Why timing matters for research into students’ GenAI misuse

A recent paper published in Science in May 2026 reported generative artificial intelligence use and misuse across 95,000 undergraduates from 20 research universities in the United States. It revealed stark differences in GenAI-assisted cheating between disciplines, with cheating among biology students identified as particularly low.

The authors claim that: “Our findings offer large-scale, discipline-specific evidence of GenAI use and cheating to better understand how students engage with this technology.” A casual reader of this study might then infer that there is something unique about biology students, concluding that they are not using GenAI to cheat as much as students in other disciplines such as computer science, business or journalism. 

• Trying to decide what, where and when to publish research?
• Stop investigating, start teaching
• The two key steps to promoting responsible use of LLMs

However, upon a closer read, the authors report that only 5 per cent of biology students use GenAI-assisted cheating when you ask them if they have submitted AI-generated content as their own work in class, knowing it may not be allowed.

So, this would include submitting a final paper, a lab report, a discussion board post or even the answer to an open-ended question on a homework assignment. What it wouldn’t include is using GenAI to answer any closed-ended question that has a specific correct answer: homework assignments, in-class polling questions or exam questions – in other words, most assessments that are used in biology courses at large research institutions, such as the ones sampled from in this study. 

But for computer science courses, business courses and journalism courses, students are likely doing more “copy and paste” of AI-generated content as code, case analyses or essays. 

An alternative explanation could be that the observed discipline-specific differences are simply an artefact of the way that the survey question was asked. Perhaps these differences are actually demonstrating which disciplines have closed-ended multiple choice assessments and which disciplines ask students to generate their own products, rather than a heterogeneity of GenAI misuse among disciplines. 

Plus, the data are two years old, which matters when we are talking about GenAI use. There are two major changes that have occurred with GenAI, the first midway through their data collection and the second after their data were collected. 

In May 2024, AI stopped being primarily a text chatbot and became multimodal, allowing users to interact with other forms of media such as images and audio. In September 2024, tasks that educators believed were AI-proof started becoming much more vulnerable to AI because of shifts in AI’s reasoning models. The GenAI that students are using or misusing in this study in 2024 is not the same GenAI that anyone is using right now in 2026. 

Research takes time, especially for large-scale studies, and the peer-review publication process is notoriously slow. Researchers studying GenAI use and misuse may need to consider whether their primary goal is to merely document student behaviour at a particular point in time, or to influence current teaching or policy decisions. While publishing in a journal such as Science is considered the gold standard for many academics, we argue that because GenAI is such a rapidly evolving technology, a two-year gap between data collection and publication makes these data immediately obsolete for current teaching or policy decisions. 

This is a problem that can be solved in several ways: journals can fast-track time-sensitive research or researchers can choose to share their research outside traditional academic journals. Many researchers have started posting preprints on free online servers to try to share findings earlier (eg, bioRxiv), but typically the official publication that has gone through the lengthy peer-review process is what garners more attention (and status) among educators and administrators. 

If researchers studying GenAI continue to use traditional publishing approaches, then they have a responsibility to be overtly transparent about the time of data collection and what has changed in the technology since then. This information needs to not be buried in the methods or limitations sections, but clearly stated in the title or abstract so the casual reader can immediately access it.

Educators themselves have a responsibility to pay close attention to the time of data collection and exactly how the data were collected or else they may inadvertently use outdated or inaccurate evidence related to student use of GenAI. In fact, for educators reading this Science paper with an eye on changing teaching or policy, we encourage caution in using this as evidence, since these are now historical data.  

Despite the large cross-disciplinary sample size and being published as a policy article in one of the most prestigious journals, the rapid evolution of GenAI over the past two years and the narrowness of the survey question about its misuse significantly weaken the conclusions that we can derive from this study, and any discipline-specific policy implications.

Benjamin G. Chan is a PhD student and Sara E. Brownell is a president’s professor, both in the School of Life Sciences at Arizona State University. 

If you would like advice and insight from academics and university staff delivered direct to your inbox each week, sign up for the Campus newsletter.