AI lacks ‘deep understanding’ of material, researchers find
ChatGPT can pass a college-level engineering course, earning a B grade with “minimal effort,” researchers at the University of Illinois at Urbana-Champaign recently found.
However, the study also revealed that while AI excels at “pattern recognition,” it lacks “deep understanding” of course material.
One of the researchers that led the study told The College Fix via email that AI struggles with complex questions likely because it hasn’t been trained on enough similar examples.
After doctoral student Gokul Puthumanaillam and Professors Timothy Bretl and Melkior Ornik programmed a robot to complete an undergraduate aerospace engineering course, they published their findings in a paper titled “The Lazy Student’s Dream: ChatGPT Passing an Engineering Course on Its Own.”
“The LLM achieved a B-grade performance (82.24\%), approaching but not exceeding the class average (84.99\%), with strongest results in structured assignments and greatest limitations in open-ended projects,” the paper states.
The researchers evaluated 115 course deliverables “using ChatGPT under a ‘minimal effort’ protocol that simulates realistic student usage patterns.”
They used various assessment platforms, including multiple-choice questions, Python programming tests, and detailed analytical writing assignments.
“Report analysis indicates stronger performance in methodology description and result presentation but weaker performance in critical analysis and design justification,” the project’s website states.
ChatGPT performed the worst in programming assignments, as it employed rigid reasoning and was unable to adapt to the problems’ nuances.
Professor Ornik, who helped lead the project, told The Fix AI ultimately lacked a deeper “understanding” of course material that should be demonstrated by top students.
He clarified “that the terminology of ‘pattern recognition’ vs. ‘deep understanding’ in the paper speaks from the perspective of human learners.”
MORE: AI is killing real education, Hillsdale student warns
“For humans, we have a decently good, if only intuitive, idea of what ‘pattern recognition’ means in the context of a class like mine — essentially solving questions that were already seen, but ‘with different numbers,’” the professor said.
“On the other hand, a deeper understanding implies a capability to apply the learned material to questions that have never been previously seen,” he said.
AI’s poorer performance on more complex issues likely stems from a “lack of sufficient training data,” as the length and originality of such questions make it challenging to link them to AI’s existing dataset.
“This challenge is certainly not unique to ChatGPT: recognition and response to ‘out-of-distribution’ tasks – that is, tasks that are significantly different from those seen before — remains a significant challenge and area of research in AI,” he said.
Turning to the broader impact of AI, the professor said “there are certainly elements of the labor market which will be affected by increased automation, the same way that artisans were affected by the Industrial Revolution or that human computers of the 1930s-1970s were replaced by automated calculations.”
He also told The Fix that programming and software development jobs are already experiencing shifts due to AI.
However, while AI may excel in certain domains, it cannot take over fields like humanities, social sciences, and art. He said that these disciplines are inherently tied to the complexities of human experience and emotions.
The professor also said it is “crucial to our nation that our federal and state governments strongly support America’s educational and research enterprises across the board.”
“Having seen the incredibly nimble efforts of some other countries, falling behind now may lead to permanent deterioration of America’s preeminence in higher education and our technological supremacy,” he said.
This study comes as more students and professors admit to using some form of AI to complete or grade course work, Inside Higher Ed reported.
“A significant number of students admit to using generative artificial intelligence to complete their course assignments (and professors admit to using generative AI to give feedback, create course materials and grade academic work),” the outlet reported.
The College Fix reached out to the University of Illinois at Urbana-Champaign via email to ask about its response to the findings and whether it plans to make any changes to its courses as a result. The school did not respond.
MORE: As AI replaces learning, concerns grow: ‘You just don’t really have to think’
IMAGE CAPTION AND CREDIT: Human using AI chat bot; Supatman/CanvaPro