Keyword: «adaptation of physics problems»

The article addresses the problem of using generative neural networks (using DeepSeek as an example) for solving school problems in thermodynamics. Based on the analysis of four classical problems on the heat balance equation, stable markers are identified that distinguish an AI-generated solution from a student's work: technical artifacts in LaTeX, transliteration of Russian words, excessive precision of calculations, presence of «verification» blocks, step-by-step numbering of processes, and the use of non-standard constant values. It is shown that problems containing a graph require visual data analysis and therefore fundamentally cannot be solved by a neural network without a textual description, which makes them an effective tool for preventing cheating. Specific methodological techniques for adapting tasks for grades 8 and 10 are proposed, reducing the effectiveness of using AI without changing the physical complexity of the problems.