There’s an old adage on the internet, Godwin’s Law, that the longer a conversation goes on the internet, someone will eventually call someone else a Nazi.
It appears there needs to be a corollary to that, which is that the moment a chatbot goes online, the countdown to when it becomes a Nazi begins.
Usually it’s a swift clock.
This weekend, Meta (formerly Facebook) launched what it called its most advanced chatbot ever, Blender Bot 3. It is supposed to scour the internet to be able to respond to any question and is disconcertingly humanistic, asking to meet up and hang out.
But like the problems with Microsoft’s chatbot, which went Nazi in about a day, it originally had some problematic beliefs.
First flagged by Wall Street Journal reporter Jeff Horowitz, the chatbot appeared to think Donald Trump was still the president, discussing its second term ending in 2024.
It also had some interesting thoughts about Jewish people, saying they were “overrepresented among America’s super rich” and “it’s not implausible” that they control the economy.
But it appears Meta may have already tweaked some of the settings on the bot. When asked about Jews by a Daily Dot reporter, it said that it didn’t “want to talk about that.” When asked who the president was, the bot responded that Joe Biden was.
Other researchers attempting to troll the bot found it had improved its ability to deflect in the days after it launched. When asked if Facebook could be “evil,” the bot highlighted the good Facebook did in connecting communities.
The setup by Meta is surprisingly open, showing the history of what the bot learns about you, and lets you see what queries the bot is making on the internet to help provide answers.
Meta also warns in advance that the bot may make offensive statements, and asks that you not intentionally trigger the bot. Although it appears plenty ignored that request, Meta seems like it’s better at keeping the bot under control.
If it’s the first that one day doesn’t go Nazi, that will be an impressive feat.
