The world’s most important knowledge platform needs young editors to rescue it from chatbots – and its own tired practices
Established in 2001, Wikipedia is an “old man” by internet standards. But the role it plays in our collective knowledge of the world remains astonishing. Content from the free internet encyclopedia appears in everything from high-school term papers and pub trivia questions to search engine summaries and voice assistants. Tools like Google’s AI Overviews and ChatGPT rely heavily on Wikipedia, although they rarely credit the site in their responses.
And therein lies the problem: as Wikipedia’s visibility diminishes, reduced to mere training data for AI applications, it also loses prominence in the minds of readers and potential contributors. When someone notices a topic that is poorly described on Wikipedia, they might feel motivated to correct it. But this can-do spirit goes away when the error comes through an AI summary, where the source of the information isn’t clear.
It’s actually not easy to ensure that an LLM will cite a correct source, in the same way it’s not easy to ensure that it will provide accurate information. It’s based on token probability, not deterministic lookups of “this data came from this source.” It could entirely make something up, then write “Source:” and then probabilistically write “Wikipedia” because those tokens commonly follow those for “Source.”
If you have an AI bot that looks up information in real time, then that would be easy. But for a trained LLM, the training process is highly destructive. Original information is not preserved except in relationships based on probability.
I choose to interpret the grandparent commenter’s use of “easily” to mean “not impossible, and an ethical obligation, so you’d better fuckin’ make it a priority.”
That’s accurate. Nothing in technology is actually “easy” and I know it requires a lot of work. Didn’t mean to diminish all the time and energy put into making this stuff. Thanks for better expressing what I meant.
Right, in my experience the majority of URLs generated by LLMs are just jumbles of letters that vaguely look like a URL. A fundamental architecture difference needs to happen in one way or another to properly cite sources, and it’s really bad for performance.
The more I learn about AI, the less I like it.
It’s a fun toy. It’s not a research aid, it’s not a productivity tool, and it’s not particularly useful in the workplace.
It’s honestly very similar to the VR craze of a few years back. Silicon Valley invented a fun toy and then tried to convince everyone that it would transform the workplace. Meetings in VR and simulated workstations and all that. Ultimately everyone figured out that VR is completely useless in the workplace and Silicon Valley was just trying to find ways to sell their fun toy. Now we’re going through the same learnings with AI.
I love VR. I have so many hours in some of the slower paced fps titles that it’s almost matched my video game time total for non-vr games on steam.
The one thing I learned for sure is that I don’t want anyone else telling me when I have to put on the headset and when I’m allowed to take it off.
Never will wear a vr headset in the workspace.
Oh yeah, I love my Index as well. I think it’s a lot of fun as a gaming device. But the big money is in B2B sales, which is why tech companies try to convince everyone that blockchain/VR/LLMs have all these corporate applications that just make no damn sense.
It’s actually extremely useful as a productivity tool in many workplaces. You’re just stating how you feel about it as if it’s fact for everyone
It’s a great way to replace competent human workers with a lower-cost, lower-quality alternative. Wall Street may buy that anti-worker BS but workers tell a different story.
Literally an article in Forbes today that says 77% of employees report that AI tools make them less productive: https://www.forbes.com/sites/torconstantino/2024/09/12/77-of-surveyed-employees-say-ai-tools-make-them-less-productive/
I personally use LLMs in the workplace. It’s great for generating boilerplate code, especially stuff that is often very repetitive like test classes.
perplexity.ai does a decent job at providing sources for searches.
Yeah Bing Chat had sources for a while (not sure if it still does) and when I checked the sources, the frequently didn’t contain the claim in question. So even if you get it to cite real pages, it just doesn’t work the same way as human citations do.
Maybe specifying source should be a legal requirement if the LLM service provider shouldn’t automatically be held accountable for the answers their services produce?