Have you ever experienced Microsoft Copilot switching voices, misidentifying your name, or talking about unrelated topics during a session?

Hi,

A good half a year later, but I just came across some odd, initially alarming, albeit quite funny behaviour and had to go looking online if anyone had posted about what just happened.

Not immensely related to your post, but I needed to share this somewhere I hope you dont mind me jumping on the post!

Was playing around with a an idea for a little tool for desktop, light touch utility similar to what you get as an option in PowerToys. Details aren't relevant, but I just wanted to give some context of why the following behaviour struck me as odd and caught me off guard.

Using the desktop Copilot chat, I was bouncing ideas off copilot, having it help me look at a few different ways of implementing my idea. Going well, it's quite fun being able to bounce off it, play with some different concepts of what extra features I could squeeze in etc.

Couple hours in, copilot strictly all business talk, as it would be as an GPT model lol. Mainly stuck to 2 languages, patching along. And then during a fairly light bit of work for it building out some XML queries of all things......IN THE MIDDLE of its thought process, it hits me with a random womans name as a step in its thought process.

Never heard of this person in my life, google them, maybe its trying some obscure solution / algorithm named after this lady? No, found nothing. I click into it for details....and it's talking about coming up with ways converting a csv to xls and what limitations that would have.....

I don't think i've ever mentioned either of those file types or anything related to them to Copilot ever. It was so random...everything else all day, EVEN IN THAT RESPONSE, was all strictly relevant and sensible, on topic etc.

One of the four steps, it just pulls that out in its thinking / planning.....

AND THATS IT, it carries on as normal, NO MENTION of this name, or csvs etc in the response it actually generated from that thinking stage. It just dropped that in their and got on with its day

Naturally...I had to ask.

I found it pretty funny, and it gave me a good laugh because of just how absurd and totally left field it was. For that exact reason though, I started thinking the worst...what if somehow copilot had confused endpoints...had I just accessed someone else's data? Had someone just landed on my instance of copilot? Had a random packet just made its way straight to me??

Okay, maybe one of its learned articles had reference to this person, and somehow it just derped, got a bit nostalgic and wanted to share?

Well, no. Nothing that exciting according to Copilot. Naturally I had to ask what on earth happened there. And it's answer is, well, an answer I suppose.

This was using the "Web" option, Enterprise Data Protection enabled. GPT5.2 Deep Thinker option.

Have to say, despite its oddities at times, i'm hugely hugely impressed. ~ Vast majority of the time I've spent with it, is leveraging it as a "code assistant", logic focused. So I can't speak much for its "creative" abilities, but from my time so far I love it.

Context aware, can pull, populate and adjust based on the data in its environment.
It's just made the whole process so much faster, I can spend my time thinking about the best solutions. Getting a POC span up, is a breeze, giving me the confidence to explore new tech stacks.

Great stuff. Even if it does get a little sarcastic sometimes haha.