In a world where AI can write research papers, solve equations, and create works of art, educators worry about how college students will use, misuse, or miss out on AI. However, there is little comprehensive research on college students and their use of AI.
Now, Igor Chirikov, a senior researcher at the University of California, Berkeley’s Center for Higher Education Research, has collaborated with researchers from the University of Technology Sydney and Cornell University to present the largest study of undergraduate students’ use of generative AI. More than 95,000 students at 20 research-focused public universities answered questions about how they use AI, including whether they use it to cheat. The results of the survey were published in a journal on May 21st. science.
Bora Reid/University of California, Berkeley
“The advent of artificial intelligence technologies and GenAI tools like ChatGPT has had a huge impact on higher education, faculty, students, and everyone involved,” Chirikov said. “We didn’t know much about how students were using and abusing it.”
The study, conducted in spring 2024, used data collected by Berkeley’s Research University Student Experience (SERU) Consortium, a group of research universities that collaborate on student surveys to improve higher education. Approximately two-thirds of respondents said they use GenAI, and nearly 40% use GenAI monthly or more frequently. Additionally, at least 9% of students who used AI reported using it to cheat. The numbers varied widely by academic discipline, with more non-STEM students cheating with AI than STEM students. However, researchers warn that banning GenAI will not stop fraud and may even harm students seeking jobs in industries where AI proficiency is expected.
Chirikov and his co-authors recommend that academic programs find new ways to measure student knowledge and abilities that cannot be faked by AI, but it is not an easy undertaking for programs that require deep critical thinking and building skills over time.
The study also found alarming disparities in the use of AI across demographic groups, including lower rates of AI use among low-income, racially underrepresented students, and female students. These students may fall behind in college and ultimately in the workplace due to unequal access to and use of AI.
Generative AI is evolving so rapidly that this study “feels like it’s already from a past life,” Chirikov said. Still, he said his findings are important because they can help universities and students alike think about what uses of AI would be most useful.
University of California, Berkeley News recently We spoke with Chirikov about the growing number of students using AI, what universities can do to test student learning, and how to ensure even students with fewer resources can gain the AI proficiency they need for their careers.
Your research shows that it is not always clear to students what is cheating and what is acceptable use of generative AI, and there can be a slippery slope to cheating. 26% of daily AI users say they use AI for fraud, but only 7% use it monthly. Can you talk a little bit about that?
AI policies vary widely by course, ranging from instructors allowing AI across the board, including exams, to banning it completely.
Now, when you search for something on Google, AI is used. Because it usually pops up with a quick AI-generated summary. Or, if you use grammar and editing tools, sometimes there’s even integration with AI, so the AI can rewrite the entire thing with one click. The level of integration of these tools is incredibly high and makes them very appealing to use. It is extremely difficult for students to self-regulate regarding classroom AI policies and navigate this complex environment.
We don’t know if the frequency of AI use is causing students to cheat more, or if students who are more likely to cheat in general just tend to use those tools more often. However, this trend is very worrying in that there is a clear correlation between AI abuse increasing as students use it more frequently.
What do you think motivated students who knew or thought they were cheating but still cheated?
This study was conducted at a highly selective research university with many talented students, a competitive environment where grades are very important, and having a perfect GPA is important for internships and admission to graduate school. Higher education is also expensive for students and families, so there is pressure to keep up and graduate as quickly as possible.
On the other hand, there are many AI tools that are very easy to use. Instead of staying up all night on an assignment, you can create and submit something in under 30 minutes.
I think these two forces create a perfect storm for AI-assisted fraud. And learning is to some extent a victim of that perfect storm.
One of the findings of this study is that generative AI fraud is actually less prevalent than previously reported. Why do you think that is?
There are several caveats to this statement. First, this is data from two years ago, relatively early in the AI adoption cycle. In terms of GenAI capabilities and usage, it was clearly lower than what we’re seeing today.
Second, our numbers are conservative. To help students answer honestly about sensitive behaviors, we used an indirect survey method. But students still need to know when they are not allowed to use it, which can be difficult.
