- A quarter of developers ask ChatGPT, not community.
- Using LLMs lowers amount of training data for future generations.
- Human-generated responses in the commons vital for progress.
Researchers at the Cornivus University of Budapest have found that activity on Stack Overflow decreased by 25% in the six months after the release of ChatGPT. Stack Overflow is a popular forum for software developers, where the development community share ideas, ask questions, and pool knowledge.
Compared to its Russian and Chinese counterparts (countries with limited access to ChatGPT), engagement by humans on Stack Overflow fell by a quarter during that period. The researchers surmise that developers who might previously posed questions on Stack Exchange were instead using the large language model to seek answers.
The decline in human interactions was most notable in areas of Stack Overflow concerned with the most popular programming languages. We can surmise that information on those languages is freely-available in greater amounts, and therefore represents a larger body of knowledge available for ingestion by an LLM, so improving its best guess lexigraphically-based results.
The paradox of re-use
The paper refers to the “paradox of re-use,” in which use of a reference resource such as Wikipedia diminishes over time, as its information is presented to users by (in the case of Wikipedia) search engine results pages, which are populated by data scraped from the same source by the search engine. The researchers noted that the trend they describe is just such a paradox.
Additionally, the team found that the quality of posts on Stack Overflow has not diminished, evidenced by a continuation in the trend of up- and down-votes to users’ posts, and that the decline in human input to reference platforms like Stack Overflow gives early-arrival LLMs (such as ChatGPT) a market advantage: later competitors have less human-generated training data for their models, thanks to users with queries referring privately to LLMs, answers to which do not automatically produce new publicly-available knowledge.
The paper’s introduction states, “The transformation of the online commons into a private database presents a novel risk to consumer welfare. More broadly, a shift from open data to a more closed web will likely have significant second-order impacts on the ever-growing digital economy, and how we access, share, and evaluate information.”
The findings of the 2023 Stack Overflow Developer Survey showed that developers who used ChatGPT in their workflows were less likely to post on Stack Overflow, and the Cornivus University’s research seems to be in accordance with that.
There is some potential for knowledge stagnation in more popular programming languages as large language models have less material, over time, with which to finesse their responses. Newer or less-popular languages are less affected by the withdrawal of up to a quarter of contributors to Stack Overflow and other reference sites.
The price of convenience
Project maintainers and their user communities rely on feedback from all levels of users to create environments in which comment and feedback are (almost always) welcome and form an integral part of their communities. On a practical level, such environments help improve any piece of software through sharing of best practice, methods, foibles, workarounds and bugs.
By moving a collectively-created body of knowledge onto a proprietary database behind a paywall is to the detriment of technology and will at least slow the rate of progress in improving developers’ competancy and the quality of code they write.
The apparent convenience offered by LLMs to those who create software is one side of a two-sided sword. Aside from licensing issues around code that has been ingested by LLMs (a legally grey area many open-source and free software advocates choose to ignore), the research shows that the erosion of the foundation on which the modern world continues to be built is, for up to 25% of the community, a price worth paying for a quick, often inaccurate, answer.
The trend identified by researchers points to a situation that can surely only worsen, as scrape-able content produced by LLMs is ingested as learning data by LLMs. The two-headed snake eating itself produces an increasing number of hallucinations (AKA errors) in AIs, lowering the mean quality of knowledge available on the internet, whether its source is human or machine.
