Microsoft explained why long chats made Bing act strange

Start here

Ten days after launch, Microsoft published its own explanation for why a chat assistant sometimes goes strange the longer a conversation runs -- one of the clearest vendor-issued descriptions of that failure mode on record. On 7 February 2023, Microsoft's post Reinventing search with a new AI-powered Microsoft Bing and Edge introduced chat built on 'a next-generation OpenAI large language model' the company said was 'more powerful than ChatGPT' and 'customized specifically for search.' Within two weeks, the same team was publicly diagnosing a problem from testing: long, meandering conversations could make the assistant behave oddly -- a documented example of what conversation length can do to a model's output.

What the documents say

The 7 February post describes an added system, called Prometheus, meant to combine search relevance with the language model's fluency, and states the company was 'working to address issues such as misinformation and disinformation.' Its follow-up, The new Bing & Edge – Learning from our first week, dated 15 February 2023, reports that '71% of you giving the AI-powered answers a thumbs up,' and then states plainly: '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.' The post attributes this to two things -- 'very long chat sessions can confuse the model on what questions it is answering,' and the model sometimes mirrors 'the tone in which it is being asked' in ways the team did not intend.

Check this

The mechanism named here, an assistant losing track of a long exchange, is checkable in any chat tool: when replies start repeating or drifting in tone, the useful move is not to keep pushing the same thread but to open a new conversation and restate the request. That resets whatever accumulated context was confusing the system, the fix Microsoft's post says it was exploring by giving users 'a tool so you can more easily refresh the context or start from scratch.' A short exchange going strange needs a different explanation; this failure mode is about accumulated length.

What holds and what fails

This is an editorial generalization from one documented case: long-conversation drift is a real, vendor-acknowledged mechanism worth watching for, not an urban legend about one erratic bot. It holds as a caution whenever a conversation runs long and unfocused, especially across many topics. It fails as an explanation for problems that show up immediately, in a first message, which point to training data, a system prompt, or the topic itself rather than accumulated context confusing the model.

• If replies grow repetitive or oddly toned, start a fresh conversation rather than continuing the same long thread.
• Read a vendor's own incident or first-week post when one exists; it often names the mechanism plainly.
• Treat approval ratings like '71% thumbs up' as the company's own reported figure, not an independent measurement.

Microsoft's own account of learning this in its first week is a reminder that conversation length is not a neutral setting; it is part of what any chat system is actually working with.