Slop on Facebook

Researchers studied 100+ Facebook Pages posting large volumes of AI-generated images (shrimp Jesus, fake children, AI "artwork," etc.); together they attracted hundreds of millions of interactions. Users did not have to follow the pages — "simply viewing several AI-generated posts was often enough for Facebook's recommendation system to fill their feeds with more." The researchers noted comments on mundane AI images (housewares, homes, children's "artwork") suggest many users are unaware of the synthetic origin. (source)

Documented spam Pages with 100,000+ followers each; operators in Pakistan, India, Vietnam, Thailand, Indonesia posting dozens of AI images per day using free tools (Microsoft AI/Bing Image Creator); some operators ran 10,000–13,000 fake accounts. Copy-pasted prompts target US audiences ("american soldier veteran holding cardboard sign that says 'today's my birthday, please like'"). (source source source)

Indian YouTuber Gyan Abhishek coached viewers to "make money through performance bonus," claiming "$100 for 1,000 likes"; a Kenyan slop creator told New York magazine he prompted ChatGPT with "WRITE ME 10 PROMPT picture OF JESUS WHICH WILLING BRING HIGH ENGAGEMENT ON FACEBOOK"; in a podcast, a spammer showed earnings of $431 from one AI image of a train made of leaves ("People don't even make this much money in a month"). (source source source)

Many slop Pages link to AI-generated "news" sites stuffed with Google ads or low-quality product sales — the engagement farm funds click farms. NewsGuard tracks 3,749+ "Unreliable AI-Generated News" sites across 16 languages. (source)

Reporting consistently notes Facebook's aging user base as especially exposed: users earnestly comment "Amen" on shrimp Jesus; older users are more likely to mistake AI images for real ones (NBC News reporting via ibtimes piece above). A 2025-era analysis found ~75% of impressions on the 30 most active scam accounts came from users 65+ (see §3).

WebProNews (Feb 2026) describes feeds "nearly devoid of posts from actual friends, replaced instead by bizarre AI imagery, recycled memes from pages they never followed." Futurism (via webpronews) notes slop turning "dark and weird" — unsettling synthetic imagery beyond simple curiosities. (source)

Facebook ranks by predicted engagement (watch time, saves, shares, comments). Slop is engineered to trigger exactly these signals: emotional imagery (Jesus, soldiers, babies, puppies), "today's my birthday, please like" captions, "Pray/Share/Rate this" prompts, repetitive "Amen"/"Beautiful" comment bait. "Content attracting comments, reactions and shares is more likely to be promoted, regardless of whether it came from a genuine community or an AI-powered spam page." ( — ibtimes.sg article above)

Meta's own AI blog states AI-driven recommendations deliver content from outside users' networks; Meta said more than 20% of Facebook and Instagram feed content is AI-recommended from accounts users don't follow (source). Zuckerberg said in 2022 Meta expected to "more than double" that share (source). Marketing analyses put unconnected content at 35–50% of feed by 2025–2026 (secondary sources, weaker evidence).

DiResta & Goldstein found users were pulled into slop via Suggested-for-You recommendations after viewing only a few AI posts — no follow required. 404 Media (2024): "Facebook's recommendation algorithms are promoting bizarre, AI-generated images being posted by spammers and scammers." ( — ibtimes.sg; 404 Media URLs above)

Rathje et al. (2021) found outgroup-hostile/negative-emotion posts shared ~2× as often on Facebook/Twitter; Guess et al. (2023, Science, Meta collaboration) found switching users to chronological feeds reduced time spent and decreased exposure to uncivil content and slur words on Facebook; Piccardi et al. (2024) showed filtering hostile content reduced engagement — i.e., provocative content is "engagement-efficient," giving platforms a revenue reason to tolerate it. (source source)

Meta's Performance/Creator Bonus program paid participants for viral posts, and slop operators explicitly optimized for it; 404 Media verified "consistent viral posting can lead to significant earnings for users in countries like India." Meta's defense to 404 Media: the program "works as intended if content meets community standards and engagement is authentic." Gizmodo (Aug 2024): "Meta explicitly allows AI-generated images and the only metrics it cares about are the ones that improve its bottom line… The slop makes everyone money." Meta spokesperson to 404 Media: "We encourage creators to use AI tools to produce high-quality content that meets all our Community Standards, and we take action against those who attempt to drive traffic using inauthentic engagement whether they use AI or not." (source source)

