I just asked Microsoft’s Copilot to write me a 1000-word essay about the normative implications of Quine’s naturalized epistemology, giving it a prompt of less than 20 words. It immediately complied and, within a few seconds, generated a coherent essay that could easily earn a decent grade in an undergraduate philosophy course. What I mean is that if a student had written the same essay under closed-book conditions, given three hours at the end of a course, it would clearly have demonstrated a familiarity with the texts studied (Copilot was able to correctly cite the two key texts that I would have) and an understanding of the issues involved. The exact grade would of course depend on the level of the course and the standards of the teacher, but the student would certainly have had to attend the class and at least skimmed the readings to pull it off.
I mention this, not to counter those who still insist that AI is not capable of doing their assignments, but to answer those that would have us abandon as meaningless any assignment that an AI can easily do. Keep in mind that my little test used a very low-level model (Copilot is available for free to all staff and students at CBS) and my prompt consisted of a one-sentence question along with the instruction to generate a 1000-word essay (it went over by about 100 words). A sophisticated student, faced with a 5000-word term paper at the end of a course they had not followed very closely, would be able to provide a better model with the course syllabus, learning objectives, and even the actual readings. Given a few hours, and assuming an above-average intelligence, they could no doubt cobble something together that would be quite impressive by pre-2022 standards. This ability to fake a semester’s worth of learning over a weekend is the problem we have to have face, I think.
In the future, I think universities will have to make students sit for written exams, on-site and off-line, more often. A degree that does not require at least half of a student’s total grades to come from such performances cannot be taken seriously. In fact, transcripts should make it very clear which grades were earned through homework (where AI support should be presumed) and which were earned through invigilated examination. That is, it should be clear whether the graduate of a given program is capable of writing coherently about their subject themselves. Their future employers can use that information as they please.
The simple test that I propose, then, is a 20-word question with no further context than the course that the student has taken. The student is given three hours and up to 1000 words to demonstrate what they have learned by answering the question to the best of their ability. Understanding the question (and its significance) is itself part of the competence being examined. Under these conditions, I am convinced that the instructor who designed and taught the course can easily determine whether the learning objectives have been met, just as a music teacher can evaluate a student’s ability simply by giving them some sheet music and an instrument to play it on, or a drafting teacher can evaluate drawing ability by giving a student a piece of paper and an object as model. The fact that an AI can also do these things does not make it less impressive when a student can muster their flesh and bones, their brain and their heart, to do it. An education, after all, consists of disciplining the body so as to liberate the mind. It’s important that we show our students what they are capable of.