With the increasing availability and capability of generative AI, a new challenge to academic integrity emerges in computer science education, as concerns grow about its potential misuse. Students may try to solve programming assignments by generating solutions based on the assignment description and even adjusting the temperature value in order to alter properties of the output to their advantage. In this paper, we evaluate the effectiveness of Large Language Models (LLMs) at generating valid, inconspicuous code for university-level programming assignments. We analyze how the temperature value affects the correctness of the generated code and its ability to evade detection by plagiarism detection systems. To that end, we systematically generate about 10,000 Java programs using the LLMs GPT-4.1, GPT-4o, and DeepSeek-V3 across four real-world assignments under varying temperature settings. Our results show that although most programs compile successfully, they frequently fail to meet the functional requirements. Moreover, we observe that generated programs fail to evade typical plagiarism detection systems. We observe that only 0.95% of all programs are both inconspicuous as well as syntactically and semantically correct, indicating limited effectiveness of cheating with generative AI.