You're wrong, it is very well possible with algorithms, and it has been a thing for many years already. For example check the Auto TLDR bot for Reddit. Here's an example of what it could look like for the article posted here: https://smmry.com/https://www.science.org/content/article/sc...

It's not even hard to do, you mostly have to count token frequencies which is hardly machine learning. Although you could also do it with BERT.