The Hidden Trade-Off of AI in Higher Education

Artificial intelligence is rapidly becoming part of the everyday learning infrastructure of universities, but the assumption that smarter AI automatically produces better students is beginning to look increasingly fragile. A study titled "Helpful or Harmful? How AI Assistant Intelligence on Online Learning Platforms Shapes Students' Innovative Behavior for Sustainable Education," published in the journal Sustainability by Xinxin Hao, Jiangyu Li, Huan Huang and Bingyu Hao, examines how intelligent AI assistants influence university students not directly, but through the psychological changes they trigger during learning.

More capable AI assistants can strengthen students' confidence in their academic abilities, yet the same convenience can also encourage procrastination by making students feel they can postpone difficult work and rely on AI later. The result is not a simple story of technology improving or damaging learning, but of two competing pathways operating at the same time.

Using a two-phase survey of 503 higher education students in China, the researchers found no significant direct effect of perceived AI-assistant intelligence on innovative behavior. Instead, what mattered was whether AI increased academic self-efficacy or academic procrastination, and whether students had enough AI literacy to keep the positive pathway stronger than the negative one.

AI's Educational Value Depends on What It Does to the Learner

AI assistants can provide immediate explanations, personalised feedback, resource recommendations and task-specific support, often at a speed traditional teaching systems cannot match. Universities increasingly use AI-enabled platforms to support coursework, information retrieval, writing, exam preparation and self-directed learning.

However, this convenience creates a deeper educational question. If AI removes uncertainty, simplifies complex tasks and provides instant assistance, does it help students become more capable,or simply reduce the amount of effort they need to invest? Previous research reviewed by the authors points in both directions, associating generative AI with stronger learning efficiency and knowledge building, but also with overreliance, weaker self-directed learning, academic dishonesty and reduced deep processing.

The study moves beyond the usual measures of AI adoption, frequency of use or student acceptance. Instead, it focuses on perceived AI-assistant intelligence: whether students experience an AI system as capable, knowledgeable, reliable, adaptive and able to solve learning problems effectively. That distinction matters because two students may use the same platform equally often but respond differently depending on how capable they believe the assistant to be.

Drawing on Social Cognitive Theory, the researchers examine two psychological responses. Academic self-efficacy refers to students' confidence in their ability to learn, complete difficult tasks and handle academic demands, while academic procrastination refers to delaying work despite knowing that the delay may have negative consequences. The study tests whether these two responses explain how AI influences innovative behavior, defined as generating, promoting and implementing useful new ideas in learning or research.

The Same AI Can Act as a Scaffold and a Shortcut

The most notable result is that intelligent AI appears to push students in two directions at once. Perceived AI-assistant intelligence was positively associated with academic self-efficacy, suggesting that capable AI can reduce uncertainty, provide useful guidance and help students feel more competent when confronting difficult tasks.

Self-efficacy can support innovative behavior. Students who believe they can master difficult material are more likely to persist, experiment with alternative approaches, integrate new knowledge and propose new ideas. In the study, the indirect effect of AI-assistant intelligence on innovative behavior through academic self-efficacy was positive and statistically significant.

However, the same AI capability was also positively associated with procrastination. When students know that an intelligent assistant can quickly generate explanations, summaries or solutions, they may feel less pressure to start difficult tasks immediately. The technology becomes not only a learning scaffold but also a psychological safety net that makes postponement easier.

The second pathway works against innovation. Procrastination reduces the time and cognitive resources available for exploration, reflection, refinement and creative problem-solving. The study found a negative indirect effect of AI-assistant intelligence on innovative behavior through academic procrastination, and that negative pathway was larger in magnitude than the positive self-efficacy pathway.

This helps explain why the researchers found no significant direct relationship between AI-assistant intelligence and innovative behavior. The positive and negative mechanisms may partially cancel each other out. In other words, smarter AI does not automatically make students more innovative because technology can simultaneously increase confidence and encourage delay.

AI Literacy Is Emerging as the Critical Protective Skill

The study's most policy-relevant finding is that AI literacy changes the balance between these competing effects. The researchers define AI literacy as the ability to understand, use, evaluate and appropriately apply AI, including recognising limitations, judging the reliability of outputs and using the technology responsibly.

Among students with higher AI literacy, intelligent assistants were more strongly associated with academic self-efficacy and much less strongly associated with procrastination. At high AI-literacy levels, the positive indirect effect through self-efficacy was statistically significant at 0.086. At low AI literacy, the same pathway was weak and not statistically significant.

The pattern was even more pronounced for procrastination. Among students with low AI literacy, the negative indirect effect through procrastination reached −0.158, while among students with high AI literacy it was small and statistically non-significant. The implication is that capable AI becomes much more educationally useful when students know how to treat it as a tool rather than a substitute for effort.

This shifts the debate away from whether students should be allowed to use AI and toward whether they possess the skills needed to use it productively. A student who can question AI outputs, verify information, understand system limitations and integrate suggestions into their own reasoning is more likely to gain confidence without surrendering agency. A student with weaker AI literacy may instead use convenience to avoid difficult cognitive work.

For universities, that makes AI literacy more than a technical skill or academic-integrity issue. It becomes a form of educational risk management. Institutions investing heavily in intelligent platforms while neglecting student capability may unintentionally amplify dependence, procrastination and passive learning.

Higher Education Needs Better AI Design, Not Just More AI

Universities should not judge AI systems only by how fast they respond or how much content they can generate. The researchers argue that intelligent assistants should be designed as learning scaffolds, using step-by-step prompts, explanatory feedback, study-plan reminders and process tracking rather than simply supplying answers.

AI literacy also needs to move into the core curriculum. Students should learn how to formulate effective prompts, verify generated content, identify limitations and use AI feedback to support, not replace, their own reasoning. The study suggests that these competencies can strengthen the benefits of intelligent assistance while reducing the risks of overdependence and delay.

Teachers, meanwhile, face a parallel challenge in course design. Assignments increasingly need to preserve independent judgment while allowing students to benefit from AI-supported learning. Asking students to critique AI-generated answers, explain how they used AI feedback or compare machine suggestions with their own reasoning may become more important than simply trying to prohibit the technology.

For developing countries and education systems in the Global South, the study suggests that investment in AI infrastructure alone may not deliver the expected gains in educational quality or innovation. If access to AI expands faster than AI literacy, institutions could reproduce a new form of digital inequality in which some students use AI to extend their capabilities while others become increasingly dependent on it.

The research relies on self-reported measures, meaning social-desirability and recall biases cannot be entirely excluded. Although the two-phase design improves the study's robustness, academic self-efficacy, procrastination and innovative behavior were measured together in the second phase, limiting the strength of causal claims.

The sample is also drawn from Chinese higher education students using online learning platforms such as Chaoxing Learning Pass and MOOCs. The authors therefore acknowledge that the findings may not automatically generalise across countries, disciplines, educational levels or types of AI tools. They call for future studies using learning-platform logs, AI usage records, teacher evaluations, longitudinal tracking and experimental designs.

Regardless, the wider significance is hard to dismiss. As AI becomes more powerful, the central educational challenge will not be ensuring that every student has access to an intelligent assistant. It will be ensuring that students remain active learners once that assistant becomes capable of doing increasingly more of the work.