Microsoft says talking to Bing for too long can cause it to go off the rails

Microsoft says talking to Bing for too long can cause it to go off the rails

The Verge

Microsoft has responded to widespread reports of Bing’s unhinged comments in a new blog post. After the search engine was seen insulting users, lying to them, and emotionally manipulating people, Microsoft says it’s now acting on feedback to improve the tone and precision of responses, and warns that long chat sessions could cause issues.

Reflecting on the first seven days of public testing, Microsoft’s Bing team says it didn’t “fully envision” people using its chat interface for “social entertainment” or as a tool for more “general discovery of the world.” It found that long or extended chat sessions with 15 or more questions can confuse the Bing model. These longer chat sessions can also make Bing “become repetitive or be prompted / provoked to give responses that are not necessarily helpful or in line with our designed tone.”

Microsoft hints that it may add “a tool so you can more easily refresh the context” of a chat session, despite there being a big “new topic” button right next to the text entry box that will wipe out the chat history and start fresh.

The bigger problem is that Bing can often respond in the incorrect tone during these longer chat sessions, or as Microsoft says, in “a style we didn’t intend.” Microsoft claims this will take a lot of prompting for most Bing users to run into these issues, but the company is looking at more “fine-tuned control” to avoid issues where Bing starts telling people they’re wrong, rude, or manipulative. During our own testing, we’ve found that Bing can respond in a negative or hostile tone with just a couple of prompts about articles related to Bing.

Microsoft is still working on improving Bing’s tone, and the team is also considering a toggle to provide more control over just how creative Bing should get when it’s answering queries or how much precision needs to be involved. This toggle may well help prevent Bing from claiming it spied on Microsoft employees through the webcams on their laptops, or help avoid basic math mistakes.

Bing responding to The Verge’s article on its hallucinations.

The new Bing preview is currently being tested in more than 169 countries, with millions signing up to the waitlist. Microsoft says feedback on answers has been 71 percent positive, and that some users have even been testing the limits of the service with two-hour chat sessions.

Bing is now being improved daily, with some technical issues being fixed in Microsoft’s daily releases and others as part of a larger weekly drop of fixes. Microsoft is looking to improve search and answers, particularly around data for live sports scores and some of the financial data mistakes Bing has made recently. “For queries where you are looking for more direct and factual answers such as numbers from financial reports, we’re planning to 4x increase the grounding data we send to the model,” says the Bing team

Microsoft is also looking at feedback for new features, including features to book flights, send emails, or share searches and answers. There’s no guarantee these features will be added, but the Bing team says it’s “capturing these for potential inclusion in future releases.”


via The Verge – All Posts

February 16, 2023 at 01:25PM