Incident: Long chat sessions confused Bing's model

Module 7 · Incident file · About 30 minutes

What happened

In February 2023 Microsoft was testing its new chat feature in Bing with a limited preview group "in over 169 countries". On 15 February the Bing team published what it had learned in the first week. One finding was about length: "in long, extended chat sessions of 15 or more questions, Bing can become repetitive or be prompted/provoked to give responses that are not necessarily helpful or in line with our designed tone" (Bing Blog, 15 February 2023).

The team gave two reasons. First: "Very long chat sessions can confuse the model on what questions it is answering", so "we may need to add a tool so you can more easily refresh the context or start from scratch". Second, the model "at times tries to respond or reflect in the tone in which it is being asked". The post also noted "a few 2 hour chat sessions".

Two days later, on 17 February, the fix shipped, and it was a hard limit: "the chat experience will be capped at 50 chat turns per day and 5 chat turns per session", where a turn is one question and one reply. "At the end of each chat session, context needs to be cleared so the model won't get confused", using a broom icon for "a fresh start". Microsoft said "only ~1% of chat conversations have 50+ messages" (Bing Blog, 17 February 2023).

The cost was the feature itself. On 21 February the team wrote that the limits came "in response to a handful of cases in which long chat sessions confused the underlying model", that such sessions "are not something we would typically find with internal testing", and that many users wanted "a return of longer chats". It raised the limit to 6 turns per session and 60 chats per day, with 100 a day planned (Bing Blog, 21 February 2023). Microsoft did not publish how its chat context was built, so this file claims nothing about that; it replays the pattern its posts describe.