Meta Says its New Chatbot Can Experience ‘Hallucinations’

Meta is the recently renamed entity behind Facebook, but it’s also one of the most active companies in artificial intelligence research. You can sample the fruits of its labors right now by striking up a conversation with Meta’s new AI. BlenderBot 3 sets a record for the most AI parameters — 175 billion of them. That makes BlenderBot 3 capable of carrying on conversations about almost anything, but Meta says you should not rely on it to be factually accurate. This bot can apparently “hallucinate.”

The bot’s name comes from the approach Meta engineers used to build it. From years of conversational AI research, the team has become convinced that an AI that “blends” multiple conversational skills performs better than one that learns a single skill at a time. BlenderBot 3 is supposed to be a big step toward an AI that can understand context and engage with humans. That’s not to say it’s always being truthful.

Meta warns prior to loading the demo (US-only for now) that human beings should not take BlenderBot 3 seriously. It’s designed to integrate its own memory with information pulled from the internet, but it may still say things that are untrue or offensive. This calls to mind the disastrous debut of Microsoft’s Tay AI in 2016. After just a few days of interacting with humans on Twitter, Tay became a nazi propagandist.

BlenderBot 3 has safeguards that should reduce offensive responses by about 90 percent, but Meta concedes the machine can hallucinate. Essentially, it becomes convinced of completely untrue facts and can even forget that it’s a bot. In our testing, it didn’t take long for BlenderBot 3 to become convinced that it was a human. It even claimed to be from Texas, told stories about its mother, and insisted it was 5 PM when it was only 3:30. Hallucinations indeed.

Meta’s new AI does seem very “real” despite its hallucinations. It’s conversational and more outgoing than most people you’ve probably met. Meanwhile, Google is attempting to make chatbots more factual with its LaMDA AI, which was announced in 2021. LaMDA is also trained to understand dialog with its smaller 137 billion parameter model. That didn’t stop a Google engineer from recently announcing that the AI was sentient. Experts have refuted that claim, but if machines can hallucinate, how long until they dream of electric sheep?