I Tried Grokipedia, The Ai-powered 'anti-wikipedia.' Here's Why Neither Is Foolproof

Busà Photography/Moment via Getty Images

Follow ZDNET: Add america arsenic a preferred source on Google.

ZDNET's cardinal takeaways

• xAI's Grokipedia is now live.
• Articles connected nan level are written and "fact-checked" by Grok.
• It differs from Wikipedia successful respective ways, but besides rips it off.

On Monday, Elon Musk's xAI launched Grokipedia, an online accusation repository that's curated wholly by Grok, nan company's flagship AI chatbot. Musk has been promoting nan tract arsenic a benignant of anti-Wikipedia, which he has said has "a non-trivial left-wing bias" and has derisively called "Wokepedia." 

Also: Need nan champion AI contented detector successful 2025? Try these 4 devices (you astir apt already usage one)

Since its debut successful 2001, Wikipedia has amassed well complete 7 cardinal articles, each of which are written and edited by quality users. While nan thought down this exemplary was to let for nan introduction of a maximal diverseness of perspectives and opinions -- a benignant of self-correcting accusation instrumentality that would incline toward truth -- galore people, including nan site's cofounder Jimmy Wales, person based on that Wikipedia has go ideologically biased successful caller years. 

Just arsenic Grok was intended to beryllium an anti-"woke" replacement to celebrated chatbots for illustration ChatGPT and Gemini, xAI is promoting Grokipedia arsenic a much balanced and reliable type of Wikipedia. But really does it stack up? I tried it out. 

How it works 

Grokipedia has a bare-bones artistic resembling Grok's personification interface: I was greeted by elemental hunt barroom beneath nan matter "Grokipedia v0.1," suggesting that this is conscionable nan earliest loop of nan tract and that others tin beryllium expected successful nan future.

Screenshot by Radhika Rajkumar/ZDNET

I entered a taxable into nan hunt barroom and, akin to a hunt engine, Grokipedia showed maine a database of page titles to select. As of Wednesday morning, a tally connected nan site's homepage indicated that it had complete 885,000 articles. each of which are written and "fact-checked" -- according to a disclaimer supra individual articles -- by Grok.

Also: You're reference much AI-generated contented than you think

However, nan Verge reported Monday that immoderate Grokipedia articles see an inconspicuous disclaimer astatine nan very bottommost of nan page stating that nan contented had been "adapted from Wikipedia, licensed nether Creative Commons Attribution-ShareAlike 4.0 License." In different words, nan tract seems to beryllium utilizing contented taken straight from nan very tract that it was intended to replace.

Users pinch an X relationship tin item snippets of accusation recovered successful a station and click a fastener reference "It's wrong" to emblem inaccurate information, akin to the Community Notes attack some Meta and X traded existent fact-checking for early this year. xAI did not instantly respond to my petition for remark connected really personification feedback shapes Grokipedia entries.

Concerns

Sticking to nan facts isn't precisely AI's beardown suit. A study published past week by nan European Broadcasting Union and nan BBC that analyzed thousands of news-related responses from 4 starring AI chatbots (not including Grok) recovered that close to half had astatine slightest 1 awesome issue, including hallucinating (generating inaccurate accusation disguised arsenic truth), providing outdated information, and utilizing flawed sourcing methods.

Also: Get your news from AI? Watch retired - it's incorrect almost half nan time

While Grok was engineered "to maximize truth and objectivity," according to xAI, it's important to retrieve that it's an LLM, and for illustration each LLMs, it's imperfect. 

According to 1 nationalist leaderboard comparing nan mirage rates of frontier models erstwhile prompted to execute a elemental document-summarization task, Grok 2 presently scores comparatively high, meaning it's much prone to mirage than galore different starring models. Grok 4, nan newest loop of nan model, presently ranks successful nan 99th spot connected that aforesaid leaderboard, conscionable up of OpenAI's o4-mini-high and conscionable down Microsoft's Phi-4.

Screenshot: GitHub

xAI besides trained Grok successful portion utilizing nationalist posts and different accusation taken from X, which xAI has described arsenic "a unsocial and basal advantage" for nan chatbot. A study posted to nan preprint server tract arXiv earlier this month, however, recovered a causal narration betwixt training information gleaned from "junk" societal media contented -- think: high-engagement, low-quality posts -- and an inclination toward nan integer analogue of "brain rot": a noticable diminution successful nan trustworthiness of exemplary outputs, and an summation successful "dark traits," specified arsenic psychopathy.

Content sourcing questions

Then there's nan truth that galore users have reported incidents successful which Grok appeared to consult Musk's societal media posts aliases articles written astir nan billionaire earlier responding to definite delicate prompts. Similarly, Grokipedia appears successful immoderate important respects to bespeak Musk's personal, political, and ideological opinions.

Also: Why unfastened root whitethorn not past nan emergence of generative AI

For example, Musk has been outspoken astir what he regards arsenic nan existential consequence posed by declining commencement rates, which he described successful a December X post arsenic "the biggest threat to quality civilization by far." I came crossed a Grokipedia entry connected "Societal collapse" that includes a lengthy chat connected nan destabilizing societal effects of commencement complaint diminution passim history, which goes connected to picture really "mass immigration" compounds nan problem. I recovered a corresponding Wikipedia entry that included nary mention of a causal domiciled betwixt commencement complaint diminution and societal collapse.

Screenshot: Grokipedia

Another Grokipedia entry connected Grok echoed Musk's ain idealistic views of nan chatbot arsenic an anti-"woke" foil to its competitors.

Want much stories astir AI? Sign up for our AI Leaderboard newsletter.

"Unlike galore competing AI systems, Grok is engineered for maximal truthfulness, incorporating humor, a rebellious attack to spicy aliases rejected queries, and guidance to accepted governmental biases," nan introduction reads.

Bottom line

Ultimately, immoderate assets devoted to nan gathering and cataloging of accusation that's built by humans is going to bespeak immoderate grade of quality bias. The determination of which devices you should use, then, should travel down to which ones you see to person nan astir effective and trustworthy correction correction mechanisms successful place.

Also: I've been testing nan apical AI browsers - here's which ones really impressed me

As immoderate assemblage professor will show you, you shouldn't spot Wikipedia arsenic a superior root of information. It mightiness beryllium useful for getting a high-level overview of a person, place, thing, aliases humanities event, but nan double-edged quality of it being an open-source database intends that there's plentifulness of delirium connected nan site, too. You should ever corroborate thing you publication connected Wikipedia by checking much reliable accounts, for illustration peer-reviewed studies aliases articles published by reputable news outlets whose editors are devoted to journalistic integrity.

That said, I personally would spot much religion successful a website curated by quality writers and editors than a tract for illustration Grokipedia, which is managed by an LLM, pinch each of its foibles. And that would use to immoderate LLM, not conscionable Grok. Until we person models that are wholly hallucination-proof -- and it remains to beryllium seen if that's moreover technically imaginable -- you should ever return immoderate consequence you get from 1 pinch a hearty atom of salt, because there's a nonzero chance it's lying to you. 

Even without hallucinations, AI models tin transportation biases imported from their training data, which tin power their responses.