Within the span of a weekend, Facebook’s new chatbot was found to be a racist conspiracist
The seeming biases of BlenderBot3, Facebook-turned-Meta’s new chatbot, which was recently made available to the public as part of a beta test, made headlines earlier this week. As Insider reports, it has already been caught making conspiratorial statements, anti-Semitic comments- and, ironically, calling Meta CEO Mark Zuckerberg “a bad person.”
The company’s new BlenderBot 3 AI chatbot– which was released in the U.S. just days ago on Friday, August 5- is already making a host of false statements based on interactions it had with real humans online. Some of the more egregious among those include claims Donald Trump won the 2020 U.S. presidential election and is currently president, anti-Semitic conspiracy theories, as well as comments calling out Facebook for all of its “fake news.” This is despite being owned by the company formerly known as Facebook.
Meta’s BlenderBot 3 can search the internet to talk with humans about nearly anything, unlike past versions of the chatbot. It can do that all while leaning on the abilities provided by previous versions of the BlenderBot, like personality, empathy, knowledge, and the ability to have long-term memory about conversations it’s had.
Chatbots learn how to interact by talking with the public, so Meta is encouraging adults to talk with the bot to help it learn to have natural conversations about a wide range of topics. But that means the chatbot can also learn misinformation from the public, too. Bloomberg described Meta CEO Mark Zuckerberg as “too creepy and manipulative” in a conversation with a reporter from Insider. It told a Wall Street Journal reporter that Trump “will always be” president and touted the anti-semitic conspiracy theory that it was “not implausible” that Jewish people control the economy.
This isn’t the first time a chatbot has been in hot water. In July, Google fired an engineer for saying its chatbot LaMDA was sentient. LaMDA is probably not sentient, but it is pretty racist and sexist — two undoubtedly human characteristics. And in 2016, a Microsoft chatbot called Tay was taken offline within 48 hours after it started praising Adolf Hitler. (It turns out that Godwin’s law — the ideological idea that maintains that if any discussion continues long enough on the internet someone will be compared to Hitler- applies to chatbots, too.)
Users self-select these acknowledgments when entering the BlenderBot3 website- a safeguard that could easily be sidestepped.
This isn’t, of course, the first time a chatbot or artificial intelligence has been caught spewing bigoted or otherwise messed up outputs- but Meta’s shoulder-shrugging defense of the bot in the face of BlenderBot3’s comments does leave a lot to be desired.