• Judy Skene, 86, Ontario: lost $900,000 after clicking a Facebook ad (summer 2025) featuring an AI deepfake of PM Mark Carney promoting a crypto platform; scammers coached her into a $300,000 mortgage. Canadian Anti-Fraud Centre: $1.2B+ in investment-fraud losses reported since 2022, mostly crypto; calls such stories "increasingly common." (source)
• Maurine Meleck, 82, Florida: lost ~$200,000 in retirement savings via a Facebook video deepfake of Dr. Pierre Kory (Oct 2024) pushing a crypto "special fund." (source)
• Wisconsin: seniors reported $92M+ in losses; FBI's 2025 internet-crime report was the first to track AI scams, with most complaints from ages 60+ (phishing/spoofing, crypto). Katie Paul, Tech Transparency Project: "they cannot execute that scam without Meta reviewing, approving, and collecting money from the ads they are running." (source)
• Scale: 30 most-active scam accounts earned ~215 million impressions in a year, ~three-quarters from users 65+ (source); Americans lost $2.1B to social-media scams in 2025; FTC: consumers lost $12.5B to fraud in 2025 (+25% YoY). Federal class action in Oakland, CA alleges Meta financially benefits from fraud on its platform.

Researchers warn older users who mistake AI images for real become targets for fake charities, scam products, and emotionally manipulative funnels that build trust before redirecting users off-platform (ibtimes.sg).

Facebook spammers posted AI-generated images of Holocaust victims with fake stories; the Auschwitz Memorial called them a "dangerous distortion" (only a handful of real Auschwitz photos exist). History-focused Facebook groups "inundated" with AI-generated "historical" photos; plant-care misinformation proliferated, with online retailers selling seeds of nonexistent plants using AI flower images. ( — Wikipedia "AI slop" via pengen.diewe.workers.dev link above)

Feeds "less like a community hub and more like a fever dream" (Futurism via webpronews); users can't reliably distinguish genuine human content from synthetic engagement bait — "one of the platform's biggest challenges" (ibtimes.sg).

CCDH's June 9, 2026 "Safety Off" report (source) found abusive comments toward US Congress members tripled (24,000→78,000), violent threats quadrupled (1,800→7,600), and hate quadrupled (6,900→30,000) in the six months after Meta's Jan 7, 2025 enforcement rollback — evidence that the 2025 pullback in proactive enforcement measurably increased harmful content exposure.

claims ~91% enforcement precision on Facebook, with violent/graphic content prevalence at 0.20–0.21% of views (Q2 2026) and bullying/harassment at 0.07–0.08% (Q1 2025). (source)

Policy analyst Metehan Durmaz found the January 2025 changes narrowed proactive moderation — violating content is now removed before a user reports it only for "the most serious and illegal" content; everything else waits on user reports, which are "generally less reliable than automated enforcement." New definitions and thresholds make Q1 2025 data incomparable with earlier quarters, so reported prevalence "likely underrepresents the actual number of lower-severity violations." Meta's claimed ~50% drop in enforcement "mistakes" is a byproduct of reduced proactive enforcement: "Meta is simply 'looking' for fewer violations, not necessarily because the platform is safer." (source)

Meta ended the US third-party fact-checking program (replaced with an X-style Community Notes model), rewrote its Hateful Conduct rules to permit previously banned speech — including calling LGBTQ people "mentally ill" and referring to women as "household objects or property" — loosened rules around immigration status, and stopped proactive automated scanning for less-severe violations, so hate, harassment, and violence/incitement enforcement now largely waits on user reports. (Meta's own announcement page blocked direct access during research; the change list is documented via Reuters, the Oversight Board, SMEX, and CCDH.) (source source)

the changes were announced "hastily, in a departure from regular procedure, with no public information shared as to what, if any, prior human rights due diligence the company performed." It issued 17 recommendations, including a human-rights impact assessment focused on Global Majority countries, LGBTQIA+ people (including minors), and immigrants. Meta declined to remove the term "transgenderism" from its hate-speech rules. (source)

nearly 8 million Facebook comments directed at 100 US Congress members, six months before vs. six months after the change: abusive comments tripled (~24,000 → 78,000); violent threats quadrupled (1,800 → 7,600); hateful comments quadrupled (6,900 → 30,000); harassment more than doubled (15,700 → ~40,000); threats against President Trump doubled. (CCDH notes its AI classification is an estimate, not perfectly accurate.) (source)

11 Instagram accounts selling antisemitic/racist "hate merch" (Nazi imagery, Holocaust denial, "Life is a race. Be a racist") were all recommended by Instagram's own algorithm, amassing 1.5 billion combined views; views rose 3.75× after the rollback, with estimated sales over $1.3M. (source)

1 in 6 users in Meta-defined protected-characteristic groups reported being targeted with gender-based or sexual violence on Meta platforms since January 2025; 66% witnessed harmful content; over 90% feel less protected. (source)

