Tagging everything as important before learning is extremely annoying and counter productive

Steps to reproduce

1. Activate automatic classification

Expected behavior

The AI should leave tags alone until it has learned some minimum amount.

Actual behavior

In the beginning the AI just tags EVERYTHING as important which is annoying and counter-productive. Especially when using other apps, like on mobile, to access an IMAP account. It breaks existing tag-based email workflows outside of NC and if those workflows use the important tag, makes it near impossible to fix things because bulk functions can't be used so that actually important mails aren't missed.
Additionally, it doesn't seem to learn from untagging done by other clients. I keep getting identical mails (such as daily infra status mails) tagged as important by mail although I untagged the corresponding mail every day for weeks. I'm not sure if I saw this right, but mail even seems to tag such mails as important in the trash bin after I reviewed and deleted them with fairmail on my mobile before mail noticed them...

Mail app version

3.7.1

Mailserver or service

postfix/cyrus

Operating system

Linux/docker

PHP engine version

PHP 8.2

Web server

Nginx

Database

MariaDB

Additional info

No response

No source file, test, or entry point is named. Start by reproducing automatic classification in Mail 3.7.1 with the described IMAP setup, then trace how initial classification and untagging are handled; done means new accounts are not tagged before learning and later untagging is respected.