When ChatGPT launched in late 2022, it took the world by storm, gaining over 100 million users in its first month. Its impact has been felt across industries, and education is no exception. AI tools are opening doors to new ways of teaching and learning, but they are also raising tough questions about how we test and prepare students.
Despite the buzz, little research has explored how large language models (LLMs) like GPT-3.5 and GPT-4 affect higher education. A team at EPFL recently set out to change that. They studied 50 STEM courses, from computer science to biology, to see how well AI could handle traditional assessments.
The findings were surprising. GPT-4 answered 65.8% of questions correctly on average, and at least one approach worked for 85.1% of questions. Even simple prompts—those anyone could use without special knowledge—produced impressive results. “No one expected the models to perform this well across so many subjects,” the researchers said.
The risks of relying on AI
While these tools are powerful, they come with risks. The researchers point to two main concerns. First, assessment vulnerability: traditional tests may no longer be a reliable way to measure student knowledge. Second, educational vulnerability: students might use AI to skip the hard work of learning, building weaker foundations for advanced skills later.
“If students rely too much on these tools, they might miss the chance to fully grasp core concepts,” the researchers warned. This raises bigger questions about what students should be learning and how education can adapt to take advantage of AI while avoiding its pitfalls.
How schools can adapt
One way forward is to redesign assessments. Instead of focusing on isolated skills, tests should require students to combine ideas and solve complex problems. “Project-based learning could help,” the researchers said. “It’s harder for AI to handle and better for students in the long run.”
AI tools, meanwhile, are getting better fast. The study used general-purpose models, but newer, specialized tools—such as those tailored for math—are even more capable. If the research were repeated today, the success rates would likely be higher.
A lesson from history
The rise of AI echoes earlier debates about calculators. When they first appeared, some worried that students would stop learning math. Today, calculators are common tools, used alongside lessons to deepen understanding. AI may follow a similar path, complementing education rather than replacing it.
The EPFL team stresses the need for collaboration between educators and AI developers. Practical solutions can help students and teachers adapt while reducing risks. “This is just the beginning of a big shift,” they said. “Education must change to use AI wisely while keeping its core purpose intact.”
As AI continues to improve, so must the way we teach and learn. If schools can strike the right balance, AI could enhance education rather than disrupt it—helping students prepare for an increasingly complex world.
