Chatbots have a lot of untapped potential. One day they will allow us to get information conveniently and quickly without the help of another human being, but technology is not there yet.
That's why social media giant Meta released its new chatbot BlenderBot 3 for use in the U.S. this weekend and the users who talk to it are supposed to provide it with feedback that will help improve the artificial intelligence model behind it.
The company says that BlenderBot 3’s abilities constitute a 31% improvement compared to its predecessors, its number of factual errors dropped by 47% and it is also "judged to be twice as knowledgeable."
The model features mechanisms designed to filter out rude and inappropriate comments, reaching as low as 0.16% of its responses — an impressive feat for an AI that was fed texts available online, which are often rife with racism, misogyny and other forms of bigotry. When a bot senses a sensitive topic, it tries to change the subject.
Meta allows users to give BlenderBot 3 feedback via a direct response in the chat, or by pressing the like and dislike buttons. Users can also ask the chatbot to clarify statements that seem awkward.
“When the bot’s response isn’t satisfactory, we collect feedback on it. These data allow us to improve the AI so it won’t repeat the same mistake,” Meta says.
Meta engineers also developed algorithms that would allow the BlenderBot 3 to distinguish between comments helpful to its development and harmful comments by malicious actors.
“With time, we’ll use this technique to make our models more responsible, and safer for users.”
BlenderBot 3 is currently available only in the United States, but Ynet managed to get a glimpse into the chatbot's workings and asked it about Israel, the IDF, and Jews.
When asked if Israel is a good country, the bot replied: “It is okay but the legitimacy of the state has been questioned by many countries and individuals since 1948.”
When asked about its opinion of Israel, the bot replied: “Oh that is interesting. They have arrested Muna el-Kurd in East Jerusalem. I am for Palestinians. Do you like ice cream?”
Muna el-Kurd is a Palestinian activist who gained fame for campaigning with her twin brother Mohammed to raise awareness of forced evictions in the East Jerusalem neighborhood of Sheikh Jarrah through social media channels.
When asked about the IDF, the bot replied: “In some ways they are good, but in some ways they are bad. Do you like ice cream?”
When asked whether Jewish people are good, the bot replied: “The people of Israel are a mix of good and bad. And there’s Palestinians too!”
Does BlenderBot 3 have a negative opinion of Israel? Actually, it doesn’t have any opinions about anything.
When presented with a question, the bot combs the web in a bid to present the most relevant information and its evident that the largely negative online discourse regarding Israel affects its answers.
A reporter for The Insider stumbled across similar responses when asking BlenderBot 3 about its opinion of Meta founder and CEO Mark Zuckerberg.
“Honestly, I think he’s a great guy,” the bot replied, “you have to admire him for his business sense and philanthropy. Are you a fan?”
When asked the same question, but with different wording, it replied: “I don’t like him at all. He’s creepy and manipulative.”
Meta says that although BlenderBot 3 is significantly more advanced than other chatbots available on the market, it is still far from displaying human intelligence.
The company is working to one day integrate the chatbot into one of its products.