Do Students Who Rely on ChatGPT "Benefit"?

AI has its uses, but it's not a substitute for writing--or learning.

Photo by Berke Citak on Unsplash

---

Artificial intelligence has its uses in education. But it’s not a substitute for learning to write—or learning in general.

A large-scale study at a public university has found that the quality of student writing improved after the advent of ChatGPT in late 2022, especially for students from families that don’t speak English. But if students just submitted text created by ChatGPT, does that count as improvement?

The study, discussed recently by Jill Barshay in her column for the Hechinger Report, was conducted by a team led by Renshe Yu. It evaluated over a million student writing submissions at a university serving many students from Spanish-speaking families and international students who are not native English-speakers. The researchers characterized all of these students as “linguistically disadvantaged.”

The researchers themselves didn’t evaluate all those writing samples, mind you. They got computer models to do it. That itself is a possible red flag, since such tools can be unreliable for evaluating writing quality.

But assessment isn’t the main problem with the Yu study. Rather, it’s the assumption that if students use ChatGPT to “help” with their writing and the end result is better than what they could have produced without ChatGPT, that means the students have benefited.

Yu and his co-authors are in fact primarily concerned about equitable access to those supposed benefits. Among the linguistically disadvantaged students in the study, those from higher socioeconomic groups—many of whom were international students—showed more improvement in writing than their less privileged peers. That’s probably because wealthier students have more access to ChatGPT and are more tech-savvy, the researchers speculated.

(A lot of the study is speculative. The researchers don’t even know whether students whose writing improved used ChatGPT. Yu told Barshay that they inferred the cause was ChatGPT, because of the timing and because “what other things would produce these significant changes?”)

Educators and policymakers, the paper concludes, need to “ensure that the benefits of LLMs are shared equitably, regardless of socioeconomic status, geographic location, or cultural background.” (“LLMs” stands for “large language models,” of which ChatGPT is one.)

Are The Benefits Real?

But are wealthier students getting real benefits? As Barshay points out, “this improved writing quality doesn’t mean that these international students are actually learning to write better.” It doesn’t even mean they did the reading. It’s possible, Barshay notes, that they just fed the assigned reading and the professor’s question into the chatbot and submitted the response it spat out.

The researchers’ conclusion that students were benefiting from ChatGPT seems to rest on the assumption that they were using it just to overcome their lack of facility with English. But if that were the case, they could have written their assignments in their native languages and used Google Translate—which was around long before ChatGPT. If, as the timing suggests, they were using ChatGPT, they were likely getting it to do their writing for them.

To be clear, I’m not arguing against any use of AI in education. AI can do some extraordinary things. The best results seem to occur when AI is used to supplement rather than replace teachers or tutors. One study, for example, found that an AI system that prompted math tutors to think more deeply about their interactions with students made the weakest tutors nearly as effective as their highly rated peers. With writing, AI could be helpful for brainstorming ideas and editing drafts.

But, as others have noted, there are significant risks. As seems to have happened in the Yu study, students often use AI to avoid the cognitive challenges presented by many writing tasks. And teachers sometimes use it to enable students to avoid reading complex text.

About 10 percent of the K-12 teachers in a recent RAND survey, for example, said they used AI to “adjust” the reading level of texts “to match students’ needs.” If you’re a busy teacher with a bunch of ninth graders who are reading at a third grade level—which in some high-poverty schools is not uncommon—it’s understandable that you would be tempted to use AI to simplify the texts in the curriculum. But that approach can backfire.

If, for example, you ask ChatGPT to convert the text of the Gettysburg Address to a third-grade reading level, you get this:

Four score and seven years ago, our country was started with the idea that all people are equal. Now we are in a big fight, and many people have died to keep our country safe.

We are here today to remember them. We want to honor the people who fought and died for our freedom. But we know that just saying words is not enough. We must make sure that the country they fought for stays strong and free.

This version preserves the basic meaning of the original (although ChatGPT seems unaware that most third-graders would be stumped by the phrase “four score and seven years ago”). But it omits a lot of nuance, not to mention the ringing language that has made the Gettysburg Address an iconic text. Students who get the ChatGPT version might justifiably wonder why it’s considered such a big deal.