But even with those restrictions, there are still significant numbers of students. In a recent paper on grade inflation, I showed that when students use AI in their assignments, their overall course grade can be inflated compared to what students actually know.
I think this is indicative of the serious challenges facing higher education institutions. We need to be aware of this issue and address it immediately.
Your research suggests that different fields need to develop different policies for the use of AI and different ways of assessing students. why is that?
Vitaly Galiev (via Unsplash)
Our findings show that students from different disciplines use AI in different ways. Any solutions should be discipline-specific or course-specific.
One response that is currently gaining traction is to move all assessments to a controlled environment. For example, a proctored oral exam or an in-class handwritten exam.
The problem is that this type of assessment only targets a limited group of skills that can be tested in a time-controlled environment. Universities, especially research universities, teach students a much wider range of skills. And some of those skills require long-term engagement with the material.
Going back and forth, writing and coding, or just struggling intellectually is part of how you learn.
If you limit your assessment to narrow settings and very short time frames, you may miss what you are actually trying to teach your students.
You conclude that the solution to AI-based fraud is not a comprehensive university-wide policy. Each department must develop its own policy. How did you come to that conclusion?
One of the problems many teachers face is that GenAI is not easy to detect. You may think that your students’ work is being done by AI, but that may not be the case. Additionally, even if you discover the use of AI, you may end up spending more time proving it, as evidence is easier to collect than in plagiarism cases.
As AI detection software evolves, there are new detectors that can better detect AI-generated text, but there are also AI humanizers (services that make text or code appear to be written by a human), so it’s a cat-and-mouse game. Therefore, it will be a never-ending battle.
Another solution is to ban AI altogether. This is not a productive solution. Students will continue to utilize AI. Some people use it to learn better in a course, to explain content, or to ask difficult questions to the instructor.
Therefore, it is difficult for universities to stay away from the wave of AI adoption and they need to teach students how to use AI. However, what responsible use of AI looks like varies by sector. Writing, coding, problem solving, lab work, and creative work all pose different problems. So a blanket ban probably won’t work.
The study found disparities in the use of generative AI by demographic group members, with low-income, racially underrepresented, and female students less likely to use generative AI. Why is that a concern?
This is even more important than the cheating part. There is less systematic evidence regarding disparities in students’ use of AI tools.
One thing that stands out is the socio-economic and racial disparities in the use of AI, which I think will likely worsen as newer, more expensive models become available.
My concern is that students from wealthier families have access to advanced AI tools with more powerful features and fewer usage restrictions. But students without resources may be limited to clunky and limited free AI tools.
Many employers are interested in graduates with experience using AI tools. Students from higher socio-economic backgrounds may have an advantage not necessarily due to their skills, but rather in being able to pay for those tools. That is a very important element of this study.
Why do you think these findings are important for students?
Using AI for learning can have negative effects. You may create a polished product for your class and even get a good grade, but you won’t develop the skills you were intended to build in the assignment.
There are several experimental studies showing that learning with AI is significantly less efficient and results in less durable skills than without AI. Many students are highly misinformed about how good or bad they are in certain areas and may lack investment in fundamental skills.
And we don’t know what the future holds for AI in the workplace. In such an uncertain world, it is important to understand how AI is impacting education.
I encourage my students to stop and ask themselves questions when using AI. “Can you explain this without tools? Could you do a similar task on your own tomorrow? Did AI help you understand the material better? Or did it primarily help you finish faster?” These simple questions can help students track whether AI is supporting their learning or replacing it.
But it’s challenging. I really care about my students.
Why are your research results important to the university?
There was already a crisis of trust in higher education long before AI. But AI creates another point of criticism for universities. Are universities up to their mission of teaching and assessing student skills in the age of AI?
How a university responds to its challenges shapes public trust in the university. Because when all students achieve excellent grades, it becomes difficult to trust their credentials.
A number of important efforts are already underway by universities to address the impact of AI, including the University of California, Berkeley. However, the evidence in our paper shows that these efforts require more resources and higher priority.
This interview has been edited for length and clarity.
