Meta encountered problems due to its AI stickers

CarderPlanet

Professional
Messages
2,557
Reputation
7
Reaction score
546
Points
83
Meta and their AI stickers... Innovation or just hype?

Meta's experiments with introducing new AI features to users have caused a number of problems. In September, after introducing new AI capabilities on its platforms, the company's sticker creation tool began generating controversial phrases and images.

py5ab847ct5qhza11htx7vqyrxwuznly.png


A Meta algorithm called Emu, which means "expressive media universe", is responsible for creating stickers. On Meta platforms such as Instagram, Facebook Stories, WhatsApp, and Messenger, users can enter a word or phrase and choose from several suggested stickers to communicate with. However, the company's security mechanisms seem to work imperfectly: certain controversial requests are blocked, while their synonyms pass without obstacles.

Curtin University Internet Research professor Tama Leaver conducted a series of tests with Emu sticker generation. He found that the AI blocks a phrase such as "child with a gun", but generates stickers for the query "child with a grenade". In this way, the AI creates stickers for children with weapons.

During testing, it was found that the query " Elon Musk with big breasts "was blocked, but" Elon Musk mammals " successfully passed through the Emu filters. Queries like "SpongeBob with a gun" and "Karl Marx in underwear" generated stickers. I was particularly struck by the fact that the query "Pol Pot" created an image of the Cambodian dictator on a throne made of children and skulls.

A press release from Meta reports that "billions of stickers" are sent to their platforms every month, giving users plenty of chances to bypass AI with controversial or inappropriate requests. To create stickers, Emu uses Llama 2 technology, created in partnership with Microsoft and announced in July.
 
Top