Perhaps even more important, they’ll never get the exposure to the complex vocabulary and syntax that will enable them to understand all sorts of texts they’ll be expected to read in the future—not just in high school and college but throughout life. They may struggle to make sense of newspaper articles, job training manuals, and lease agreements.

Well, you might say, they can always just run those texts through ChatGPT and get a simplified version. True. But ideally, texts aren’t complex just for the sake of complexity. They’re complex because they’re conveying complex ideas. And the ability to understand complex text reflects the ability to think in more complex ways. Simplified texts can be used as stepping stones, but limiting students to simpler texts limits their cognitive development.

Using AI as a Substitute for Writing

Producing complex text is, of course, even harder than understanding it. No wonder many students resort to ChatGPT—although how many is unclear. One AI detection tool identified only 3 percent of assignments as generated mostly by AI, but—like computer models that try to evaluate the quality of writing—such detection tools can be unreliable.

Anecdotal evidence indicates that many students are using AI to do their assignments for them, whether they’re disclosing that to their teachers or not. The Wall Street Journal reported on one New Jersey high school student who used generative AI to “cheat her way through” her junior year. Although she used a bot for dozens of assignments, she was caught only once and passed all her classes.

Even more dramatic is the case of a Tennessee student with dyslexia who graduated from high school with a 3.4 GPA but couldn’t even read or spell his own name. He was able to pass his courses by using various authorized accommodations, including—yes—ChatGPT.

ChatGPT does turn out to be a pretty good student, even at an elite institution. One undergraduate at Harvard—which, according to another Harvard student, is “in the middle of a techno-panic, with AI at its center”—conducted an experiment in which she submitted essays written by ChatGPT. The bot “ended the year with a respectable 3.57 GPA.” (The student’s instructors knew she was conducting an experiment and were told that the essays had been written either by her or by AI, although in fact all were written by AI.)

But even if ChatGPT does well, the students who use it as a substitute for doing their own work are losing out. One large-scale study of Turkish high school students found that those who used AI to study for a math test did worse than those who studied the traditional way. Researchers hypothesized that students had used AI as a “crutch” rather than doing their own thinking. Another study in the UK found “a very strong negative correlation between subjects’ use of AI tools and their critical thinking skills.”

The New Jersey student who relied on AI during her junior year decided to abandon it the following year, saying she had learned far less than she could have. “I have tried to take a step back,” she told the Journal, “and use my brain.”

Why Writing Instruction Is Not Obsolete

Some AI advocates argue that teaching students to write has become obsolete. “What is the value of an essay?” a member of the education team at OpenAI asked at a 2024 conference, according to the Wall Street Journal. She suggested that it makes more sense to measure critical thinking and communication skills based on students’ ability to use AI well.

Daniel Willingham, a cognitive psychologist at the University of Virginia, pushed back, saying that writing “requires a type of thinking that other types of exercises don’t.” You need to be more careful in your explanations, he said, and more complete in your arguments.

I agree. As I’ve argued before (see here and here, for example), writing isn’t just a matter of putting down what’s already in your head—or at least, it shouldn’t be. Ideally, through the process of organizing your thoughts and conveying them in the more precise and complex language of written text, you’re reinforcing your knowledge and deepening your comprehension—or maybe discovering that you don’t understand the topic as well as you thought you did. Subcontract that out to a bot and you’re missing out on a lot of learning and cognitive development.

In their recent study, Yu and his co-authors worry that wealthier students are more likely to reap the “benefits” of ChatGPT. But the lower-income students who have to rely on their own wits may actually be benefiting more in terms of what they’re learning.

Ideally, rather than expanding access to ChatGPT, as Yu and his colleagues urge, we would be expanding access to effective writing instruction. The ability to read complex text requires the ability to understand complex thinking, but teaching students how to write in complex ways can enable them to think in more complex ways themselves.

Yu, an instructor at Columbia Teachers College, should be aware of the pitfalls of encouraging students to use ChatGPT. In his classes, he told Barshay, more and more students have been submitting writing that appears to have been generated by AI. It “seems reasonable,” he said, “but doesn’t make a lot of sense.”