Meta removed a ~80,000-member Facebook group used to warn Chicago-area residents about ICE sightings, following Department of Justice demands and without warnings, flags, or restrictions — swift action there, while hateful content flowed under the loosened rules. (source)

analysis of 812,000+ anti-vaccine posts found 65% of anti-vaccine content traceable to 12 individuals — 73% on Facebook alone — and platforms failed to act on 95% of reported COVID/vaccine misinformation despite repeated terms-of-service violations. (source)

during the Delta surge — cases up 70% week-over-week, deaths up 26%, concentrated in low-vaccination areas — the White House publicly pressured Facebook while flagging posts to the company. Biden later clarified: "These 12 people are out there giving misinformation... It's killing people." (source)

MIT/UPenn researchers extrapolated survey experiments to all 13,206 vaccine-related URLs popular on Facebook in Jan–Mar 2021: flagged misinformation got 8.7M views — just 0.3% of 2.7B vaccine-related link views — while unflagged vaccine-skeptical content (often factually accurate but implying harm) was viewed hundreds of millions of times and reduced vaccination intentions 46× more than flagged misinformation (2.28 vs. 0.05 percentage points). A single vaccine-skeptical Chicago Tribune headline reached 50M+ users — six times the views of all flagged misinformation combined. Estimated ~3 million Americans potentially dissuaded from vaccinating. (source)

no documented individual death in 2024–2026 was found directly attributable to Facebook COVID misinformation — the harm evidence for the window is at the population (hesitancy/exposure) level.

originated in the private Facebook group "Springfield Ohio Crime and Information" (~8,000 members) — a claim that Haitians were "carving up" a resident's lost cat to eat, with no evidence and no eyewitness; the author later deleted the post and said she regretted sharing an unsupported rumor. It spread from the Facebook group to the presidential debate stage, where Trump repeated it on live TV (Sept 10, 2024) to ~67 million viewers. Consequences: 33 bomb threats forced evacuations of public buildings; state police deployed to every school; Haitian residents said they felt unsafe. City officials, the mayor, and Ohio's Republican governor all declared the claims false. (source)

the "Epping Says No" Facebook group (1,500 members), administered by three members of the Homeland Party — a splinter of the neo-Nazi group Patriotic Alternative — organized anti-asylum-hotel protests from July 13, 2025; the late-August bank holiday saw coordinated actions at 26 migrant hotels, triggered in part by unverified social-media rumors (e.g., that asylum seekers were being moved to a four-star London hotel). (source)

within hours of Axel Rudakubana murdering three children, false claims named the attacker a Muslim migrant/asylum seeker (he was born in Cardiff to Christian Rwandan parents); far-right organizers used the false name to mobilize a march that turned violent, and unrest spread to mosques and asylum-seeker hotels nationwide (~600 arrests). The false name spread primarily via X and Telegram — Facebook's documented role in the initial spread is thinner, though Full Fact documented false Rudakubana claims circulating in Facebook posts with thousands of reactions. (source source)

the UN Fact-Finding Mission (2018) said social media played a "determining role": "As far as the Myanmar situation is concerned, social media is Facebook, and Facebook is social media." 650,000+ Rohingya fled to Bangladesh amid military "clearance operations." Facebook's resourcing failure: ~2 Burmese-speaking moderators for 20M+ users while its recommendation systems fueled anti-Rohingya content. (source)

false Facebook rumors — including a viral video purporting to show a Muslim restaurateur admitting to mixing "sterilization pills" into Sinhalese-Buddhist customers' food — contributed to the worst anti-Muslim riots in the country's history: mosques, Muslim-owned shops, and homes burned across multiple towns; the government banned Facebook. One viral status read: "Kill all Muslims, do not spare even an infant, they are dogs." Facebook later apologized after an independent investigation found its role in instigating the violence and its failure to respond to nearly a decade of warnings. (source)

false child-abduction rumors spread via WhatsApp (Facebook-owned) and Facebook posts triggered mob lynchings — at least 27 people dead. (source)

on April 3, 2025, a Kenyan court ruled it has jurisdiction to hear a £2bn case alleging Meta fueled ethnic violence in Ethiopia. Lead claimant Abrham Meareg Amare says his father, Professor Meareg Amare Abreha, was murdered after Facebook posts targeted him as Tigrayan — publishing his photo, home address, and false claims he was aiding the Tigray People's Liberation Front. He was shot twice outside his home; armed militants blocked his family from helping for seven hours; he died in the street. The petition alleges Meta repeatedly failed to take down posts calling for violence, including posts promoting concentration camps and rape as a weapon of war. Separately, a May 2025 DAIR study (incl. Timnit Gebru) found platforms "failed to adequately moderate genocidal content" during the Tigray war; Facebook supported only 2 of Ethiopia's ~100 languages for integrity systems. (source source)

internal documents showed Facebook acted on only 3–5% of hate content and 0.6% of violence/incitement content; in one 30-day dataset, 98% of hate-speech removals were reactive (after user reports), only 2% proactive. Haugen testified Facebook misled the public about progress on hate speech and misinformation. (source source)

