Mozilla's AI Help infects web documentation with GPT hallucinations

Alfonso Maruccia

Posts: 917   +280
Facepalm: Formerly known as Mozilla Developer Network, MDN Web Docs is a documentation repository providing learning resources for web developers. The site strives to be a reliable source of information "for developers by developers," but Mozilla has now decided to feed it to generative AI algorithms – and developers aren't happy in the slightest.

Mozilla recently announced AI Help, a generative AI-based tool designed to be a new "problem-solving companion" for web developers searching for answers on MDN Web Docs. The site hosts documents about CSS, HTML, JavaScript and other web technologies since 2005, and it has essentially become an authoritative resource with contributions coming from volunteers, big corporations like Microsoft and Google, and of course Mozilla docs as well.

Since 2017, the MDN service also hosts all of Samsung's web documentation after the Korean company decided to shut down its own documentation projects. Needless to say, when Mozilla decides to bring some big, unexpected addition to MDN, developers can react in unpredictable and passionate ways.

According to MDN director Hermina Condei, AI Help was conceived to optimize developers search processes, making it "quick and easy" to find the information they needed. AI Help uses OpenAI's API to feed users prompt to ChatGPT, and the generative AI should retrieve the "most pertinent information" from MDN's comprehensive documentation repository.

As everyone and their dog know by now, generative AI shouldn't be considered a reliable source of information of any kind. The algorithm has no intelligence, no awareness of anything, and it just puts words together to try and find the most statistically sound answer to users' textual prompts.

Mozilla was asking for feedback about the AI Help introduction, and developers answered back in droves.

A GitHub issue opened for Yari, which is the platform code powering the MDN service, clearly depicts the sorry state this "AI Help" feature is in right now. "Eevee," the developer who opened the issue, describes the generative AI feature as a "strange decision" for a technical reference where the human-like answers "may be correct by happenstance, or may contain convincing falsehoods."

A lot more developers chipped in the discussion, describing AI Help as a "snake oil" potion which is "worse than useless" for the reasons described by Eevee and then some.

The most benign comments said that the AI help feature was likely to cause "much more damage than it would possibly help," while other, less sympathetic developers were quick to dismiss the feature as a "deeply misled" addition that would produce disinformation at scale – "like all other LLM applications."

Reading through the Yari issue, you get the idea that no one, absolutely no one seems to like the idea of a generative AI algorithm answering questions about actual web documentation. In the end, an MDN core maintainer named "sideshowbarker" said that AI Help seems to be something that Mozilla decided to do on their own "without giving any heads-up of any kind" to any other MDN stakeholders.

This new AI feature is a "monumentally bad idea," sideshowbarker remarked, and he promised to personally escalate the issue internally at Mozilla "as high as I can," with the aim of getting it removed "absolutely as soon as possible." The AI Help button has now been paused, and it won't be available anymore – for the time being, at least.

Permalink to story.

Mozilla is kind of like Microsoft, except they just have the political agenda and self-delusion without the revenue streams.
I couldn't understand the problem from this article. Had to go to the GitHub issue link which clarified it immediately. Bad writeup, IMO.
I couldn't understand the problem from this article. Had to go to the GitHub issue link which clarified it immediately. Bad writeup, IMO.

Paragraph 4 summarizes the root issue, in case you weren't aware of the LLMs shortcomings, or you hadn't follow all the conversation about Chat GPT & Co. for the last months. Assuming that you understand those, then the article is about applying it to Mozilla's Docs, which sounds dumb just by stating it (LLMs are great for creative work, though, although not what you'll want on your documentation search).