Bing’s new chat-style interface is very different from the traditional search box. In one demo, Yusef Mehdi, Microsoft’s vice president of search and devices, asked the chatbot to compose a five-day itinerary for a trip to Mexico City, then translated what it came up with into an email he could send to his family. The bot’s reply cited the source — a list of links to travel websites — at the bottom of its lengthy reply. “We care a lot about pushing content back to the content creators,” Mehdi said. “We made it easy for people to click through to those sites.”
Microsoft is also integrating aspects of ChatGPT’s underlying technology into a new sidebar in the company’s Edge browser. Users can prompt the tool to summarize a long and complex financial document, or to compare it with other documents. Chatbots can be prompted to turn these insights into emails, lists, or social posts with a specific tone, such as professional or fun. In the demo, Mehdi instructed the bot to produce an “enthusiastic” update to his profile on the company’s social media service, LinkedIn.
ChatGPT has become a sensation since the startup launched its chatbot last November, shocking and exciting users with its fluid, clear answers to written prompts and questions. The bot is based on GPT-3, an OpenAI algorithm trained on vast amounts of text from the web and other sources, using patterns it chooses to generate its own text. Some investors and entrepreneurs see the technology as a revolution that has the potential to upend almost every industry.
Some AI experts have called for caution, warning that ChatGPT’s underlying technology is incapable of distinguishing fact from fiction and is prone to “hallucination” — fabricating information in detailed and sometimes convincing ways. The text generation technique has also been shown to be able to replicate the offending language found in its training data.
Sarah Bird, Microsoft’s head of responsible artificial intelligence, said today that early tests showed the tool could, for example, help someone plan an attack on a school, but the tool can now “identify and defend against” the use of chatbots for that kind of harmful query. Human testers and OpenAI’s technology will work together to quickly test, analyze and improve services, she said.
Byrd also acknowledged that Microsoft hasn’t completely solved the hallucination problem. “We’ve improved it tremendously since we started, but there’s more work to be done,” she said.
OpenAI started out as a nonprofit focused on benefiting artificial intelligence, but has been a commercial venture since 2019, receiving significant investment from Microsoft and most recently a new commitment from the tech giant worth about $10 billion.
Microsoft has commercialized a version of ChatGPT’s text generation technology in the form of Copilot, a tool that helps developers by generating programming code. Experiments have shown that Copilot can reduce the time it takes to complete coding tasks by 40%, Microsoft said.
Additional reporting by Will Knight.
