When students use artificial intelligence to learn new subjects, their underlying motivations will determine whether they actually retain the information or just retrieve surface facts. A recent experiment found that learners who were instructed to focus on personal understanding absorbed more knowledge than those who were instructed to outsmart their peers. Students who focused on competing with other students experienced greater anxiety and approached learning tasks with shallow strategies. The study was published in Applied Cognitive Psychology.
In the field of educational psychology, experts rely on a concept called achievement goal theory. This framework defines the internal reasons why individuals put effort into their studies. Students may read textbooks to truly understand how the universe works, or they may read textbooks just to secure a passing grade.
Instructors often form these internal goals by emphasizing various measures of success in the classroom. This environmental influence is known as goal structure. Teachers can change this structure through task design, grading methods, or the words they use in instructions.
The structure of mastery objectives allows students to focus on personal progress. In this environment, the main objective is skill development and deep understanding. Criteria for success can be absolute or personal. That is, students simply compare their current knowledge to a past baseline.
Conversely, the performance goal structure encourages students to prove their abilities by outperforming their peers. The focus is entirely on managing how others perceive your abilities. Success is based on a normative curve, and getting good grades means beating your classmates and avoiding unfavorable evaluations.
Past educational research has shown that mastery environments tend to foster deep learning activities. Students are more likely to show persistence and seek true understanding. Performance environments often encourage shallow learning habits in which learners prioritize appearance of competence over actual understanding.
Artificial intelligence tools are now rapidly penetrating mainstream education. Generative chat programs offer a highly personalized learning method that acts like an on-demand tutor. However, early studies of these text generators have shown very different results.
Some studies have shown that these programs improve academic performance. Other studies have found that performance decreases when students rely on software.
Laura Schmidt, an education researcher at Germany’s Ruhr University Bochum, wanted to understand the reasons behind this variation. People may think that students’ technical skills regarding rapid engineering determine their overall success. However, Schmidt suspected that students’ personal motivations could shape their technical approach to software.
Together with his colleagues Niklas Obergassel and Julian Roel from the University of Münster, Schmidt designed a study to see whether the traditional concept of goal structure still applies in the age of automation. Because large-scale language models respond directly to user instructions, students’ internal motivations can immediately determine the types of questions they ask during a learning session.
To test this idea, the research team recruited 104 university students and conducted a supervised online experiment. Participants were told to use ChatGPT to learn four specific social psychology concepts. These included ideas such as the mere exposure effect, which explains the tendency for people to develop a preference for an object simply because they are familiar with it.
Students were given a test that measured their existing knowledge of social psychology concepts. They were then divided into two slightly unequal groups. The first group received instructions to create a mastery goal structure.
These participants were told that the purpose of using the software was to expand their personal knowledge. The researchers instructed the students to deeply understand the concepts and pointed out that the learning session would be considered successful if they felt they had learned a lot.
The second group was placed in a performance goal structure. The researchers told the group that their real purpose was to outshine the other participants in the study.
The instructions clearly stated that you should try to gain more and better knowledge than your colleagues. The session is considered a success if you perform better than other participants on the final test. To ensure that students understood these specific instructions, I asked everyone to rephrase the instructions in writing before we began.
Both groups then worked with a text generator for 20 minutes to learn about the four concepts. The researchers reminded the students of a specific goal every five minutes. Participants were not allowed to take personal notes.
After the learning phase, participants reported their intrinsic motivation and emotional state. They then took a final test consisting of 12 open-ended questions designed to measure both conceptual knowledge and deeper understanding. The researchers also collected chat records and categorized all the prompts that students entered into the software.
The results demonstrate that how tasks are framed can dramatically change the way people interact with artificial intelligence. Students in the mastery group gained significantly more conceptual knowledge than students in the performance group. They provided better definitions and explanations of core psychological concepts.
Chat logs reveal how the performance group’s desire to appear smart disrupted their learning process. These students wanted to be better than their peers, so they asked the chatbot for small details that might sound impressive in a social setting.
They frequently requested the specific name of the researcher who coined a particular term or the exact year in which a famous study was published. We asked for the titles of related books and magazines. These facts may help a person appear knowledgeable in casual conversation, but they contribute little to the actual understanding of core psychological principles.
During the final test, the performance group included a high percentage of these unimportant details in their written answers. This phenomenon is known as criterion adherence, and learners attempt to meet external assessment standards in a superficial manner, ignoring the actual learning objectives.
In contrast, the proficient group used their time more efficiently. Rather than searching for obscure trivia, they turned to software for memorization aids to help them retain content better. They actively sought to build a structural understanding of information.
Beyond differences in knowledge acquisition, performance goal instructions placed an emotional burden on users. Students in this group reported increased levels of pressure and tension during assignments.
The fear of possible failure put a lot of stress on the learning process, as they were worried about external evaluations and normative test standards. They experienced a general increase in anxiety. Proficient groups avoided this emotional burden by focusing solely on their own intellectual progress.
The researchers also measured positive aspects of intrinsic motivation, such as feelings of enjoyment, interest, and autonomy. Surprisingly, the differences between the two groups on these positive measures were not found to be statistically significant. The researchers noted that both groups showed very high baseline levels of mastery orientation.
Many college students naturally want to learn, and this internal desire can coexist with external performance pressures. Although performance instruction did not eradicate their natural curiosity, it did create negative pressure that hindered their progress.
We found similar results when testing deeper understanding. Differences between groups on deep understanding tasks that required students to apply psychological concepts to new scenarios were not statistically significant. The researchers believe that a 20-minute session may not have been enough time to achieve a high level of understanding. Learning how to properly apply complex psychological phenomena will probably require a long period of study.
Future research should address several limitations of this experiment. Participants likely had varying levels of experience with artificial intelligence. Most students used the program to request basic instruction, but few asked the software to generate practice questions or simulate assessments. Learners’ previous technical skills can easily change the way they pursue different goals.
The researchers also point out that this particular experiment only tested declarative knowledge, which relies on rigorous factual and conceptual understanding. Asking for peripheral details is an easy way to disguise your ability when studying basic texts. However, this strategy may not translate well to learning problem-solving skills or advanced mathematics, where superficial details alone provide no benefit. Future research could test different types of learning materials to see if the effects persist.
Finally, the researchers had no way to track how deeply students thought about the text generated by the software. Without eye-tracking technology or think-aloud protocols, it is difficult to know whether the proficient group actively read the text more thoroughly or whether their higher test scores resulted purely from differences in the initial prompt.
Despite these unknowns, the lesson for educators remains simple. Encouraging students to compete with each other takes their focus away from mastering the content. When teachers incorporate artificial intelligence into their classrooms, they should avoid structuring tasks around comparisons with peers.
The study, “AIming High: Do Goal Structures Matter in Learning With ChatGPT?”, was authored by Laura Schmidt, Niklas Obergassel, and Julian Roelle.
