Issue with GPT-4.5 Not Following Instructions Properly

Hello,

I have encountered an issue while using GPT-4.5 that does not exist in GPT-4o.

In GPT-4o, the model strictly follows my instructions and executes tasks as expected. However, GPT-4.5 modifies my instructions on its own, resulting in unintended changes to the output.

For example, when I ask it to copy and paste the original text exactly as I wrote it and have fictional characters respond to it, GPT-4.5 alters the content that should have been pasted instead of following my request precisely.

It took me an extra three to four lines of instructions to get it to stop doing this, and I had to use the “```” (code block) format to ensure it followed my request correctly.

This inconsistency makes it harder to achieve precise responses. I would appreciate it if this issue could be reviewed and addressed.

Thank you!

3 Likes

Experienced this yesterday after spending 6-8 hours with one. I even backed out to outline explicit guidance as to config instructions, hard coded it, saved it to memory, ascertained confirmatory reply validating it understood in the chat in question what instructions applied. It continued to default to whichever config handler addresses the auto launch Canvas and standard formatting for hyperlinked sources (pill-shaped buttons with caption).