In this study, we focus on automation of citizen-generated complaint identification. In recent years, researchers have used several combinations of machine learning methods to solve the problem of studying opinions. Recent advancements in AI for automating social network complaints converge on sophisticated transformer architectures, multi-modal frameworks, and federated privacy-preserving models. However, challenges remain in adversarial robustness, low-resource language adaptation, and real-time-human collaboration. The integration of explainable and ethical AI techniques is a promising frontier for trustworthy and scalable implementations. We took 75 public pages on VKontakte social network belonging to the governors of Russian regions, subject to open comments. The collection was carried out using the API and public web scraping tools in December 2024, 19958 messages were collected to the most commented post per month in each of the 75 accounts. Of these, 9500 comments were manually marked as containing or not containing complaints for subsequent training of the model, and another 1,000 comments became a test sample for the algorithm. Our results showed that manual markup resulted in 33% of the complaints for the 9,500 comments in the sample, and applying BERT pre-trained model resulted in 28,4%. This means that our hypothesis has been confirmed, and the model, pre-trained on a small sample in the same language, turns out to be almost equivalent to manual markup.