Meta made its own AI-generated clickbait news feed

Facebook has agelong been filled pinch feeds of clickbait articles. Now, Meta is making its ain clickbait articles pinch AI.

The standalone Meta AI app now has a “For You” conception that populates a database of clickbait-style stories for you to read. But the topics, images, and matter are each AI-generated — and arsenic questionable arsenic you’d expect from AI-created works.

The Meta AI app first launched successful April 2025 pinch its attraction connected a nationalist “Discover” provender that showed AI-generated images and conversations from different users (who often seemed unaware that they were being made public). That’s each disappeared. The app now has a modular chatbot interface, positive a For You page that’s been coming for astatine slightest a fewer months, displaying a watercourse of suggested article prompts that, erstwhile tapped, make full “stories.”

When targeting me, a newsman based successful London, the prompts were aggressively British, involving topics for illustration tea, manners, pubs, royals, shot — sorry, shot — and, naturally, the creation of queuing. Suggested stories included “A royal butler yet settled the beverage first debate” (the beverage goes first, apparently), “The psychology of joining a queue without knowing why,” “The anatomy of the devastating British tut,” and “Inside the utmost athletics of visiting each UK pub.” Some made moreover little sense, for illustration “When a spot of a pickle intends full disaster.”

My colleague, meanwhile, appears to person been placed firmly wrong the luxury watch aficionado bracket by the algorithm. His provender suggested stories called “My clone Rolex experiment” and “The sadistic mathematics down the Rolex waitlist illusion.”

The AI-generated matter publication for illustration puffy filler, offering small constituent beyond many times restating the premise of the prompt. Sourcing was besides nonexistent.

I tried to way down wherever these “stories” whitethorn person originated. The royal butler beverage communicative appears to trace backmost to a 2018 BBC Three drama bid called Miss Holland, which follows a fictional beauty queen from a mini Dutch municipality arsenic she travels to Britain and learns “how to beryllium posh and classy” from existent erstwhile royal butler Grant Harrold. The “Rolex experiment” story, meanwhile, appeared to beryllium a complete fabrication, generated successful our chat container arsenic a first-person communicative without a byline, aft a spot of accustomed whirring that happens erstwhile a chatbot is generating. Other stories leaned connected vague references to unnamed experts aliases fictional research.

When I tapped the aforesaid cards much than once, the generated stories stayed wrong the unsmooth bounds of the punctual and each were intelligibly versions of the aforesaid thing, but somewhat different. Typing the aforesaid header into a abstracted chat produced a wholly different response. The clearest giveaway came from my chat history. It showed the hidden, suggested prompts that were expected to trigger the procreation of articles. One began:

“You are a adjuvant conversational assistant. The personification is responding to a proactive provender paper that was shown to them. The paper discourse beneath provides inheritance connected what prompted the user’s message,” followed by what appeared to beryllium references to soul instructions, information, and metadata.

The articles had images attached. A batch of these were harmless — bland mush of cartoony people, landscapes, and food. But immoderate depicted existent people, including nationalist figures, and were riddled pinch errors. “Who really pays for the royal family successful 2026?” featured 2 Queen Elizabeth IIs, contempt her decease respective years anterior and her beingness arsenic only 1 person.

Around the Queen clones were group who seemed to beryllium approximations of different royals: a vaguely Princess Kate-ish look to the left, a unusual effort astatine Prince William astatine the back, and a sort-of King Charles successful the mediate who bore an exaggerated resemblance to his precocious father. Other images had accustomed AI tells for illustration intolerable hands and bodies leaning astatine unnatural angles. One image really turned retired to beryllium a GIF of an older mates dancing and making limb movements nary quality assemblage could make.

It wasn’t clear whether the app should beryllium capable to make AI images of existent group successful accordance pinch Meta’s own, alternatively opaque rules, but it was. The institution has antecedently said it wants “people to cognize erstwhile they spot posts that person been made pinch AI” and that it automatically adds labels to immoderate user-generated content erstwhile AI is detected. Despite this, location was nary evident denotation aliases explanation successful the provender aliases articles that immoderate worldly was AI-generated.

Meta declined to reply galore of my questions astir the feature’s purpose, whether the institution considers the output news aliases fiction, what safeguards are successful place, and whether images of existent group and nationalist figures comply pinch its ain AI-content policies.

“We’re testing a regular provender that proactively shares tips, content, and recommendations tailored to your interests,” Meta spokesperson Tracy Clayton said successful a little statement. “The extremity is to propose what’s astir applicable to you – specified arsenic fittingness advice, repast plans, aliases different insights – earlier you moreover person to ask.”

Clayton later sent a astir identical “updated” statement, mysteriously removing the connection “proactively.”

A 3rd connection from Clayton followed later successful the day: “This was a trial for a constricted number of users and it will beryllium deprecated. Meta has nary plans to move guardant pinch this feature.”

This leaves maine pinch further questions. How was this trial constricted if, too me, astatine slightest 3 of my colleagues astatine The Verge had entree to the aforesaid characteristic serving AI clickbait? What did “proactively” moreover mean? And, of course, who asked for immoderate of this successful the first place?