Whitespaces Don't Lie: Feature-Driven and Embedding-Based Approaches for Detecting Machine-Generated Code

Detection is currently highly effective: Both feature-based and embedding-based approaches achieved over 99% accuracy (ROC-AUC scores of 0.995 and 0.994 respectively) at distinguishing AI-generated code from human-written code, suggesting that reliable detection tools are technically feasible today.

Simple patterns reveal AI authorship: Whitespace and indentation patterns are surprisingly strong indicators of machine-generated code, meaning detection doesn't require complex analysis—this makes it practical to implement in real-world settings like academic institutions and software companies.

Different detection methods have different strengths: Feature-based detectors are more interpretable and easier to understand (you can see exactly why code was flagged), while embedding-based detectors are slightly more precise but work as "black boxes," presenting a trade-off between transparency and accuracy that organizations must consider.

Generalization remains an open question: While the paper demonstrates strong performance on a large benchmark dataset, it doesn't fully address whether these detectors will remain effective as LLMs evolve and potentially adapt to evade detection, suggesting ongoing vigilance will be needed.

Multiple safeguards are now possible: The existence of two complementary detection approaches means organizations can implement layered defenses—using both methods together could provide additional confidence in flagging AI-generated code for academic integrity and responsible AI governance purposes.