Grokipedia 能为 Wikipedia 提供哪些关于 AI 应用的经验教训

AI & Journalism AI Search & Platforms Search & Platforms Audience & Creators Audience & Creators Local News Local News Reporting & Production Reporting Business & Revenue Revenue

About Nieman Lab Subscribe Archives Contact

Nieman Foundation Nieman Fellowships Nieman Reports Nieman Storyboard

Appearance

Auto Light Dark

Text size

A A A

Get Nieman Lab in your inbox.

All the latest headlines and insights from the world of digital journalism and publishing. Our daily newsletter is a must-read from newsrooms to executive suites.

Leave this empty

Email address Name (optional)

We will never sell your address. Prefer a weekly email?

Nearly there — check your email.

We have sent a confirmation link to . Click it and you are on the list. If it has not turned up in a minute or two, it may be in your spam folder.

Subscribe another address

Elon Musk’s Grokipedia is littered with mistakes. But Wikipedia could still learn some things from it.

Grokipedia shows that AI is actually pretty good at finding tidbits in the far corners of the internet.

By Bill Adair Nov. 2, 2026, 10:59 a.m.

When I saw Grokipedia’s articles about me, the memories came rushing back: my 19 years in minor league baseball, including that impressive season with the Montgomery Rebels when I hit 20 home runs. I was lean back then — 6 feet tall and 168 pounds — and earned a career batting average just under .300. I became a minor league coach and managed the Chicago White Sox for a short stint in 1970, which Grokipedia said “highlighted [my] readiness for higher-level responsibilities.”

I then attended Arizona State University, switched to a career in journalism, created the fact-checking site PolitiFact, won a Pulitzer Prize, and now teach journalism at Duke University.

Ah, the memories!…except that I never played minor league baseball, never managed the White Sox, and can only dream about being 6 feet tall. (I’m 5’8″.) Grokipedia, the AI-written encyclopedia of Elon Musk, has confused me with Marion Danne “Bill” Adair, who died in 2002.

That article and another one about “Bill Adair (journalist)” — which really is me — illustrate the many flaws of Grokipedia, which launched a year ago. The articles are littered with mistakes and superfluous details. They are terribly written and ridiculously long. If Grok, the AI engine that wrote this turgid prose, were a student in one of my writing courses, I would give it a D.

Still, the articles provide valuable lessons for Wikipedia editors who are debating how and when to use AI. Grokipedia adds new content and context that Wikipedia often lacks. It shows that AI is actually pretty good at finding tidbits in the far corners of the internet and summarizing them in sweeping prose. It can also provide balance to articles that need it.

Grok can’t write worth a lick, but it shows how Wikipedia can take advantage of AI at a time when overworked editors are trying to keep the encyclopedia alive.

Musk launched Grokipedia as an alternative to Wikipedia, which he said was too liberal. (He called it “Wokepedia.”) Still, he trusted it enough to have Grok copy hundreds of thousands of its articles, which is permitted under Wikipedia’s Creative Commons license, and then add content from elsewhere on the internet.

With his usual bluster, Musk said his encyclopedia used a new version of his large language model that was good at “cogent analysis.” He said programmers told it to “cycle through the million most popular articles in Wikipedia and add, modify, and delete.”

When Grokipedia launched, Musk said he had big aspirations to overtake Wikipedia, which has 68 million articles in more than 300 languages. He said his encyclopedia would be more complete and more balanced and would make smarter use of images and video.

Maybe someday.

For now, Grokipedia is a massive collection of overly long articles that are often out of date. Its flaws have been well-documented: Articles are littered with mistakes, peddle conspiracy theories, rely on low-credibility sources, and sometimes skew to the right. And although Grokipedia shows edit requests from readers and whether they are approved, the site is not as transparent as Wikipedia, which reveals details about its editors and every change they make.

Last spring, Grokipedia mysteriously stopped updating, according to a report by researchers Renée DiResta and Ronald Robertson. It resumed in late September, still a hot mess, but with a more colorful homepage that always seems to feature its article about Musk.

Grokipedia has two articles about me: the conflated baseball player one and “Bill Adair (journalist),” the real one, which is incredibly boring, even to me. Both borrow heavily from Wikipedia’s articles about us and enrich them with details found on assorted websites.

Wikipedia’s article about me is nicely concise — just 700 words. It is written in the encyclopedia’s signature dry style and doesn’t contain any errors. The article recounts my career, some of the major news events I’ve covered, the two books I’ve written, and the awards I’ve won. But it highlights trivial events from my career, fails to emphasize larger ones, and lacks any criticism of my work. (I have not tried to improve the article because Wikipedia’s conflict of interest policy prohibits people from editing their own bios.)

By contrast, Grokipedia’s “Bill Adair (journalist)” is painfully long — 5,600 words. Its greatest value would be to help someone suffering from a sleep disorder.

Like a college student trying to meet a word count, Grok stuffed the article with needless details and redundant descriptions. The article tells you about my work with PolitiFact and then tells you again and again. It explains the site’s rating scale eight times and mentions my academic title four times.

Despite labels that say “fact-checked by Grok,” both articles had lots of errors. I spotted 32 in the journalist one and so many in the baseball mashup that I couldn’t count them all. Mistakes are not surprising in AI writing, but I discovered interesting patterns that reveal the weaknesses of large language models.

In the article about the real me, only seven of the errors were hallucinations in which Grok seemed to create a detail out of thin air. No, I did not win an award from Manhattan College, and no, PolitiFact did not begin by fact-checking the Tampa mayor’s race.

