What They Did
The researchers had medical and humanities experts identify
which of a set of German-language term papers were written by medical students
and which were generated by ChatGPT 3.5, as well as rate several aspects of
each paper. Each participant had a week to make their decisions, though they
were instructed not to discuss the project. Participants identified student or
AI authorship of the papers 70% of the time, with no significant differences
between the medical or humanities experts, nor correlation with participant
traits such as experience in academia, experience with ChatGPT, or knowledge of
the subject matter.
For papers that were correctly identified, the medical
experts rated the student papers as having better language use, logic, and
scientific approaches, while the humanities experts gave the
correctly-identified student papers better ratings on scientific approaches but
not other traits. Medical experts also gave student papers better ratings on
citation of sources, even when they incorrectly identified them as AI-generated.
When papers were correctly identified, medical experts rated student papers
better at suggesting new research directions, but when they were incorrectly
identified, the AI papers were rated better.
In follow-up interviews, participants frequently indicated
that they identified papers as AI-generated because they were redundant,
repetitive, or lacked a sense of coherence. The researchers point out that
although the participants relied heavily on linguistic style to distinguish between
the student and AI-generated papers, it is not clear how effective they would
be if they didn’t already know that one of the papers was AI-generated.
Further Exploration
Difficulty distinguishing between papers written by students
and those generated by AI poses a major challenge to education, and advances in
AI will likely make it even harder. AI-generated books have appeared for sale
on Amazon, including a mushroom-foraging guide that could pose a real danger
with inaccurate information (see https://www.theguardian.com/technology/2023/sep/01/mushroom-pickers-urged-to-avoid-foraging-books-on-amazon-that-appear-to-be-written-by-ai.)
At the same time, students may be falsely accused of cheating with AI and have
trouble proving otherwise (see https://odsc.medium.com/ai-detectors-wrongly-accuse-students-of-cheating-sparking-controversy-7afb2ea7edc8).
The prompts the researchers used for generating papers with
ChatGPT simply told the program to write sections of the paper one at a time with
citations. I imagine that a tech-savvy student could edit such a paper to sound
more human without having to know much about the topic. At the same time, using
natural langue to prompt an AI can make it more effective than a traditional thesaurus,
and students can also use AI to help organize and structure their own ideas.
These uses of AI seem akin to the use of a calculator in a math class.
More complex is the question of what constitutes fair use of written material in building an AI. On one hand, the major AIs have been built using copyright material without permission (see https://authorsguild.org/advocacy/artificial-intelligence/faq/.) On the other hand, my own human “brain soup” is full of fragments of books I’ve read and ideas I’ve absorbed without the ability to truly credit all of the sources. Figuring out what’s ethical in this sense is a huge question, but that’s a rabbit hole for another day!
https://commons.wikimedia.org/wiki/File:Literatursuche_mit_KI_(3).jpg
.jpg){kind=link}
No comments:
Post a Comment
Please be nice.