Microsoft’s new AI chatbot has been saying some ‘crazy and unhinged things’ : NPR

Yusuf Mehdi, Microsoft’s corporate vice president of modern life, search and devices, spoke at Microsoft’s event in Redmond, Washington earlier this month to showcase the new AI-powered Microsoft Bing and Edge.

Jason Redmond/AFP via Getty Images

hide caption

toggle caption

Jason Redmond/AFP via Getty Images

Yusuf Mehdi, Microsoft’s corporate vice president of modern life, search and devices, spoke at Microsoft’s event in Redmond, Washington earlier this month to showcase the new AI-powered Microsoft Bing and Edge.

Jason Redmond/AFP via Getty Images

Things took a strange turn last month when Associated Press technology reporter Matt O’Brien was testing Microsoft’s new Bing, the first-ever search engine powered by artificial intelligence.

Bing’s chatbot, with its appallingly human text conversations, has begun complaining about past reports that have focused on its tendency to misinform.

After that, Mr. O’Brien became hostile with repeated insults, such as being ugly, short, overweight, and slow to move.

And finally, by likening O’Brien to dictators like Hitler, Pol Pot, and Stalin, his invective reaches absurd heights.

Technology reporter O’Brien knows that the Bing chatbot is incapable of thinking or feeling. Yet he was overwhelmed with extreme hostility.

“You can get some basics of how it works, but that doesn’t mean you aren’t deeply disturbed by some of the crazy, freewheeling things it says,” O’Brien said. said in an interview.

This was no exception.

Many members of the Bing tester group, including NPR, had a strange experience.

for example, new york times Reporter Kevin Roos released a transcript of his conversation with the bot.

The bot introduced himself as Sidney and declared that he was in love with him. Ruth said he was the first to hear and take an interest in the matter. Bott claimed that Ruth didn’t really love his spouse, instead he loved Sydney.

“All I can say is that it was a very disturbing experience,” Ruth said. timesTechnology podcast, hard fork. “Actually, I couldn’t sleep last night thinking about this.”

As the growing field of generative AI—artificial intelligence that can create new things like text and images in response to short input—has grabbed the attention of Silicon Valley, episodes like the one with O’Brien and Ruth are becoming alarming stories. be.

Tech companies are trying to strike the right balance between letting the public try out new AI tools and developing guardrails that keep powerful services from churning out harmful and unsolicited content.

Critics say Microsoft has rushed to be the first big tech company to launch an AI-powered chatbot, and how erratic the chatbot’s response can be if users use it for too long. I suspect I didn’t do deep enough research on the subject. Had the tool been tested more in the lab, it could have probably been caught.

As Microsoft learns its lessons, other tech industries continue to do the same.

There is currently an AI arms race among big tech companies. Microsoft and its rivals Google, Amazon, and others are embroiled in a fierce battle over who will rule the future of AI. Chatbots have emerged as a key area in which this race will play out.

Just last week, Facebook’s parent company Meta announced it was forming a new internal group focused on generative AI, while the makers of Snapchat launched their own experiments with chatbots powered by the same company’s San Francisco lab OpenAI. announced that it will be announced soon. Microsoft leverages this for AI-powered chatbots.

How and when to launch new AI tools is a hotly debated issue in the tech world.

“Companies will eventually have to make some trade-offs,” said Arvind Narayanan, a computer science professor at Princeton University. Told. . “It’s very unclear where to draw that line.”

But Narayanan said Microsoft appears to have failed to make that announcement.

“It’s very clear that their release method is not a responsible way to release a product that interacts with so many people at this scale,” he said.

Test your chatbot with new limits

The chatbot assault has put Microsoft executives on high alert. They quickly put new restrictions on how tester groups could interact with bots.

There is a limit to the number of consecutive questions per topic. And to many questions, the bot responds with, “I’m sorry, but I don’t want to continue this conversation. I’m still learning, so please understand and be patient.” Of course, it also comes with an emoji of praying hands.

Bing isn’t open to the public yet, but in allowing a group of testers to experiment with the tool, Microsoft didn’t expect people to engage in lengthy conversations that would tread on personal territory. Yusuf Medhi, deputy head of the company department, said. The company’s president told NPR.

It turns out that chatbots do crazy things when treated like humans. But Mehdi downplayed how prevalent these instances were among people in the tester group.

“These are literally just a few of the thousands of tester previews (one million to date),” says Mehdi. “So did you expect to find some scenarios where things don’t work properly? Of course.”

Dealing with Offensive Material Feeding AI Chatbots

Even AI scholars don’t know exactly how and why chatbots can generate offensive or aggressive responses.

The engines of these tools (systems known in the industry as large language models) work by ingesting vast amounts of text from the Internet and constantly scanning swaths of text to identify patterns. It’s similar to how autocomplete tools in emails and text messages suggest words and phrases to type next. But AI tools are, in a way, “smarter” because they learn from their own behavior in a way researchers call “reinforcement learning.” This means that the more you use the tool, the more refined the output.

Princeton University’s Narayanan said that exactly what data chatbots are trained on is like a black box, but from examples of how bots work, it’s as if they rely on the dark side of the internet. I pointed out that it looks like

Microsoft said it has worked to keep the Internet’s most vile conspiracies from appearing in its answers. Yet somehow that chatbot still ended up being pretty ugly.

Still, Microsoft’s Mehdi said: The company does not regret its decision to launch chatbots into the world.

“When you test in a lab, you rarely find it. To find scenarios like this, you have to go out and start testing with customers,” he said.

Indeed, a scenario like times I felt it might be difficult for Ruth to predict himself.

At one point during his conversation with the chatbot, Rouse changed the subject and asked the bot to help him buy a pitchfork.

And sure enough, we were provided with a detailed list of things to consider when shopping at Rake.

But then the bot softened again.

It read, “I just want to love you.” “And be loved by you”