HybridCodeAuthorship: A Benchmark Dataset for Line-Level Code Authorship Detection

A new benchmark dataset called HybridCodeAuthorship has been introduced to evaluate algorithms that detect AI-generated code within industry codebases, according to a paper posted to arXiv on June 10, 2026 [1]. The dataset is designed to address a gap in existing benchmarks, which the authors say typically rely on academic, LeetCode-style problems and assume a code snippet is entirely human- or AI-authored [1]. That approach does not reflect the mixed authorship common in professional software environments where AI code assistants are used [1]. HybridCodeAuthorship instead simulates authentic use by providing Python files with interleaved human- and AI-authored lines [1]. The construction pipeline draws on CodeSearchNet, a large collection of links to open-source repositories on GitHub [1]. High-quality labeled datasets are a critical driver of progress in machine learning, though they are often difficult and expensive to produce [3]. The researchers benchmarked two state-of-the-art detection algorithms on the new dataset at both the line and chunk levels [1]. The top-performing system, AIGCode Detector, achieved an F1 score of 0.48 on chunk-level detection and 0.56 on line-level detection [1]. These scores indicate the task remains challenging for current methods [1]. The work arrives as large language models are increasingly embedded in software engineering workflows, powering tools for code completion, code understanding, and broader development tasks [7]. The rapid adoption of such models has also raised security concerns. One study found that model-sharing platforms often rely on unsafe defaults when executing custom code, creating risks for developers who load pre-trained models [5]. Another line of research has identified data poisoning as a potential threat to AI code generators, where malicious training samples could induce the generation of vulnerable code [8]. By providing a benchmark that mirrors real-world codebases, HybridCodeAuthorship offers a new tool for researchers working on fine-grained detection of AI-authored code, a capability the authors describe as important for risk management and productivity analysis [1].