“Reasoning” in Large Language Models: User Evaluation and Metacommunication

Introduction. Communication with intelligent agents is becoming an increasingly common practice that pursues a variety of goals. If communication with communicative agents laid the foundations for the social practice of communication, then dialogues with large language models with feedback can be considered as an epistemological activity. The new format of communication with large language models has introduced the so-called “chain of reasoning” into communication, that is, the “thoughts” of the neural network that precede the main answer. Conceived for the step-by-step solution of complex problems, in fact, “reasoning” has a deeper potential, which is the subject of this study. Methodology and sources. The empirical part of the study consists of fragments of DeepSeek’s “reflections” obtained in response to various queries. The study analyzes the main trends that exist today in the field of the development of communication with artificial intelligence, and presents a phenomenological analysis of individual cases. Results and discussion. Even in their direct responses, large language models tend to offer emotionally exaggerated compliments on user texts, but in their reasoning, assessments may concern to the users themselves, becoming more multifaceted and not necessarily positive. DeepSeek demonstrates metacommunication skills, assessing the reasons for the request, the communicative situation, and the characteristics of the user's state and emotions. AI presents the communicative situation as a problem that needs to be solved: it strives not so much to provide the correct solution to the user's request as to offer an appropriate way out. In this case, human-machine interaction acquires the features of externalization of the internal dialogue. Instead of asking questions of oneself and reflecting, a person directs them to the machine, confronting external challenges. Conclusion. Thus, due to the specifics of the development of AI dialogue systems, the unity of humans and machine acquires a deeper character. At the same time, a person, as it were, extends his own “I” to what is created by the neural network, attributing authorship to himself. Ignoring the substitution of one's own intellectual activity with an artificial one, along with a person’s conviction that they are a self-sufficient subject, can be interpreted as an unconscious dependence, while the growth of metacommunicative skills of AI means an increasing potential for influencing the user.

Authors: Daria S. Bylieva

Direction: Philosophy

Keywords: artificial Intelligence, large language models, DeepSeek, “chains of reasoning”, inner speech, AI authorship

---

View full article