I think the issue is largely that in the age of AI, learning a skill requires one to be deliberate and dedicated, but the entire reason grades and exams are so prominent is because most students need the threat of near term failure to learn.
Open ended projects were always my favorite ones because I was able to utilize some of my personal projects on them. Profs also enjoy seeing students' passion for their topic. That kind of student is probably still doing well.
Those kinds of projects are a big trap for certain kinds of people. If I had used ChatGPT in college, maybe I wouldn't have come up with overly-ambitious projects that seemed doable at the time but instead took several full weeks of work and basically made my semester miserable.
Ensures that students have a plan, that the goals are achievable, and can help to spread out the project's points.
Also helps when profs make it clear that they're more concerned about seeing what you know/learned and how you approached the problem, not whether or not you achieved exactly what you promised.
LLM over-reliance has an analogous problem: People start using LLMs to start projects that get in over their heads, then when the LLM has gone through all of the easy and medium problems they're left with only the hardest problems to solve on top of a project where they didn't learn much doing the easy and medium parts first.