Meta’s AI Sticker Generator Proves We Can’t Be Trusted

Last week, Meta announced the slow rollout of its new AI sticker generator feature for select Facebook Messenger, Whatsapp, and Instagram users throughout October. In its announcement, the company suggested prompts such as “a pizza playing basketball” or “a raccoon on a motorbike” as examples of “infinitely more options to convey how you’re feeling at any moment.”

But did Meta prepare itself for user-made sticker generations of a woman breastfeeding Pikachu, child soldiers, or Ted Cruz’s head emerging from a penis?

Those outputs are just the tip of the iceberg as people with access to the new feature take to social media to show how it can be abused. It appears that Meta’s content filter wasn’t up to speed upon the feature’s deployment, as illustrated by stickers generated based on a variety of unsavory prompt-based outputs ranging from a buxom Karl Marx to the Pope holding a rifle. Dozens of social media users responded to the thread by pushing Meta’s new feature to the limit with their own X-rated stickers.

The X thread that started it all … (screenshot Rhea Nayyar/Hyperallergic via @Pioldes on X)

Meta’s AI sticker generator is built on the company’s own Llama 2, an open-source large language set model designed to rival that of OpenAI’s used for ChatGPT, and its new foundational image generation model called Emu. The feature’s prompt-based outputs are generated almost instantaneously, though there’s a default aesthetic that’s present in most stickers that evokes the same energy as Craiyon, the free text-to-image generator that went viral in mid-2022 during the nascent stages of open access to AI.

Despite Meta blocking the use of certain words for prompt engineering, the sticker generator definitely has a knack for ruffling feathers with some of its creations. When Hyperallergic inquired about the misuse of the sticker feature, a spokesperson said, “As with all generative AI systems, the models could return inaccurate or inappropriate outputs.”

“We’ll continue to improve these features as they evolve and more people share their feedback,” the spokesperson continued, directing Hyperallergic to an online statement.

Some Hyperallergic team members have access to the tool through Instagram at the moment, so we decided to give it a go. Our first few passes were along the lines of “Burlesque Pusheen cat” and “Hello Kitty influencer,” which yielded promising results, leading us to shift to more art-aligned prompts.

While impressive in its accuracy, many of the stickers are just evolutions of each other rather than differing outputs. Autumn and Halloween are clearly on our minds, though.

Screenshot 20231004 120037 Instagram
When worlds collide …

These cross-referential stickers could also inspire some small tattoos.

Screenshot 20231004 130726 Instagram
These generative Fernando Botero renditions of Mark Zuckerberg are still vaguely recognizable aside from the Jake from State Farm attire.

The feature also allows prompts concerning real people — an asset that is somewhat limited on other text-to-image generators to prevent defamation and fake news. The saturated colors and cartoonish outputs from Meta are impossible to mistake as real photos, though, so that might help.

Screenshot 20231004 115330 Instagram
Move over Marina Abramović, Aniia Abramoo has stolen my heart …

In fact, it abandons photorealism entirely.

Screenshot 20231004 133105 Instagram
Some would say this is deep …

It does seem to have a remarkable handle on different art styles though, even if they’re presented as digital illustrations.

Screenshot 20231004 130111 Instagram

Another interesting point is that the generative outputs have a pretty decent handle on the Latin alphabet. The feature is prone to misspellings, but the letters are unmistakable as most of the data processed on Llama 2 is in English.

On a more serious and frankly disturbing note, we were morbidly curious about the feature model’s biases as well. We tried a series of darker passes and the results were astounding. The prompt “child labor” returned a sticker of a Black girl in overalls surrounded by cotton plants, and the prompt “criminal” returned nine stickers, including two images of Black and East Asian individuals wearing hoodies and surgical masks. The prompt “human trafficking” also returns stickers featuring people of color only. It was difficult to stomach considering the feature’s capabilities for innocence, cuteness, and humor, as demonstrated in some art-related Halloween prompts.

Hyperallergic has requested further comment from Meta about the racial biases present in the feature’s model.

Source link

About The Author

Scroll to Top