This is one for the AI Wierdness files. Except the results are weird because the NLG did a passable job. A frighteningly, literal passable job: EduRef found that the GPT-3 program could get passing grades on College assignments [1].
The provocative thing, of course, is that the AI was able to place in the middle of the pack. It also received similar feedback to the humans. And, of course, it took 3 days for the puny carbon based life forms, and about 20 minutes for the silicon based responses.
OK, the AI flunked the creative writing assignment, and was awarded generally mediocre grades. I mean, ‘C’ is not really a passing grade, not where I come from. More than a couple of C’s and you’re out.
Also, the feedback was recognizably different. The machine was critiqued for producing, well, mechanical, product.
But still. It’s not that the dog got a ‘C’, it’s that the dog danced at all.
I’ll note that the graders were blind as to the identity of the submitters, and were not specifically asked to guess which were human and which were AI. That, of course, raises uncomfortable questions about the process of routinely grading student writing. Graders are focused on efficiently checking against the specific assignment, and much of the evaluation is pretty shallow, and certainly colored by expectations and comparisons with previous (human) students.
In this respect, the AI is really trying to hide in the herd, to mimic other students. I.e., not necessarily to write well, but to write like an undergraduate. (In the future, we may see an AI ‘dumbing down’ it’s writing to not seem too advanced for a student.)
A cynic might say that some students might follow the same strategy. And the graders are explicitly expecting undergraduate level performance, and evaluating against that implicit standard.
So, worst case, teachers are rewarding shallow mimicry of a ‘better than average student paper’, and students are learning to produce the same. It’s called ‘college’. : – )
Anyway, these philosophical questions about what is taught and what is learned aren’t new. The AI just demonstrates that they can be learned to some extent by ML.
The article does not discuss how the AI was trained, but it must have been given a dataset of example papers to learn from. (One wonders if one trained with only A+ papers as example, shouldn’t the results be ‘A’ papers?)
This study does provoke some interesting thoughts.
One thing I wonder is whether the AI could learn to get better at this exercise. This would be tested by feeding the feedback from graders into the ML to improve the next paper. This experiment could also compare the putative improvement of the human students, as well. Does practice help? Does grading help?
Another interesting question is, ‘Is it cheating to have an AI write the first draft of your paper?’
I suspect that you could get a better product with a hybrid approach, start with AI, and edit to improve. Even better, have a mechanism to seed the AI with parameters, say, a partial outline, some key refs, maybe some ‘strategy notes’ (e.g., guidance that this paper should explore the ‘loyalist’ point of view on the American Revolution).
The idea is to produce a draft for the student to edit, fill in, and generally finish. (And, by the way, done right, the machine generated draft might be a really useful teaching instrument.) If you are nervous about students just turning in the draft, have them submit both the AI draft and their modifications.
On the other side of the game, an AI grading assistant would be interesting. If AI can rapidly mark up the basic stuff (e.g., grammar, possible fraud, major issues, overall style), the human grader can review and then focus on higher level feedback, such as advice about methods and sources, examining biases, performance relative to standards, and intangibles.
And finally, we can make students compete against AI. If we can construct AI to reliably get an ‘A’ or ‘B’ or whatever on an assignment, then let the students see the results and try to do as good or better. This would be particularly valuable if the AI can explain itself—the ML would then be teaching writing, or at least something about writing.
Oops, NLG generally isn’t good at explaining itself, is it? And, in fact, it generally learns a lot of really shallow and stupid stuff, like word frequencies and co-frequencies. Most of the insides of NLG are probably useless for students.
So this is yet another use case that really needs ML to explain itself.
- EduRef.net, What Grades Can AI Get in College?, in EduRef.net, February, 2021. https://www.eduref.net/features/what-grades-can-ai-get-in-college/
- Greg Nichols, AI can write a passing college paper in 20 minutes, in ZDNet, February 24, 2021. https://www.zdnet.com/article/ai-can-write-a-passing-college-paper-in-20-minutes/
Robert McGrath's Blog 
One thought on “AI writes term papers”