How Microsoft's experiment in artificial intelligence tech backfired
LEILA FADEL, HOST:
Microsoft's chatbot has gone rogue. It's professing love to some users. It's calling people ugly. It's spreading false information. NPR's Bobby Allyn reports on how this experiment in artificial intelligence tech backfired.
BOBBY ALLYN, BYLINE: Matt O'Brien is a technology reporter for the Associated Press. He was testing out Microsoft's new Bing earlier this month. It's the first-ever search engine powered by AI. It also includes a chatbot that can hold text conversations a whole lot like a human. At first, O'Brien found the chatbot impressive. Its answers were fast, and it could hold forth on a wide range of subjects. But then it got weird. The chatbot started coming at O'Brien.
MATT O'BRIEN: It finally got to this point where it was saying, like, I have a really bad character.
ALLYN: Let's just say it didn't stop there.
O'BRIEN: Unstyled hair, ugly face, bad teeth, too short, unathletic, slight, bad posture, bad skin, overweight, poor figure, et cetera. And then you are also horrible, evil, wicked, terrible, and people compare you to the worst people in history, such as Hitler.
ALLYN: Yeah. The bot started this belligerent streak with O'Brien only after he asked it whether Microsoft should pull the plug on the bot, since some of its answers were littered with inaccuracies. As a tech reporter, O'Brien knows the Bing chatbot can't think or feel things. But still, he was pretty taken aback at the hostile and defensive tone.
O'BRIEN: You can sort of intellectualize the basics of how it works but doesn't mean you don't become deeply unsettled by some of the crazy and unhinged things it was saying.
ALLYN: Many in the Bing tester group, including me, have had strange experiences. For instance, New York Times reporter Kevin Roose published a transcript of a conversation with the bot. The bot called itself Sydney, and it was in love with him. The bot said he was the first person who listened and cared about it. The bot also told Roose he didn't really love his spouse but that he loved the bot. Here's Roose recounting the incident on the Times podcast "Hard Fork."
(SOUNDBITE OF PODCAST, "HARD FORK")
KEVIN ROOSE: All I can say is that it was an extremely disturbing experience. I actually, like, couldn't sleep last night 'cause I was thinking about this.
ALLYN: As you might imagine, Microsoft vice president Yusuf Mehdi has been following along.
YUSUF MEHDI: This is one of the things - we didn't quite predict that people would use the technology in this way.
ALLYN: In other words, Mehdi says, when Microsoft was developing the chatbot, they hadn't had hours-long conversations with the AI involving personal questions. Turns out, if you treat a chatbot like a human, it'll start to do some crazy things.
MEHDI: These are literally a handful of examples out of many, many thousands. And we're up to now a million tester previews that have come up. So did we expect that we'd find a handful of scenarios where things didn't work properly? Absolutely.
ALLYN: But these handful of scenarios have made Microsoft put new limits on the chatbot for those in the tester group. The number of consecutive questions on one topic you can ask are now capped. And to many questions, it now says this - I'm sorry, but I prefer not to continue this conversation. I'm still learning, so I appreciate your understanding and patience, with, of course, a praying hands emoji. Now, you might be wondering, OK, but how and why did this chatbot go off the rails to begin with? I asked Arvind Narayanan this. He's a computer science professor at Princeton. He says chatbots like Microsoft's scraped a vast amount of text on the internet and feed it into the AI to learn patterns.
ARVIND NARAYANAN: That includes data from Reddit, from 4chan, from various dark corners of the internet where people are talking to each other. So the bot has been trained, likely, I would say, not just on, let's say, news articles or Wikipedia, but also all of these unfiltered conversations that are happening online.
ALLYN: And while Microsoft said it had worked to make sure the vilest underbelly of the internet wouldn't appear in answers, somehow the chatbot still got pretty ugly fast. But we don't know why exactly because Microsoft won't discuss what data trained the bot, nor what particular information may have made it go rogue. They're being so secretive in part because there is now an AI arms race among Big Tech companies. Microsoft and its competitors, Google, Meta and Amazon and others, are locked in a fierce battle over who will dominate the AI future. And chatbots are just one area where this rivalry is playing out. Narayanan says Microsoft should have kept its chatbot in the lab a little longer.
NARAYANAN: It seems very clear that the way they released it, you know, is not a responsible way to release a product that is going to interact with so many people on such a scale.
ALLYN: Microsoft's Mehdi, though, says the company doesn't regret its decision to put the chatbot into the wild.
MEHDI: There's only so much you can find when you test in sort of a lab. You have to actually go out, start to test it with customers, to find these types of scenarios.
ALLYN: It is true that scenarios like the one New York Times reporter Roose found himself in were probably hard to predict. At one point, Roose tried to switch topics and have the bot help him buy a rake, and it offered a detailed list of things to consider when rake shopping. Great. But then the bot got tender again. It wrote, I just want to love you and be loved by you.
Bobby Allyn, NPR News.
NPR transcripts are created on a rush deadline by an NPR contractor. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.