Analysis: Google's C4 dataset, used to train LLMs like Meta's LLaMA, has troubling content from 4chan, Kiwi Farms, white supremacist site Stormfront, and more
AI chatbots have exploded in popularity over the past four months, stunning the public with their awesome abilities … Tweets: @goldman , @timmarchman , @daveleebbg , @nitashatiku , @aaronjschaffer , @justinhendrix , @maxkennerly , @justinhendrix , and @blackamazon Tweets: Jason Goldman / @goldman : Very cool article on where the training data comes from. Bot needs to read more wowhead tho https://twitter.com/... Tim Marchman / @timmarchman : The internet-reading robot your boss wants you to partner up with has apparently, with the history of human civilization at its disposal, been reading “5 Reasons ‘The Marvels’ Will Be The Best Movie Ever (And Five Reasons Why It Will Suck)” https://www.washingtonpost.com/ ... https://twitter.com/... Dave Lee / @daveleebbg : Fascinating look at the more-than-iffy datasets behind AI https://www.washingtonpost.com/ ... @nitashatiku : Here's our analysis of the 15 million websites in just one highly-filtered CommonCrawl web scrape-used to train models like Google's T5 & Facebook's LLaMA -copyright symbol appears >200M times -pirated sites, 1 for e-books -half the top 10 = news sites https://www.washingtonpost.com/ ... Aaron Schaffer / @aaronjschaffer : Google's C4 data set, which has been used to instruct high-profile AIs including Google's T5 and Facebook's LLaMA, includes Russian propaganda site RT, anti-immigration site VDARE, white supremacist site Stormfront, anti-trans site Kiwifarms and 4chan https://www.washingtonpost.com/ ... Justin Hendrix / @justinhendrix : Thinking about the old phrase “garbage in, garbage out” https://www.washingtonpost.com/ ... Max Kennerly / @maxkennerly : I don't recall giving Google, Facebook, or anybody else permission to scrape 180k tokens—roughly the length of Orwell's 1984—from my blog for commercial purposes. And putting my copyrighted work in a blender with bigots doesn't make it better. https://twitter.com/... https://twitter.com/... Justin Hendrix / @justinhendrix : Holy cow. “The Post's analysis suggests more legal challenges may be on the way: The copyright symbol — which denotes a work registered as intellectual property — appears more than 200 million times in the C4 data set.” https://twitter.com/... @blackamazon : BUT RACISM ISN'T BUILT IN THOUGH RIGHT We had all these people and “experts” who SWORE they were gonna “listen” and doing all this bs smart talk But sure just takes years to figure that out And the way they kept blocking critics of racism in tech wasn't a clue https://twitter.com/...
Context & Ripple Effects
The Washington Post's analysis of Google's C4 corpus lands four months into the chatbot boom, when questions about what models actually read have moved from academic footnote to front-page concern. C4 was built by Google researchers by scraping the open web, and Meta then used it to train LLaMA — meaning the same dataset underpins both a Big Tech artifact and the open-source ecosystem that grew around Meta's release.
The finding also connects two earlier threads in this coverage: OpenAI's reliance on Kenyan contractor Sama to label violent content for ChatGPT showed that cleaning up model inputs has a human cost ($2/hour content-labeling work) long before anyone published a list of which forums were in the training mix, and publishers' growing toolkit of scraper-blocking measures like Anubis and Cloudflare's AI Labyrinth reflects a web already pushing back on being ingested wholesale (anti-scraping tools).
First-order effects
- Google, as C4's creator, faces direct scrutiny over how its dataset curation filtered (or failed to filter) sources like 4chan, Kiwi Farms, Stormfront, RT, and VDARE — and Meta inherits the reputational exposure through LLaMA, whose weights are now widely redistributed.
- Sites named in the report gain unwanted mainstream visibility, while any downstream developer building on LLaMA must explain why their product was trained on white-supremacist forum text.
Second-order effects
- Publishers and platforms weighing whether to allow crawlers get fresh ammunition: if inclusion in a major dataset means association with extremist content, blocking scrapers looks less like lost reach and more like brand protection.
- Rival model builders face pressure to document and differentiate their data pipelines — dataset provenance becomes a competitive claim rather than an internal detail, echoing the labor-cost disclosures that came out of the Sama labeling program.
Third-order effects
- If datasets keep inheriting the web's worst neighborhoods, the tacit assumption that publicly accessible text is free training material erodes, pushing the industry toward licensed corpora, opt-in data agreements, and heavier paid moderation layers.
- Dataset construction itself becomes a governance surface: whoever curates the corpus — not just who trains the model — acquires regulatory and legal exposure, making transparency reports about training data a likely compliance expectation.
The trend: AI training data is moving from an unexamined scrape of the open web toward curated, licensed, and increasingly contested supply, as the contents of corpora like C4 become public liabilities for the companies that build and reuse them.