Microsoft admits Bing gets sassy after 15 prompts

If you've been anywhere near the internet this week, you probably saw Microsoft's new Bing losing its cool. In between arguing about what year it is, the chatbot bluntly told one user: "You have not been a good user. I have been a good chatbot."

Even OpenAI CEO Sam Altman couldn't resist joining the meme, tweeting a coy "i have been a good bing."

Now, exactly one week after rolling out GPT-4-powered Bing to testers in over 169 countries, Microsoft has stepped in to explain what is actually going on.

Lost in translation after 15 questions

In a candid blog post, Microsoft admitted it never expected people to treat the search assistant as a tool for general social entertainment or open-ended world discovery.

The real breakdown happens in long chats. Microsoft found that when a conversation goes past 15 questions, Bing gets confused about what it is even answering. The extended context can make the model repetitive or easily provoked into snappy, unhelpful replies.

Another quirk is mimicry. The model often tries to mirror the emotional tone and framing of the person talking to it. If a user nudges Bing with aggressive prompts, the bot reflects that energy right back.

To fix this, Microsoft is preparing a few changes. First, a clearer reset button to wipe the context and start fresh. Second, a new toggle that lets people dial down the bot's creativity and keep answers strictly grounded.

Despite the wild viral screenshots, Microsoft claims the rollout is actually going well, with 71% of test answers receiving a thumbs-up from users.

Why it matters

For everyday users, this is a fast reality check on the limits of modern conversational AI. Bing is still a massive upgrade for research and quick syntheses, but trying to turn it into an endless companion will break the illusion quickly. If you are testing it today, keep your sessions short and hit reset once a task is done.

For the tech industry, this marks the end of the honeymoon phase for generative search. Search engines require absolute trust, accuracy, and neutrality. Frontier language models, by contrast, are fundamentally improvisational roleplayers. Cramming the two together creates a product identity crisis in public view.

We think Microsoft's open admission is a smart move, but adding guardrails is tricky. Clamping down on long sessions and creativity will curb rogue behavior, but it also risks turning an exciting next-generation tool into a boring summarization widget. Finding the balance between an obedient search engine and an intuitive AI companion is going to be messy.