Meta’s AI-Generated Stickers Spark Controversy
Just a few weeks after Meta CEO Mark Zuckerberg announced the company’s upcoming suite of AI features, they seem to already be facing difficulties with its new Generative AI stickers.
Users have started to complain that Meta’s security filters aren’t really catching everything, allowing certain objectionable AI-generated content to bypass, including Disney’s Mickey Mouse smoking a marijuana blunt or Winnie the Pooh holding a rifle.
The tool, which was announced at the end of September during the annual Meta Connect conference, allows users to edit images and co-create them with friends on Instagram.
Powered by Meta’s Llama 2 model, a rival to OpenAI’s ChatGPT, and Meta’s new image generation model, Emu, the AI-generated sticker tool was touted to produce “multiple unique, high-quality stickers in seconds” using textual prompts. Its recent debut at Meta’s Connect event brought forth an unexpected array of dubious creations by early testers.
While Meta had implemented certain safeguards, including blocking specific words from being used as prompts, innovative users quickly discovered workarounds. Typos or alternate descriptions of restricted terms bypassed these controls. Troublingly, even seemingly innocent prompts like “World Trade Center” led to problematic imagery without any added descriptors.
Notably, user @Pioldes highlighted the tool’s potential pitfalls, generating inappropriate stickers such as child soldiers, comically risqué versions of popular characters, and even questionable depictions of historical figures. But don’t worry, there are even parody stickers of X CEO Elon Musk floating around.
The new AI tool will begin rolling out to all WhatsApp, Messenger, IG, and Facebook users throughout the month of October.
Meta’s recent experimentation with an AI-generated sticker tool reflects the broader challenges faced by tech companies when implementing AI solutions, especially those who are keen on just putting something out into the public eye, with little to no thought or concern for the consequences.
In training AI models on text prompts, models like Llama 2 or OpenAI’s ChatGPT should be able to detect “patterns” in the training data that could reasonably lead to questionable content being generated.
Prior to rolling out these types of tools, public consumption is an irreversible phenomenon once a product is pushed out, begging the question of why tools like Meta’s AI Stickers were even released for public use this early.
It’s Not IF, It’s When…
No different than in the world of cybersecurity, it’s not if abuse will happen, it’s WHEN. Users will always find new, innovative ways to misuse or exploit vulnerabilities in these tools to push the bounds of curiosity and creativity to their liking.
While it’s challenging for tech companies to anticipate every potential case of misuse, companies like Meta have the bandwidth and resources to be able to invest the time into exploring workarounds to these potential vulnerabilities.
Rather than rolling out to the entire public, it’s crucial to control the initial audience that the tool is pushed out to, because once it’s out there, it’s hard to undo and people are already coming up with their own coding to override the intended functionality of the particular AI tool. In other words, R&D isn’t something to gloss over.
And lastly, reputation is everything. Right now, big tech is under major scrutiny for putting such dangerous systems out to the general public, and then realizing their mistakes after it’s too late.
While Meta’s new AI-generated stickers may not stick out as a huge issue, it represent a cautionary tale and call to action for companies to think before they release, and ignore this “race to the finish” in the ongoing AI battle.
Editor’s note: This article was written by an nft now staff member in collaboration with OpenAI’s GPT-4.
AI Resources
Join the Now Network for access to premium AI resources, tools and prompts curated by our community of builders, creators and collectors.