ChatGPT
vs Professor: The Good, Bad, and Bizarre of Machine Writing
The new
ChatGPT tool from OpenAI has rapidly garnered media attention
and millions of users. Although less effective sentence generators
have existed for years, ChatGPT's novelty has drawn both the
public and the reporters' interest, and its possibilities
have led to it becoming the fastest growing app in history.
But when examined more closely, its seemingly miraculous capabilities
are ultimately disappointing, at least when it comes to the
academic writing of my discipline. Like many of my peers,
I wanted to test ChatGPT's capabilities against the strictures
and demands of my field. A physics professor asked it thorny
questions about cosmology, and a programmer used it to generate
code, but I was more interested in whether the machine had
been programmed carefully enough that it could be used by
my students to cheat on their papers.
I tried
it with several prompts from a list of typical essay topics,
and found the AI generated essays were summary-dependent,
vague, fluent in terms of diction but in substance nearly
vacuous, and filled with elementary compositional missteps
as well as factual errors. A decent command of diction and
the ability to generate grammatically correct sentences is
not quite the same thing as producing meaning. I tried the
AI on three novels by H. G. Wells and Charlotte Perkins Gilman's
Herland, as well as several short stories by Thomas King and
others.
Although
I had been initially interested in how my students might use
the tool to plagiarize, I soon became intrigued by ChatGPT's
mistakes. Just as errors in meaning expose a student's facility
with argument, the AI's missteps tell us much more about its
nascent capabilities than the accuracy of its grammar or spelling.
It glibly
invented material it could not access, indulged in falsehood
when confronted, and wrote essays with a gleeful disregard
for logic and physics. Ultimately, to rework a quote from
H. G. Wells, it has developed in the most wonderful way the
distinctive silliness of a social-media-obsessed humanity
without losing one jot of the natural folly of a semi-informed
monkey on a typewriter.
ChatGPT
vs Professor: Struggling with Fiction and Poetry
Although
some people are using ChatGPT to write children's books for
sale, most people are looking for information or entertainment
when they type in a prompt. In this study, I became more interested
in what type of narrative it might create with a specific
or vague set of instructions. For the most part, its stories
were trite and hackneyed generalizations befitting a mere
sentence generator, but when it strayed from the norm, it
sometimes indulged in the wondrous and the odd.
This was
a rare occurrence, and instead I focused on how the machine
made its choices, or rather how it rationalized those choices.
I was interested in how it decided who would be hero and who
would be foe, what the gender of a character might mean, and
curiously, what stories it chose not to tell. I used the tool
to generate a list of story requirements, and then showed
that it didn't follow any of its own dictates. Instead, it
generated fairy tales with flippant ease, as though developing
a narrative was merely a matter of following a formula and
jamming words together which statistically cohere.
Although
on the whole the machine didn't impress me with its ability
to create clichéd stories about shallow characters caught
in typical circumstances, some elements of the stories rose
above the rest. When the machine didn't have a good handle
on what the prompt meant, or didn't know how to manage the
characters, it encouraged warfare and incest, nonsense and
incoherence, and in other ways showed that it still had much
to learn about both humanity and our stories.