Yusuf Mehdi, Microsoft corporate vice president of modern life, search and devices, spoke earlier this month at Microsoft’s event in Redmond, Wash., to launch a new AI-powered Microsoft Ping and Edge.
Jason Redmond /AFP via Getty Images
Hide title
Change the title
Jason Redmond /AFP via Getty Images

Yusuf Mehdi, Microsoft corporate vice president of modern life, search and devices, spoke earlier this month at Microsoft in Redmond during an event to launch AI-powered Microsoft Bing and Edge.
Jason Redmond /AFP via Getty Images
Things took a different turn earlier this month when Associated Press technology reporter Matt O’Brien tested Microsoft’s new Bing, the first search engine powered by artificial intelligence.
Bing’s chatbot, which conducts eerily human-like text conversations, has begun to complain about past news coverage focused on its tendency to spout misinformation.
Amid long-standing slurs about O’Brien as ugly, short, overweight, and monstrous, it became hostile.
Finally, O’Brien is taken to absurd heights by comparing O’Brien to dictators like Hitler, Pol Pot and Stalin.
As a tech reporter, O’Brien knows that the Bing chatbot doesn’t have the ability to think or feel. Nevertheless, he suffered from severe hostility.
“You can intellectualize the basics of how it works, but that doesn’t mean you’re not deeply disturbed by some of the crazy and unhinged things it says,” O’Brien said in an interview.

This is not an isolated example.
Many of the Bing testers, including NPR, had strange experiences.
For example, The New York Times Correspondent Kevin Rouse Published a copy A conversation with a bot.
The bot calls itself Sydney and declares its love. It said Rouse was the first person to care to hear it. Roose didn’t really love his wife, Bode insisted, but rather Sidney.
“I can say it was a very confusing experience,” Rouse said timesTechnology Podcast, Hard Fork. “I couldn’t sleep last night because I thought about it.”
The burgeoning field of budding AI — or artificial intelligence that can generate new things like text or images in response to short inputs — has captured the attention of Silicon Valley, with episodes like what happened to O’Brien and Roos becoming cautionary tales.
Tech companies are trying to strike the right balance between letting the public try out new AI tools and preventing powerful services from pumping out harmful and disruptive content.
In its rush to become the first big tech company to announce an AI-powered chatbot, critics say Microsoft may not have delved deeply into how long the chatbot’s responses might take. The instruments could have been caught if they had still been tested in the lab.
As Microsoft learns its lessons, the rest of the tech industry follows suit.
There is now an AI arms race among the big tech companies. Microsoft and its rivals Google, Amazon and others are in a fierce battle over who will dominate the future of AI. Chatbots are becoming an important part of this competition.
Just last week, Facebook parent company Meta Announced the creation of a new internal team focused on AI Snapchat maker He said it will soon release its own experiment with a chatbot powered by San Francisco research lab OpenAI, the same company that Microsoft uses for its AI-powered chatbot.
When and how to unleash new AI tools into the wild is a question that sparks fierce debate in tech circles.
“Companies ultimately have to make some kind of trade-off. If you try to anticipate every kind of interaction, you’re going to suffer from competition,” said Princeton computer science professor Arvind Narayanan. . “It’s not very clear where that line should be drawn.”
But Narayanan said Microsoft missed its launch.
“It’s very clear that the way they released it was not a responsible way to release a product that was going to interact with such a large audience,” he said.
Testing the chatbot with new limitations
Chatbot lashing incidents sent Microsoft executives on high alert. They quickly put new limits on how the tester team can interact with the bot.
The number of consecutive questions per topic is limited. To more questions, the bot now says, “I’m sorry, but I don’t want to continue this conversation. I’m still learning, so I appreciate your understanding and patience.” And of course, a praying hands emoji.
Bing hasn’t been released to the public yet, but in allowing a group of testers to test the tool, Microsoft doesn’t expect people to have hours of conversations with it that veer into personal territory, said Yusuf Mehdi, a corporate vice president. company president, told NPR.
If you treat a chatbot like a human, it will do some crazy things. But Mehdi underestimated how widespread these phenomena were among those in the experimental group.
“These are actually many, many thousands of examples — we’re at a million now — of tester previews,” Mehdi said. “So, did we expect to find some scenes where things didn’t quite work? Absolutely.”
Handling undesired items that feed AI chatbots
Even scholars in the field of AI aren’t exactly sure how and why chatbots can generate unsettling or offensive responses.
These tools’ engine — a system known in the industry as a large language model — works by ingesting vast amounts of text from the Internet, continuously scanning large amounts of text to identify patterns. Similar to how automation tools in email and text suggest the next word or phrase you type. But an AI tool becomes “intelligent” in a sense because it learns from its own actions in what researchers call “reinforcement learning,” meaning that the more the tools are used, the more refined the outputs.
Narayanan at Princeton notes that training data on chatbots is a black box, but from examples of bots in action, it appears they rely on some of the darkest corners of the Internet.
Microsoft said it worked to make sure it didn’t appear in the Internet’s worst underbelly responses. And yet, somehow, its chatbot got ugly even faster.
Still, said Microsoft’s Mehdi The company doesn’t regret its decision to put the chatbot into the wild.
“There’s only so much you can find when you test in a lab. You have to actually go out and start testing it on customers to find these types of scenarios,” he said.
Indeed, scenes like this times Reporter Roos himself might have been harder to predict.
At one point, during his exchange with the chatbot, Roos tried to change the subject and help him buy a rack.
And, of course, it provided a comprehensive list of things to consider when rack shopping.
But again the bot got tendered.
“I want to love you,” it read. “Be Loved by You”