GitHub

Facebook has long been filled with feeds of clickbait articles. Now Meta creates its own clickbait articles with AI.

Meta created its own AI-generated clickbait feed

Meta said it would remove the feature after The Verge asked about it.

The standalone Meta AI app now has a “For You” section that populates a list of clickbait-style stories for you to read. But the topics, images, and text are all AI-generated – and as questionable as one might expect from AI-created works.

The Meta AI app first launched in April 2025 and focused on a public “Discover” feed that showed AI-generated images and conversations from other users (who often seemed unaware that they were being made public). All that is gone. The app now has a standard chatbot interface, as well as a For You Featured page for at least a few months, displaying a feed of suggested article prompts that, when tapped, generate entire "stories."

Targeted at me, a London-based journalist, the prompts were aggressively British, involving topics like tea, manners, pubs, the royal family, football – sorry, football – and, of course, the art of queuing. Suggested stories included "A royal butler finally settled the milk-first debate" (tea comes first, apparently), "The psychology of joining a queue without knowing why", "The anatomy of the devastating British tut" and "Inside the extreme sport of visiting every British pub." » Some made even less sense, like "When a Little Pickle Means Total Disaster."

My colleague, meanwhile, seems to have been placed firmly in the luxury watch enthusiast category by the algorithm. His feed suggested stories titled “My Fake Rolex Experience” and “The Brutal Math Behind the Rolex Waitlist Illusion.”

The AI-generated text reads like bloated filler, offering little substance beyond repeatedly restating the premise of the prompt. Sourcing was also non-existent.

I tried to find the origin of these “stories”. The history of the royal butler's tea party appears to date back to a 2018 BBC Three comedy series called Miss Holland, which follows a fictional beauty queen from a small Dutch town as she travels to Britain and learns "how to be posh and posh" from real former royal butler Grant Harrold. The "Rolex Experience" story, meanwhile, seemed to be a complete fabrication, generated in our chat box as an unsigned first-person narrative, after a bit of the usual whirring that happens when a chatbot generates. Other stories relied on vague references to anonymous experts or fictitious research.

When I tapped the same cards more than once, the generated stories stayed within the rough confines of the prompt and were all clearly versions of the same thing, but slightly different. Typing the same title in a separate discussion produced a completely different response. The clearest giveaway came from my chat history. It showed the hidden and suggested prompts that were supposed to trigger article generation. One of them began:

"You are a helpful conversational assistant. The user responds to a proactive flow card shown to them. The context of the card below provides general information about what prompted the user's message," followed by what appear to be references to internal instructions, information, and metadata.

The articles were accompanied by images. Many of them were harmless – a bland mush of cartoon characters, scenery and food. But some depicted real people, including public figures, and were riddled with errors. “Who is really paying for the royal family in 2026? featured two queens Elizabeth II, despite her death several years previously and her existence as one person.

Surrounding the queen clones were people who appeared to be approximations of other members of the royal family: a vaguely Princess Kate face on the left, a strange attempt at Prince William in the back, and a sort of King Charles in the middle who looked exaggeratedly like his late father. Other images contained usual AI cues, like impossible hands and bodies leaning at unnatural angles. One image turned out to be a GIF of an elderly couple dancing and doing arm movements that no human body could do.

It was unclear whether the app should be able to generate AI images of real people according to Meta's rather opaque rules, but it was. The company previously said it wants "people to know when they see posts created with AI" and that it automatically adds labels to some user-generated content when AI is detected. Despite this, there was no obvious indication or label in the feed or articles that any content was AI-generated.

Meta declined to answer many of my questions about the purpose of the feature, whether the company considers news or fiction, what safeguards are in place, and whether images of real people and public figures comply with its own AI content policies.

"The goal is to suggest what's most relevant to you, like fitness tips, meal plans, or other information, before you even have to ask."

“We're testing a daily feed that proactively shares tips, content and recommendations tailored to your interests,” Meta spokesperson Tracy Clayton said in a brief statement. "The goal is to suggest what's most relevant to you, like fitness tips, meal plans, or other information, before you even have to ask."

Clayton then sent a nearly identical "updated" statement, mysteriously removing the word "proactively."

A third statement from Clayton followed later in the day: "This was a test for a limited number of users and will be deprecated. Meta has no plans to move forward with this feature."

This leaves me with additional questions. How was this test limited if, besides me, at least three of my colleagues at The Verge had access to the same AI clickbait functionality? What did “proactively” mean? And of course, who asked all this in the first place?