Five other errors showed Grok was a lazy researcher that did not look broadly enough to check a claim. Grok said I am the director of Duke’s journalism program, but I left that job five years ago. The sources cited in the article don’t explain Grok’s mistake, so it seems likely it remembered an old announcement and failed to see a newer one that named my successor.

The most prevalent errors were in a broad category I’d all cluelessness. These mistakes showed Grok didn’t fully understand what it read, like someone who skimmed an article too quickly. My favorite in this category was when I got credit for a big scoop in 2000 that George W. Bush had a DUI conviction, which the Grokipedia article said had an impact on “late-election discourse.”

I wish! Yes, I was among the many journalists who covered that story, but I had no scoop.

Together, the errors reveal that Grok is a knucklehead. It doesn’t know what’s important and what’s not. It is simply mashing together details from many web articles about me. Grok only knows what it found on the internet and it lacks experience and perspective to sort out what’s important.

The writing is dreadful. The articles read as if Grok was trained on too many online resumes and corporate memos. Grok uses words and phrases like “commenced” and “transitioned” and “multidisciplinary practicum” and “hypotheses on fact-check efficacy.” A typical sentence: “Adair’s approach in this position reflected a commitment to undiluted sourcing and causal analysis of policy impacts, influencing subsequent innovations in political journalism.”

Take a close look at that sentence and you’ll see it’s just a mush of buzzwords. I don’t know what “undiluted sourcing” is and have never used the phrase, nor would I refer to “causal analysis”…of anything. Grok is trying to impress us with an MBA vocabulary but is really just an epic bullshitter.

Bernadette Meehan, CEO of the Wikimedia Foundation, which provides money and technical infrastructure for the encyclopedia, has called AI “an existential threat” to Wikipedia.

The 25-year-old encyclopedia has long been a source and beneficiary of Google searches. Wikipedia provides the answers and in turn benefits from readers clicking through to read articles. But when people make a query to a chatbot, they often get the answer they need in an AI summary and don’t click through to read the encyclopedia article. That threatens donations because fewer readers see Wikipedia’s fundraising pitches.

Wikipedia has banned content generated by LLMs, but that hasn’t stopped many contributors — some well-intentioned, some not — from dumping an avalanche of AI-written content on the site. That has put a huge burden on the site’s “New pages patrol,” the editors who approve articles, because they must weed out the lousy content. (Wikipedia has a helpful guide of LLMs’ favorite words and phrases, such as “tapestry” and “meticulously” and sentences that start with “additionally.”)

Still, terrible as it is, Grokipedia offers lessons in how AI can help Wikipedia.

Grok’s voracious internet searches sometimes find things that human editors don’t. The long article about me includes a substantive discussion of my research about lying in politics and my work with fact-checkers around the world, neither of which are in the Wikipedia version.

Grokipedia also provides some welcome balance. When it offers context, Grokipedia can be a little heavy with hype (“His work has been lauded for innovating transparency in media…”) But it also included a “Criticisms and Controversies” section that details complaints that PolitiFact is biased against Republicans. That section has a healthy back and forth about the critiques and my response. That kind of balance could be helpful for Wikipedia, which has taken heat from Musk and others for being too lefty.

Musk’s encyclopedia also has a playful spirit that Wikipedia could learn from. Grokipedia’s homepage features random samples of recent edits that provide a glimpse at how the site is being improved. Another Grokipedia feature reads articles aloud in an AI voice you select: Eve, who sounds like a BBC anchor; Ara, a serious woman with crisp diction; Leo, who might have been a professor at Oxford; and Sal, whose name suggests a sports bar but whose voice is NPR.

My big takeaway from Grok is the importance of human judgment in curating knowledge. Grok is a bot trying to behave like a human editor. Its many mistakes and lousy writing show that AI is not ready to replace Wikipedia’s talented volunteers. Still, Grok shows that AI can be a tool that helps them research and write better articles.

After seeing an AI encyclopedia up close through my experiences with Grok, I come away with hope. Yes, AI can be a sloppy researcher and an awful writer. But Grok’s performance, shaky as it was, highlighted AI’s readiness for higher-level responsibilities.

Bill Adair, the Knight Professor of the Practice of Journalism and Public Policy at Duke University, is writing a book about how Wikipedia editors are adapting to the age of AI.

Get Nieman Lab’s Daily Digest

The most important stories about journalism and the future of news, delivered to your inbox every afternoon.

November 2, 2026

See more onAI & Journalism

The latest from Nieman Lab

Press Forward gave $20 million to 205 local newsrooms over two years. Here’s how some of them spent it.

The unrestricted funding helped small outlets plug holes and try new things, but publishers ran up against the limits of what two $50,000 grants could accomplish, especially when it came to hiring.

Sophie Culpepper

Only 32% of Americans have high confidence that news orgs will report midterm election results accurately, survey finds

Sarah Scire

The Nieman Lab News-About-News Quiz: October 30, 2026

Need another reason to read that 47,212th article about paywall strategy? What if it earned you a point?

Joshua Benton

The New York Times is adding more AI features to its Cooking app, and users aren’t thrilled

Joshua Benton

The Independent Journalism Atlas looks at “wells” and “springs” in news deserts

Sophie Culpepper

Help advance the Nieman Foundation's mission “to promote and elevate the standards of journalism” by making a donation.

Nieman Foundation for Journalism at Harvard
Walter Lippmann House, One Francis Ave. Cambridge, MA 02138
(617) 495-2237

Get Nieman Lab in your inbox.

All the latest headlines and insights from the world of digital journalism and publishing. Our daily newsletter is a must-read from newsrooms to executive suites.

Leave this empty

Email address Name (optional) Subscribe We’ll never sell your address. Prefer a weekly email?

Nearly there — check your email.

We’ve sent a confirmation link to . Click it and you’re on the list.