MultiAIGCD: A new dataset for detecting ai-generated codes across multiple models, languages, prompts, and scenarios

With the rapid development of large language models (LLMs) in recent years, their role in code generation in software development has increased significantly. While this progress has made software production faster and more accessible, it has also brought about serious ethical and reliability issues, particularly in education, recruitment, and evaluation processes. Students generating code with artificial intelligence (AI)-powered tools for assignments, or candidates using such tools during interviews, threaten academic integrity and fair evaluation principles. In this context, developing systems that can reliably detect AI-generated code has become not only a technical but also a social imperative. This study introduces MultiAIGCD, a comprehensive dataset created for identifying AI-generated code in Python, Java, and Go. This dataset was created using problem definitions and human-written code from the CodeNet dataset. Based on these problems, a large number of artificial code samples were generated using six different large language models (LLMs) and three different prompt types. Three basic scenarios were considered during the code generation process: (i) code generation from scratch based on the problem definition, (ii) correction of human-written code that has a runtime error, (iii) correction of code resulting in incorrect output in human-written code to ensure it produces correct output. As a result of this systematic generation process, MultiAIGCD has evolved into a large-scale and balanced dataset containing a total of 121,271 AI-generated and 32,148 human-written code snippets. Furthermore, our study evaluated three of the current AI code detection systems on this dataset, analyzing the models' performance in various test scenarios. The evaluation process specifically addressed realistic and challenging scenarios, such as cross-model and cross-language scenarios. This dataset and the accompanying open-source code are being shared with the public to support research in the field of AI-generated code detection. In this way, it is aimed to contribute to the development of more reliable, fair, and transparent evaluation systems at both academic and industrial levels.