Stack Overflow has introduced OverflowAI, a service under development which merges AI with its huge database of technical questions and answers.
The company will also provide a plugin for Visual Studio Code (VS Code) for users of its Teams product, a private version of Stack Overflow, to enable AI-driven chat from within a code editor. A similar chatbot will be provided for Slack, called StackPlusOne.
OverflowAI search will be part of the Stack Overflow public platform. Currently a search on Stack Overflow provides results in the form of question and answer threads, with votes showing which answers are most helpful according to the community. Questioners can also mark an answer as accepted. In the new form, OverflowAI will provide a summary answer using generative AI, showing the question threads which inform the answer as clickable citations underneath. If the AI fails to answer the question, it will compose the draft of a new question for posting.
The VS Code plugin works in a similar way, but within the IDE, and adding private sources including GitHub repositories, Confluence and Google Drive, for organization-specific answers, as well as drawing on the public Stack Overflow data. Answers may include code samples, but from what we have seen OverflowAI does not attempt to write code in the manner of other plugins such as GitHub Copilot, AWS Codewhisperer, or the new JetBrains AI Coding Assistant.
Stack Overflow has also introduced a new topic area called GenAI which is for questions about generative AI in general.
According to CEO Prashanth Chandrasekar, the goal of OverflowAI is “to make it possible for public platform users to receive instant, trust-worthy, and accurate solutions to problems using conversational search powered by GenAI.” The risk of inaccurate or invented information is a known weakness of generative AI, and the idea is that by linking this to attributed answers that risk will be much reduced.
In addition, the Teams version will serve as a form of enterprise search, though currently limited in the sources it is able to query.
DevClass spoke to Joy Cicman Liuzzo, VP of Product Marketing. Is there a risk that if developers rely on an AI-generated summary, they may miss the nuances that emerge in the kind of detailed critique that often appears in a complete answer thread? “The AI is always going to reach a complexity dip, where it can’t handle all of the nuances, and that’s where the Stack Overflow community is already there and ready to go,” she told us.
Is limiting the OverflowAI sources to just Stack Exchange and private sources going to reduce its value, compared to other coding assistants that draw on more data? “We’re leaning towards accuracy and trusted community validated content,” said Liuzzo. “That is the trade off.”
Nevertheless, OverflowAI could still get things wrong or even hallucinate content – but the limited sources along with citations make that less likely.
Will the public OverlowAI remain free or could there be a subscription? “I can’t comment on that,” said Liuzzo. What about a VS Code plugin for non-Teams users? “Just for teams right now, but we have heard that there’s interest in bringing it out into the public platform,” she told us.
Stack Overflow is suffering reduced traffic, with the likely reason being code assistants integrated into IDEs. How will this solve the problem, which could mean less engagement and therefore less data to draw on, especially if more developers stay within their IDE? “The future of the internet is not just measured by web traffic,” Liuzzo said. “It’s about the quality of the data, the trust in the data, and the community of experts that are curating that information.”
Liuzzo said the new tools are designed to be complementary to existing options like Copilot, rather than aiming to replace them.
When will we get OverflowAI? “Alpha testing begins in August and then general availability dates will be announced when we have those.” The sign up for the early preview is here.
The subject of StackOverflow and AI is a contentious one, since the unreliability of AI-generated answers and the desire of moderators to use tools to help detect and remove them was at the heart of a moderator strike, though whether or not the new launch has any bearing on or relevance to those concerns was not something Liuzzo was willing to discuss.
We should note though that the company has recently posted the full text of its message to moderators, that sparked the discontent and was originally kept private. The company also stated that it is “committed to the long-term (foreseeable future) survival of the data dumps, the API, and SEDE [Stack Exchange Data Explorer]”, another concern of the moderators, though with the additional note that “We retain the right to place guardrails around them to ensure that companies constructing language models, etc, are charged for access.” – something which may be hard to enforce.
There was also some positive comment in a discussion on the topic on the Stack Exchange site. “My initial reaction to this is joy. Assuming the quality of AI content is good and that it works fine, this actually works well for both the SE communities and the page hits … ideally, the communities will be left alone to create meaningful content. Those who want to engage will engage. The rest will have a quick answer to their problem,” said one optimistic user.