Richter, CedricHaltermann, JanJakobs, Marie-ChristinePauck, FelixSchott, StefanWehrheim, HeikeEngels, GregorHebig, ReginaTichy, Matthias2023-01-182023-01-182023978-3-88579-726-5https://dl.gi.de/handle/20.500.12116/40105Finding and fixing software bugs is a central part of software development. Developers are therefore often confronted with the task of identifying whether a code snippet contains a bug and where it is located. Recently, data-driven approaches have been employed to automate this process. These so called neural bug detectors are trained on millions of buggy and correct code snippets to learn the task of bug detection. This raises the question how the performance of neural bug detectors and software developers compare. As a first step, we study this question in the context of variable misuse bugs. To this end, we performed a study with over 100 software developers and two state-of-the-art approaches for neural bug detection. Our study shows that software developers are on average slightly better than neural bug detectors – even though the bug detectors are trained specifically for this task. In addition, we identified several bottlenecks in existing neural bug detectors which could be mitigated in the future to improve their bug detection performance.enBug detectionvariable misuse bugsempirical studyText/Conference Paper1617-5468