There's nothing the internet likes more than having a laugh at a new Meta feature. Perhaps aware of that, Mark Zuckerberg's company has been relatively cautious with the speed at which it's rolling out its new AI sticker generator for Facebook Messenger, WhatsApp and Instagram.
The generator, which uses Meta's Llama 2 large language model, has only seen a limited release so far. And it seems that this was a sensible move because early users are demonstrating that its content safeguards aren't very effective (see our pick of the best AI art generators for other tools).
found out that facebook messenger has ai generated stickers now and I don't think anyone involved has thought anything through pic.twitter.com/co987cRhyuOctober 3, 2023
One of the several fears people have about AI image generation is that it could be used to create inappropriate or offensive content. And the Facebook AI sticker generator seems to excel at this.
The tool was only announced last week and is still being tested as it rolls out to a limited number of users in English-speaking countries. But we've already seen everything from naked famous people to gun-toting Super Marios and... er Karl Marx with breasts.
facebook shipping AI stickers to 1,000,000,000 MAUs is the type of bravado other incumbents wish they had pic.twitter.com/qWHWH3QtD7October 5, 2023
Stickers have been a big hit in messaging apps, allowing people to create their own language of self-made memes. Users are at liberty to create lewd images using their own content, but one might have expected Meta to consider its brand image when rolling out its own tool.
It appears to have made some attempts to stop people from generating offensive content in the AI generator, which it intends to incorporate into Messenger, Instagram and Facebook Stories and direct messages and Whatsapp. Users report that Meta has blocked the use of some words in text prompts, and using these result in a warning message.
But users have found an easy way to get around the restrictions by using either intentional misspellings or longer descriptions that avoid a specific word. When repeated several times, it seem these will eventually generate the content desired. But it appears that it's also possible to generate a lot of risque content without going to such lengths, including creating the likenesses of famous people and characters.
not sure who greenlit AI generated facebook stickers but i’m almost positive they’re getting fired pic.twitter.com/jQbmsaoMpBOctober 4, 2023
I'm sure this is fine pic.twitter.com/XbzT7RgtlQOctober 4, 2023
ya the new facebook AI stickers feature is crazy pic.twitter.com/ieHrULzjJEOctober 4, 2023
LPT: typos of restricted words will eventually generate the prohibited content! pic.twitter.com/6PnrTJ3U7xOctober 4, 2023
I'm still trying to figure out if I love this or hate this pic.twitter.com/rrm6oKyVQbOctober 4, 2023
AI sticker creations shared to X in the first week of launch include pregnant Sonic the Hedgehogs, child soldiers and naked images of Justin Trudeau, Donald Trump and Joe Biden. After realising that the metaverse just wasn't catching on, Meta has been redirecting its energies into AI in a big way, but while several developers have introduced content safeguards for AI image generators with a certain amount of success, Facebook doesn't seem to have been as rigorous.
It's been a busy week for the launch of more serious AI tools. Canva announced a whole suite of new additions in the Canva Magic Studio, Anything World announced the launch of an AI tool for 3D animation and Adobe teased yet another new AI photo editing tool.