Facebook Recommends 490,000-Member Anti-Vaccine Group After Ending Fact-Checking

Facebook recommended a 490,000-member anti-vaccine parenting group to a reporter who never searched for it, The Guardian reported this week — three months after the group launched in March 2025, with Meta’s own AI summarizing hundreds of posts claiming vaccines are poison.

The group is called Crunchy Moms. The investigation, originally reported by The Guardian’s Michelle R Smith and carried by Genetic Literacy Project, found commenters routinely calling vaccines “poison” and insisting they cause autism and sudden infant death syndrome — claims Meta’s own policies still prohibit.

The stakes are not abstract. In August, a parent in Lancaster County, Pennsylvania asked the group whether to vaccinate a three-year-old during a measles outbreak. Members replied that children “cannot die of measles,” The Guardian reported. Pennsylvania has recorded six measles-associated deaths this year, two of them babies in Lancaster County.

This is the clearest signal yet of what changed when Mark Zuckerberg ended U.S. fact-checking in early 2025 and shifted moderation for lower-severity violations to user reports. Meta still bans claims that vaccines cause autism or that measles isn’t deadly. Enforcement, apparently, is another matter.

55 reports, zero responses

A Guardian journalist reported 55 posts and comments plus one video that appeared to break those stated rules. Facebook responded to none of the reports, The Guardian reported. Nearly all of the content stayed up.

Meta’s own systems saw it too. When the Lancaster County question drew 589 replies, Meta’s AI summarized the thread, noting commenters “claim the vaccine is unnecessary or even harmful.” The summary did not trigger a takedown.

Then there’s the recommendation engine. Facebook pushed Crunchy Moms toward users who never went looking for it — the reporting journalist among them. That reverses a 2019 pledge to stop recommending anti-vaccine groups entirely, a commitment made after measles outbreaks tied to unvaccinated communities.

“We were documenting Facebook recommending anti-vaccine groups to parents almost a decade ago,” said Renée DiResta, a Georgetown University professor who studies online manipulation. “I’m afraid that we’re going to see that same kind of dynamic all over again.”

The measured cost of algorithmic amplification

The behavioral impact of this class of content isn’t speculation. A study co-authored by Jennifer Allen — who teaches at NYU Stern and previously worked at Facebook — and published in Science calculated that during the Covid-19 pandemic, Facebook material expressing vaccine skepticism deterred about 5 million Americans from getting their shots, with roughly 21,000 deaths as a consequence.

“The fact that this group seems to exist and is being recommended to people, and there is so much clear misinformation in it, it seems to me that that means that there’s a failure along the way of Facebook,” Allen said, The Guardian reported.

Read that estimate against the current setup. The 2021 study covered a period when Meta still maintained fact-checking partnerships. Today the company has removed that layer and routes lower-severity violations to user reports — and the user reports, in this case, went unanswered.

What the founder says versus what the policy says

Group founder Sara Rupe told The Guardian she doesn’t censor members. “I’m assuming that if Meta sees that it doesn’t agree that it will just take it down,” she said.

That assumption is the whole problem. The moderation pipeline Zuckerberg described in 2025 — user reports flagging lower-severity violations — only works if the reports get processed. In this case, 55 reports produced zero responses. Rupe’s deference to Meta and Meta’s reliance on reports form a closed loop where nobody acts.

In a statement to The Guardian, Meta said it takes down misinformation that could plausibly lead to immediate physical danger and said that it was examining the instances the paper had flagged. That’s the company’s position as of the reporting. The group remained live and the content largely untouched when the investigation published.

The engineering gap between policy and enforcement

There’s a technical story here that goes beyond moderation staffing. Meta has spent years building AI systems capable of classifying content at scale — the 589-reply summary proves the comprehension layer works. What failed is the handoff between detection and action. The model correctly identified that commenters claimed the vaccine was “unnecessary or even harmful,” yet that classification produced no enforcement outcome.

Under the pre-2025 architecture, third-party fact-checkers formed a redundant check on exactly this kind of failure. Removing them didn’t just cut costs; it removed the only mechanism that operated independently of Meta’s own risk calculus around engagement. Group recommendations drive retention, and a 490,000-member parenting community is exactly the kind of asset a recommendation system is tuned to promote.

The 2019 reversal is instructive. Meta then acknowledged that its recommendation engine was actively funneling parents toward anti-vaccine communities and pledged to stop. Six years later, the same behavior is documented again — not as a bug the company denies, but as an outcome its own AI summarized without flagging.

Why the timing matters

The group launched in March 2025, less than three months after Zuckerberg announced the end of U.S. fact-checking. Its membership of more than 490,000, combined with Facebook actively recommending the group to users, suggests the post-fact-checking environment didn’t merely tolerate this content — it accelerated it.

For parents searching Facebook for vaccine guidance during an active measles outbreak, the platform now surfaces a community where the top-voted consensus contradicts public health guidance, and where the company’s stated prohibition on that content is unenforced.

Crunchy Moms has drawn the label of “the most dangerous” group on Facebook, and the timeline of the reporting behind this story is itself telling: the Guardian filed its 55 reports in May, July and September, and none received a response. The platform’s own AI summaries of the group’s content are not new, either — screenshots taken in October documented more than 10 such summaries, some dating back as far as December 2025.

Meta says it is reviewing the examples The Guardian provided. The six deaths in Pennsylvania this year — including the two babies in Lancaster County, where a parent was told children “cannot die of measles” — are the measure of what that review is weighing.

Photo of author

Technology Desk

Technology Desk at Archyde.

Asian Shares Brace for Worst Week Since 2020 Pandemic

Small Group of Isolated Awá-Guajá Briefly Emerges in Brazil, Asks for Food