Let’s say you were scanning the recent headlines and wanted to know when the East Wing of the White House was constructed.
You might ask ChatGPT. Or maybe you’d just google it.
Either way, you’d get a good answer — although slightly different versions. (Google’s AI answers has a snippet from Wikipedia that notes the East Wing was originally constructed in 1902; ChatGPT offers the 1942 two-story version.)
The point is, I don’t actually have to visit Wikipedia’s website to find out, even if the information is sourced from Wikipedia.
That’s starting to become a real concern for people who care about Wikipedia. Traffic is down 8% over the past few months when compared to the same period last year, according to Diff, the blog of Wikipedia’s parent organization, the nonprofit Wikimedia. AI-driven chatbots, fed by large language models that hoover up information from sites like Wikipedia, are largely to blame for the decline, Wiki says.
That sounds bad — is it? And what does that mean for the future of Wikipedia? It’s so ingrained into the web that it feels too big to fail — but could it?
I asked Marshall Miller, senior director of product at Wikimedia, to make sense of the data. He told me it’s certainly true that AI needs Wikipedia, which is one of the most crucial datasets that make up the backbone of the LLMs’ knowledge.
“Our mission is to spread free knowledge, and so it is a good thing when that knowledge reaches people through new platforms, whether that’s AI chatbots, search engines, or social media,” he said. “People need access to reliable, neutral knowledge, and we are happy to see the knowledge reach more people in new ways.”
But it also puts the organization in a difficult position when LLMs are using Wikipedia’s knowledge base without referring the traffic to the site that can help keep its ecosystem healthy: new contributors, new editors, and — crucially — new donors to the nonprofit Wikimedia Foundation.
Why a drop in traffic matters to Wikipedia
A key difference between a news publisher like Business Insider or The Wall Street Journal seeing a drop in traffic, and Wikipedia seeing a drop, is that the business models are quite different.
Most news outlets run ads or rely on subscriptions, or both. Wikipedia is run by a nonprofit, relying in large part on donations. So for Wikipedia, fewer people visiting it each month means fewer people prompted to donate.
It could also mean fewer contributors and editors to add to and improve on Wikipedia’s articles. I’m slightly skeptical of that threat, however. I suspect the kind of person who actually edits a Wikipedia article is a specific type of power user, not a casual drop-in who just wanted one piece of information they obtained by an AI overview.
Also, as the Diff blog post points out, the site’s traffic dip isn’t only because of AI. There are other trends at play, like younger people getting information from video and skipping traditional web searches altogether.
AI eating into traffic isn’t the only concern on the table at the moment for Wiki: A more potentially urgent threat is political. This summer, Republican lawmakers launched an investigation into Wikimedia’s alleged left-leaning bias. And Elon Musk, who has accused Wikipedia of having a left-leaning slant, just launched his Grokipedia.
But there are bright spots: Its site and apps still draw more than 10 billion views a month, and the Wikimedia Foundation brought in $170.5 million in donations last year.
And there’s a mechanism for the AI companies to actually pay Wikipedia for its information: Wikimedia Enterprise sells an API subscription. (There’s a decent question about how much leverage Wikimedia has when a lot of AI companies have already helped themselves to a huge part of its data for free.) But it’s not impossible to see a path toward deals. OpenAI, for instance, has made deals with various news publishers, including Business Insider’s parent company, Axel Springer.
“Generative AI depends on Wikipedia’s human-created knowledge,” Miller said. “Wikipedia is one of the highest-quality datasets used in training LLMs, and studies have shown that the outputs from AI models are significantly lower quality when Wikipedia is not used as a dataset.”
How you might encounter Wikipedia when using AI
I’ve been thinking about the ways that people might encounter Wikipedia information in an AI setting, vs. actually browsing Wikipedia itself. And I think there are some big differences — which should be good news for Wikipedia.
I can attest that a lot of the factual-type questions I might ask ChatGPT are being sourced from Wikipedia, and I bet that’s also the case with a lot of the homework-type questions that younger users are using AI for. “What was the cause of the 30 Years’ War?” is a great question for ChatGPT. It’ll give you a reasonable and concise answer. (Kids, please do your homework yourself.)
But consider instead what you find on the Wikipedia page for the Thirty Years’ War. So many interesting blue words to click on! So many names and places and things! You could get lost there for weeks.
We all love a good Wikipedia rabbit hole to get sucked down — that’s a very different experience from wanting a quick answer. It’s like sitting down to a meal at a restaurant vs. microwaving a frozen burrito. I say this as someone who eats a fair amount of frozen burritos — sometimes you just want the quick and easy thing! The two versions coexist!
That experience — the browsing experience where you absorb information you didn’t even think to ask for — is what sets Wikipedia apart from an AI answer of a simple fact. And that’s what the people who donate or edit keep coming back for.
Read the full article here
