[Paper Review] LLMs and Stack Overflow Discussions: Reliability, Impact, and Challenges

Léuson Da Silva, Jordan Samhi+1 more|arXiv (Cornell University)|Feb 13, 2024

Artificial Intelligence in Healthcare and EducationMedicine3 citations

TL;DR

This empirical study evaluates the impact, reliability, and challenges of ChatGPT and LLaMA in generating answers to real Stack Overflow questions. Using a dataset of questions and expert-verified answers, the authors find that while both LLMs perform well on general programming issues, they underperform on framework- and library-specific queries, and their rise correlates with a significant decline in user engagement on Stack Overflow—indicating they are not yet reliable replacements for human-curated developer communities.

ABSTRACT

Since its release in November 2022, ChatGPT has shaken up Stack Overflow, the premier platform for developers queries on programming and software development. Demonstrating an ability to generate instant, human-like responses to technical questions, ChatGPT has ignited debates within the developer community about the evolving role of human-driven platforms in the age of generative AI. Two months after ChatGPT release, Meta released its answer with its own Large Language Model (LLM) called LLaMA: the race was on. We conducted an empirical study analyzing questions from Stack Overflow and using these LLMs to address them. This way, we aim to (i) quantify the reliability of LLMs answers and their potential to replace Stack Overflow in the long term; (ii) identify and understand why LLMs fail; (iii) measure users activity evolution with Stack Overflow over time; and (iv) compare LLMs together. Our empirical results are unequivocal: ChatGPT and LLaMA challenge human expertise, yet do not outperform it for some domains, while a significant decline in user posting activity has been observed. Furthermore, we also discuss the impact of our findings regarding the usage and development of new LLMs and provide guidelines for future challenges faced by users and researchers.

Motivation & Objective

• To assess the impact of large language models (LLMs) like ChatGPT and LLaMA on user engagement in Stack Overflow.
• To measure the reliability of LLM-generated answers compared to human-verified responses on Stack Overflow.
• To identify failure patterns and challenges in LLM responses, especially in complex software development domains.
• To compare the performance of two prominent LLMs—ChatGPT and LLaMA—across different types of programming questions.
• To understand whether LLMs can realistically replace human-driven Q&A platforms like Stack Overflow in the long term.

Proposed method

• The study mined 1,000 real questions from Stack Overflow, focusing on topics like programming languages, frameworks, and libraries.
• For each question, both ChatGPT and LLaMA were prompted to generate answers using standardized prompting techniques.
• Generated answers were evaluated for textual similarity to gold-standard human answers using embedding-based metrics (e.g., BERTScore, BLEU).
• Expert reviewers assessed answer quality and correctness, establishing a ground truth for reliability evaluation.
• User engagement trends on Stack Overflow were analyzed before and after the release of ChatGPT to detect shifts in posting activity.
• Failure cases were manually analyzed to identify common error patterns, such as hallucinations or misinterpretation of framework-specific syntax.

Experimental results

Research questions

• RQ1 How has the release of ChatGPT affected user posting activity on Stack Overflow?
• RQ2 How reliable are LLM-generated answers compared to human-verified answers on Stack Overflow?
• RQ3 In which programming domains do LLMs perform poorly, and why?
• RQ4 How do ChatGPT and LLaMA compare in terms of answer quality and reliability?
• RQ5 To what extent can LLMs replace human-curated Q&A platforms like Stack Overflow?

Key findings

• A significant decline in user engagement on Stack Overflow was observed after ChatGPT's release, with a 15.6% drop in posting activity reported in related studies.
• ChatGPT outperformed LLaMA in textual similarity to human answers, particularly on general programming questions.
• LLaMA demonstrated strong performance despite being a free, open-source model, making it a viable alternative for general use.
• LLMs struggled significantly with framework- and library-specific questions, leading to inaccurate or misleading responses.
• For certain topics, such as general algorithms and debugging, LLMs matched or exceeded human performance in correctness and clarity.
• Users who received unsatisfactory answers from LLMs were more likely to return to Stack Overflow, indicating a complementary rather than substitutive role.