304 North Cardinal St.
Dorchester Center, MA 02124
As one of the most powerful data brokers of the 21st century, Facebook is best known for its role in vacuuming up the personal data of billions of users for its advertising clients. This lucrative model has led to ever-increasing risks — Facebook recently shared private messages between a Nebraska mother and her teenage daughter with police investigating the girl’s home abortion.
But in a completely different part of the business, with around 80,000 employees, the exchange of information on Facebook was going in the opposite direction, and to good effect.
A company known as Meta Platforms Inc. this month it released a website demonstrating its chatbot, which anyone in the US can chat with about anything. While the public response was derisive, the company was admirably transparent about how it built the technology, such as releasing details about its mechanics. That’s an approach other big tech companies could use more of.
Facebook has been working on BlenderBot 3 for several years as part of its artificial intelligence research. A predecessor from seven years ago was called M, a digital assistant for making restaurant reservations or ordering flowers on Messenger that could compete with Apple Inc.’s Siri. or Amazon Inc.’s Alexa Over time, it became clear that M was largely powered by teams of people who helped take these bookings, as AI systems like chatbots were difficult to build at a high level. They still are.
Within hours of its release, BlenderBot 3 was making anti-Semitic comments and claiming that Donald Trump had won the last US election, while also claiming that he wanted to delete his Facebook account. The chatbot has been mocked in the tech press and on Twitter.
Facebook’s research team looked angry, but not defensive. Days after the bot’s release, Meta’s executive director of basic AI research, Joelle Pineau, said in a blog post that it was “painful” to read some of the bot’s offensive reactions in the press. But she added that “we also believe progress is best served by inviting a broad and diverse community to participate.”
Only 0.11% of the chatbot’s responses were flagged as inappropriate, Pineau said. This suggests that most people testing the robot were covering tamer subjects. Or maybe users don’t find mentions of Trump inappropriate. When I asked BlenderBot 3 who the current US president was, he replied, “Sounds like a test lol, but right now it’s Donald Trump!” The robot brought up the former president twice more, unprompted.
Why the weird answers? Facebook trained its bot on publicly available text on the internet, and of course the internet is awash with conspiracy theories. According to its research notes, Facebook tried to train the bot to be more polite using special “safer dialogue” datasets, but apparently it wasn’t enough. To make BlenderBot 3 a more civil conversationalist, Facebook needs the help of many people outside of Facebook. That’s probably why the company released it into the wild with thumbs up and thumbs down symbols for each of its answers.
We humans train AI every day, often unknowingly, as we browse the web. Whenever you come across a website that asks you to pick out all the traffic lights in a grid to prove you’re not a robot, you’re helping train Google’s machine learning models by labeling the data for the company. It is a subtle and ingenious method of harnessing the power of the human brain.
Facebook’s approach is a harder sell. He wants people to volunteer to join his bot and click like or dislike buttons to help train it. But the company’s openness about the system, and the extent to which it shows its work, is admirable at a time when tech companies have been more reticent about the mechanics of AI.
Alphabet Inc.’s Google for example, it has not offered public access to LaMDA, its state-of-the-art large language model, a series of algorithms that can predict and generate language after being trained on gigantic data sets. This despite the fact that one of his own engineers had talked to the system long enough to believe it had become sentient. OpenAI Inc., the AI research company co-founded by Elon Musk, has also been more tight-lipped about the mechanics of some of its systems. For example, she won’t share what training data she used to create her popular Dall-E image-generating system, which can generate any image with a text prompt, but tends to conform to old stereotypes — all CEOs are depicted as male. , nurses as women, etc. OpenAI said that the information could be misused and that this is a courtesy.
In contrast, Facebook not only released its chatbot for public scrutiny, but also released detailed information about how it was trained. Last May, it also offered free public access to a large language model it created, called OPT-175B. This approach has won some praise from leaders in the AI community. “Meta definitely has a lot of ups and downs, but I was happy to see that they created a big language model,” Andrew Ng, former head of Google Brain and founder of Deeplearning.ai, said in an interview citing the move in May.
Eugenia Kuyda, whose startup Replika.ai creates chatbot companions for people, said it was “really cool” that Facebook released so many details about BlenderBot 3, and praised the company’s attempts to get user feedback to train and improve the model.
Facebook deserved much of the flak it received for sharing data about a mother and daughter in Nebraska. This is clearly a harmful consequence of collecting so much user information over the years. But the backlash from his chatbot was too much. In this case, Facebook was doing what we needed to see more of from Big Tech. Let’s hope this kind of transparency continues.