Microsoft says talking to Bing at length confuses it

The new version of Bing which uses theAI he’s causing quite a stir for the strange answers he sometimes gives – e Microsoft stepped in to explain why. Indeed, it seems that, in particularly long conversations, the AI ​​gets confused and you start answering incorrectly, going so far as to lie and “scold” users. But the Redmond company promises to be already working to resolve the situation.

Microsoft explains why Bing’s AI gets confused

After seven days of testing with a small number of users (we are still waiting to be able to access the beta), Microsoft has taken stock of how the version of Bing that leverages ChatGPT-like AI.

Users are using AI mostly for conversation and less for research. But he also assessed that, after 15 or more applications, the Bing’s AI model starts to get confused. Especially when questions become “repetitive or solicit/provoke answers that aren’t necessarily helpful or in line with the tone we’ve designed.”

Microsoft explains that it might add “a tool to more easily update the context” of a chat session, despite there being a large “new topic” button next to the text box.

The main problem is that Bing responds incorrectly during these longer sessions – although Microsoft just says “not in the style we intend”. Some users have posted on social interactions in which the Bing’s tone became almost hostile when users pointed out some shortcomings. The Verge for example points out that Microsoft’s Bing has AI defined as “partisan” the journalist who wrote an article reporting his problems.

Microsoft is reportedly working on a gradient of tone and “creativity” of Bing, so that users can set different levels of interaction. And in general, the company is applying fixes every day, even implementing new features such as the ability to book flights or send emails. The road to Bing AI still seems long, but Microsoft thinks it can improve it to make it a great tool for online searches – even if the attention in verifying the sources and evaluating the answers on our part cannot fail.