Meta has a culture of "see no evil, hear no evil"; safety features built in response to public outcry were "placebo... in name only to placate the press and regulators"; at "almost every turn the company chose policies that boosted usership and revenue over safety." (source)

submitted ads containing real-world hate speech inciting violence/genocide — 8 in Burmese (Myanmar): Meta accepted all 8. 12 in Amharic (Ethiopia): accepted all 12; after Meta admitted they "shouldn't have been approved," 2 more Amharic hate ads were submitted — accepted again. Meta refused to disclose moderator numbers for non-English countries. (source)

Announced April 6, 2024; from May 2024 Meta labeled video/audio/images "Made with AI" when industry-standard indicators (C2PA/IPTC metadata) detected or users self-disclosed; simultaneously stopped removing most manipulated media, moving to label-and-contextualize except high-risk deceit. In July 2024 the label was downgraded to a smaller "AI Info" badge after photographers complained real photos were mislabeled. (source source)

source Meta acknowledged "too much spammy content is crowding out authentic creators"; lowered reach for accounts posting long/distracting captions; aggressive action against spam networks coordinating fake engagement; took down 23M+ impersonating profiles in 2024; tested anonymous comment downvotes. Notably, the announcement made no mention of AI-generated slop (source).

H1 2025 — action on ~500,000 accounts for spammy behavior/fake engagement (demoted comments, reduced distribution, demonetization) and removal of ~10M profiles impersonating large creators; repeated reuse of others' videos/images/text without meaningful transformation → removal from monetization + reduced distribution; duplicate-video detection; guidance distinguishing reaction videos/trends (allowed) from repetitive reposting (penalized). (source source)

source "In 2025, we removed more than 20 million accounts impersonating large content producers"; clearer originality guidelines (creator-filmed = original; duplicative/minor-edit = unoriginal and deprioritized from Feed/Reels recommendations); content-protection tooling; appeals process.

ended the third-party fact-checking program (US, started 2016), replaced with X-style Community Notes (testing began March 18, 2025; built on X's open-source algorithm); kept fact-checking only for "illegal and high-severity violations"; Zuckerberg said Meta would "catch less bad stuff." Also announced more recommended civic/political content, reversing the 2021 reduction — i.e., more algorithmic recommendation surface for low-quality viral content. (source source)

"AI Info" depends on C2PA/IPTC metadata and uploader self-disclosure — the slop farms' images (Bing/Midjourney outputs reposted across fake accounts) typically carry neither. The Oversight Board ruled (2026, Scottish-politician deepfake case) that reliance on voluntary disclosure is ineffective against malicious actors and subtle badges insufficient. (source)

"Researchers found some pages disappeared after their findings were published. But new pages continue appearing just as quickly." ( — ibtimes.sg)

via its bonus program while claiming to fight it (§2).

in the April 2025 spam initiative; unoriginal-content rules target reposted content, while original-but-synthetic engagement bait is explicitly permitted ("We encourage creators to use AI tools…").

Meta shut down CrowdTangle on Aug 14, 2024; its replacement, the Meta Content Library, was found by a Proof News/Tow Center/Algorithmic Transparency Institute investigation to have fewer features and restricted access (academics/nonprofits only; journalists excluded). CrowdTangle's ex-CEO Brandon Silverman called the timing "incredibly irresponsible"; Mozilla + 140 civil-society groups protested. This blinds independent measurement of slop prevalence. (source source source)

After the Jan 2025 rollback, Meta reported low "prevalence" of violating content and ~50% fewer enforcement mistakes (Q1 2025 integrity report); analysts (Network of Centers/SMEX; OHPI) argue the bar was simply lowered — prevalence understates what users actually see, since proactive automated enforcement was cut (Meta reduced hate-speech removals ~79% per OHPI). (source source)

Meta publishes no slop-specific prevalence metric; the best evidence is researcher samples (120+ pages, hundreds of millions of interactions) and Meta's enforcement counts. CrowdTangle's shutdown makes independent longitudinal measurement harder.

The 500,000-accounts / 10M-profiles figures are confirmed by multiple secondary outlets but I did not open Meta's official post live; the adjacent official posts (April 2025 spam, March 2026 creators) were .

No public time-series data; researchers report new pages appear as fast as old ones are removed.

(2026) — no independent evaluation found in this research pass.