Do we really need Wikipedia at all? That would have been considered a bizarre question even five years ago, but today the answer is less clear. There is already one AI-based online encyclopedia, Grokipedia (grokipedia.com), which launched in 2025; tomorrow there could be ten, with potentially each AI company, each search engine, and maybe even political organizations and entire governments offering their own alternative. Grokipedia itself feels like a proof-of-concept at this point, with few images, and availability only in English. These are solvable technical issues: for example, translating the contents of any particular AI-generated encyclopedia into dozens of languages seems simple enough, for Grokipedia or any other "vendor". (Covering over 300 languages, as Wikipedia does, may be out of reach at the moment.) For AI companies, it would be an advertisement for the strength of their LLM. For search engines, it would be an extension of what they already do: if you phrase your Google search in the form of a question, for example, there is a good chance that an "AI Overview" will appear that is not entirely different from a short encyclopedia article, complete with both external and internal links. And for governments and activist organizations, it could be a way to promote their worldview on controversial topics. (Some countries have already funded or at least promoted their own non-AI Wikipedia alternatives, although for now this has been the domain only of authoritarian countries like China, Russia and Cuba.)

Even assuming that Wikipedia always offers comprehensive and accurate information, there are still advantages that AI encyclopedias can offer. Grokipedia was explicitly created in order to offer a more right-wing slant than Wikipedia’s (its creator, Elon Musk, has referred to Wikipedia as "Wokepedia"). Other AI encyclopedias could be configured to output themselves in a variety of different ways: taking any other position on the political spectrum, for example, or making articles shorter or longer, or citing more or fewer questionable sources like YouTube videos.

In fact, it’s entirely possible that an AI encyclopedia in the future could offer this sort of customization directly to the user, with the user specifying a series of preferences, from language to political standpoint to preferred article length, and the contents of each article then being tailored to those preferences. To take a relatively anodyne example, on the English-language Wikipedia there is a sort of perpetual slow-motion war between those who want to keep and expand "In popular culture" sections, listing all the times the subject of an article has been referenced or depicted in movies, songs, etc., and those who want to remove or greatly reduce these listings on the grounds that they are frivolous. Both sides act on the assumption that there is an ideal set of popular culture references for each article – but what if there isn’t? What if some readers would like to read about all the times that, say, Charlemagne has been mentioned on TV shows, but others do not, and that’s how it will always remain? For that matter, what if the same person sometimes wants to see ancillary information like this, and sometimes not? An AI-based encyclopedia could easily offer that customization with a series of checkboxes and buttons: no more need to wage battles.

One analysis holds that more than 50% of the articles on Grokipedia are copies, in whole or in part, of the equivalent article on the English-language Wikipedia. Perhaps in the future, though, AI-generated encyclopedias will stop relying on Wikipedia altogether and instead go directly to their own set of (hopefully) reliable sources: books, newspapers, scholarly articles, etc. There is nothing sacrilegious about this idea: Wikipedia openly states that it should not be considered a reliable source (Wikipedia articles cannot cite other Wikipedia articles, for example). So presumably a machine-based encyclopedia, if it wanted to get as close as possible to the truth, would skip the middleman and go directly to the actual reporting.

Where does all this leave Wikipedia? Perhaps only as a relic of, as they say, a simpler time, when wanting to impart a piece of knowledge to the world meant going to a keyboard and typing it all out.

On the other hand, perhaps things will stay mostly as they are. For all the talk of constant progress in AI, there is also the perception among some that generative AI is hitting a plateau in its abilities: that five years from now, it will undoubtedly be better at everything but its output will still be a little "off", with images that look a little too painterly, videos that look unnatural in a dozen different ways, human vocals that sound lifeless, and text that is unnecessarily wordy. Wikipedia, though its readership may continue to shrink, will remain the definitive source for encyclopedic information. This seems to be the view of the average Wikipedia enthusiast, perhaps stated in an increasingly defiant tone: "Knowledge is human", proclaimed the closing credits of the 2026 "Wikipedia 25 virtual birthday party" online video – a statement that seems to hinge on a very specific meaning of the word "knowledge". (And perhaps accidentally insults animals.)

If Wikipedia goes away or declines, one big danger is that AI will increasingly use the output of other AI as its text training data. There is a term for this phenomenon: "model collapse", and the fear is that hallucinations and other errors will magnify over time, leading to all sorts of chaotic output (though still stated with a high degree of certainty, of course). LLM makers are keenly aware of this risk, and in some cases have sought out assurances from the Wikimedia Foundation that AI will not be used to generate the text on Wikipedia. If even AI companies want Wikipedia to stay as it is, that’s in a sense as big an endorsement of human-generated content as they come.

So what would Wikipedia look like, in a world with an ever-greater usage of AI, but with Wikipedia remaining as the standard encyclopedia of choice? Of course, the main answer is that it would look like it does now. There will undoubtedly be more "digital offerings" around Wikipedia. There are already dozens of podcasts in which the hosts read out and/or discuss Wikipedia articles (some of the former are marketed as sleep aids, for what it’s worth). The next step is probably short-form videos around Wikipedia content, of the kind that have taken over social media. These would presumably have to be AI-generated in order to have enough of them to make the whole enterprise worthwhile. Ironic, perhaps – but it’s certainly possible that AI will emerge as a tool that’s useful for some tasks more than others.

Wikipedia does not exist in a vacuum: its images and other files are hosted on the site Wikimedia Commons, and for certain languages (though not English), the majority of infobox data comes from the site Wikidata. In the future, some amount of the text (again, probably much less so for English) may come from the planned site Abstract Wikipedia, written in a highly technical-looking syntax (currently nicknamed "Abstractese") that is planned to be automatically translatable to each language. What if these three resources – Commons, Wikidata and Abstract Wikipedia – ultimately become a more important source for LLMs than Wikipedia itself? The use of Wikidata, especially, hearkens back to the dream of the "semantic web", which was conceived in 2001 and peaked in hype around 2010. But perhaps LLMs will eventually see this dream reach its fruition. Structured data is much less ambiguous than text, and less likely to be the source of hallucinations. It also lends itself to more sophisticated querying – a question like "What are the 10 biggest lakes in Angola?" can much more accurately be answered by a SPARQL query (of Wikidata, presumably) than scouring its memory for facts involving lakes in Angola.

None of this is news to anyone who has ever shown interest in the semantic web. In fact, for the semantic web, SPARQL querying is part of a bigger vision that involves querying a massive variety of structured data sources, along with some agentic capabilities, with the ultimate goal being a system that can execute commands such as "Book me a vacation next month near any of Angola’s 10 biggest lakes". Of course, it’s conceivable that a purely text-trained AI system could accomplish this as well, but having RDF/SPARQL-based querying, rather than simply word matching, would presumably be more likely to provide a successful outcome.

But back to encyclopedias. A system that used AI to generate articles with text, images and infoboxes generated from some combination of Commons, Wikidata and Abstract Wikipedia could represent the best of both worlds: the reliability of human-created (or at least human-supervised) content along with the customizability of a computer-generated end product.