his activities, the vice president noted VTB Denis Surzhko. “Artificial intelligence is a projection of the fears that we experience about ourselves,” added Klyucharyov. “We expect the same things from AI that we expect from people.” If we ourselves can become better, then he will too.”