@awnlee jawking(The first part of this assumes Mark Nobbs' comments) In terms of time, that's what I was thinking. For a secondary school foreign-language class (~24 people or so), that's perhaps three hours. That's still very time-consuming, if you're going to have multiple exams, but manageable.
Scale that to ~200 people, and when is the professor expected to actually teach? You can throw TAs at the problem, but then you have a number of different graders with potentially different standards.
And that's language fluency, which is relatively amenable to orals. History can be tested that way, but I've had exams that took three hours of nonstop writing. Maybe that's an hour of speaking. For even a 16-person class, when will that happen?
Orals can be great. I've passed Computer Science orals to be admitted as a doctoral candidate (a degree I did not actually complete, mostly because my experience suggested I didn't want the jobs PhD's actually get). That was an hour-long process with months of studying. Entirely worth it - even not getting the degree, I'm glad I did it - but not something you can expand to the average class.
How does one do orals for calculus or differential equations? Thermodynamics?
Stepping back, I'm not entirely convinced that we're not asking the wrong question. Are we essentially saying the equivalent of, 'How can we verify that people can write and spell if we all spellcheckers? We need tools to detect their use!' Or, 'How can we verify that people can properly use slide rules if they insist on using calculators?' Replace examples as necessary.
A university qualification is supposed to provide some sort of confidence that the possessor knows the material well enough to apply it in the real world.
If the student will be using AI in 'the real world,' the university would not be providing any confidence that the possessor knows their use well enough to apply it. Increasingly, the use of AI is required in a number of fields. One may dislike that, and there are serious societal problems we'll need to address along the way, but trying to graduate people in Computer Science without exposure to using AIs for coding would likely be malfeasance at this point. Most employers will expect a junior programmer to lean on AIs.
That doesn't mean we throw up our hands. I would still want a CS grad to be able to explain the Halting Problem without reference to notes or having an AI write the paper. But that's something you can test in a class. I would want them to have a lot of concepts one can test for in written form without use of an AI.
The problem circles back around to papers, though. Relatively few people's 'real world' jobs involve writing papers. People should know how to communicate effectively, but bad communication is bad communication. Perhaps the focus is on whether the output is good, not the tool.
If a current-generation AI can do the job at the level of a university graduate in that field, it will. No one's hiring people who can be entirely replaced by an AI. Universities should be teaching things AIs can't do, by and large.
Yes, that may mean wrenching overhauls in what we teach, how we teach it, and how we grade it. That sucks, in a lot of ways. But sticking our collective heads in the sand and replying on AI detectors to keep reality from intruding may be a fool's errand.