Stack Exchange dabbles with AI answers – despite Stack Overflow banning them in 2022

Stack Exchange dabbles with AI answers – despite Stack Overflow banning them in 2022

Stack Exchange is experimenting with AI-generated answers, despite a ban on using AI-generated content on Stack Overflow, the section of the site aimed at developers, that is still in place.

The Answer Assistant, designated as a “Labs experiment,” uses AI-generated answers but hides them until they are “verified, edited and curated” by the community. 

The introduction of Answer Assistant is cautious in its approach, no doubt because of the sensitivity of this topic within the Stack Overflow community. The experiment will initially be restricted to Stack Exchange sites where the moderators have agreed to participate. Sites are able to restrict questions where AI-generated answers are allowed, according to certain criteria such as the age of the question and whether or not it yet has answers. The AI answer will remain private until approved by users with sufficient reputation, which is a score of 50 or more according to the introductory video. Once verified, it becomes public and will be attributed to an answer bot.

AI answers will need to be human-approved before becoming public

The experiment will only appear on a few Stack Exchange sites, currently Arts and Crafts, Raspberry Pi, and User Experience. The Stack Overflow developer site is not yet included. There are hundreds of Stack Exchange sites on topics ranging from coffee to 3D printing, though the Stack Overflow site is the biggest by a huge margin, with 1,200 questions per day versus 155 per day for the second most popular (Mathematics).

The dominance of Stack Overflow within Stack Exchange makes it unlikely that a major new feature such as this will not end up there eventually, if the experiment makes its way into production.

According to the official post, the goal of the Answer Assistant is to reduce the number of unanswered questions and to increase participation within each site’s community. 

The context for this is that traffic to Stack Overflow is in sharp decline, likely caused by developers using AI assistants directly from their IDE rather than visiting the site. Stack Overflow has already set itself up as a source of data for AI assistants, and uses AI to help users frame questions via the experimental Question Assistant.

The site though has until now resisted AI-generated content. A ban was put in place in 2022 with the help document explaining that there are issues with AI content that make it “unsuitable for use on Stack Overflow.” Users expect human-authored answers, the document states, and cannot reliably cite sources. The document also makes the point that getting an AI answer is easy elsewhere, and therefore “if a user wanted an answer from an artificial intelligence, they may already have sought one, and so it does not make sense to provide one here.”

Another question is the extent to which AI answers might be driven by data drawn from the Stack Exchange sites themselves.

Using AI trained on AI-generated data is known to be risky. A paper on the subject from Rice University in Houston, Texas concluded that “without enough fresh real data in each generation of an autophagous [self-consuming] loop, future generative models are doomed to have their quality (precision) or diversity (recall) progressively decrease.” 

As an experiment the answer assistant may never proceed into widespread use; yet it may seem a tempting strategy for the site to pursue in response to declining traffic.

The Stack Exchange team posted an introduction to the Answer Assistant in the Meta Stack Exchange community forum, with a largely negative response so far.  This “seems to be a deviation from the SE mission,” said one; while another suggested that “StackGPT is a better name.”