Generative AI and Humans Join Forces to Reshape Higher Education

Generative artificial intelligence has moved from the margins of academic curiosity to the center of a debate that will define the next generation of universities. In an editorial published in Frontiers of Digital Education, Fei Wu of Zhejiang University’s College of Computer Science and Technology and Jingyuan Chen of the university’s College of Education argue that the arrival of large language models such as DeepSeek represents nothing less than a paradigm shift for higher education. Their central claim is deceptively simple: the future of teaching and learning will not be a contest between machines and people but a structured collaboration between the two. Rather than treating generative AI as a threat to academic integrity or a shortcut that undermines effort, the authors frame it as a partner whose strengths complement, rather than replicate, human cognition. The argument arrives at a moment when institutions worldwide are scrambling to write policies for tools that students are already using daily.

The technical foundation of this shift lies in the architecture of large language models. These systems are trained on vast text corpora and learn statistical patterns of language that allow them to generate coherent explanations, solve problems, draft essays, and answer questions across virtually every academic discipline. Unlike earlier educational software, which followed rigid, pre-programmed rules, generative models produce novel responses tailored to each prompt. Wu and Chen point to DeepSeek, an open-source large language model developed for Chinese education research, as an example of how accessible, high-capability AI can empower learners across an entire society rather than remaining locked behind expensive proprietary systems. Open-source approaches matter because they allow universities to inspect, adapt, and integrate models into their own platforms without depending entirely on a handful of commercial vendors, a consideration that touches on cost, sovereignty, and pedagogical control.

One of the most concrete applications the editorial highlights is cognitive diagnosis, the task of figuring out precisely what a student knows and where their understanding breaks down. A companion study by Dong, Chen, and Wu introduces a model-agnostic framework that uses large language models together with the Solo taxonomy, a classification scheme for levels of understanding, to diagnose learners’ cognitive states. Related work by Ma and colleagues demonstrates that large language models can act as zero-shot cross-domain diagnosticians, meaning they can assess knowledge in subject areas they were never explicitly trained to evaluate. The technical significance is considerable: traditional diagnostic systems required large amounts of labeled response data for each specific domain, whereas language-model-based approaches can generalize across subjects. For universities, this means assessment could evolve from infrequent, high-stakes examinations toward continuous, fine-grained feedback that identifies misconceptions the moment they appear.

Mathematical reasoning remains a demanding benchmark for these systems, and the editorial’s reference list acknowledges the challenge directly. Liu and colleagues present MathEval, a comprehensive benchmark designed to evaluate how well large language models handle mathematical reasoning across difficulty levels and topic areas. Mathematics exposes both the power and the fragility of generative models: they can often produce elegant solution steps, yet they may also fail in ways that look convincing to a novice. This is precisely why the human-in-the-loop framing matters. In a collaborative paradigm, the AI proposes, explains, and drills; the instructor verifies, contextualizes, and corrects. The division of labor plays to the strengths of each side, with machines providing tireless availability and breadth while humans supply judgment, accountability, and the ability to recognize when an answer is plausible but wrong.

Personalization is the second pillar of the paradigm shift. Li, Chai, and colleagues explore how detailed student portraits, structured profiles built from learning behavior, performance data, and preferences, can drive personalized learning at scale. Tu, Chen, and Huang examine the mechanisms by which generative AI empowers personalized learning, along with the challenges and pathways institutions must navigate. Combined with diagnostic models, such portraits allow an AI tutor to adapt explanations to an individual student’s level, pace, and prior knowledge in ways a single lecturer facing hundreds of students never could. The editorial situates these capabilities within the characteristics of Generation Z, a cohort that has grown up with interactive digital environments and often expects immediate, individualized feedback. For these students, a static textbook and a weekly lecture are no longer the default interface to knowledge; conversational, adaptive systems are.

The editorial also engages with the geopolitical and cultural dimensions of educational AI. A commentary by Wang argues that DeepSeek’s Chinese-style innovation reinforces China’s educational confidence, while Wu’s own companion piece describes DeepSeek as a vehicle for global education empowerment for a whole society. The underlying point transcends any single model: when capable AI systems are open and affordable, they can reach institutions and learners who have historically been excluded from elite educational resources. The authors suggest that the paradigm shift is therefore not only pedagogical but also structural, potentially redistributing access to high-quality tutoring, feedback, and content creation. At the same time, the emphasis on open models reflects a recognition that educational AI embedded in one nation’s values and curricula will not automatically transfer to another, making locally adaptable systems essential.

None of this enthusiasm is presented without caution. The editorial and the studies it accompanies repeatedly stress that generative models hallucinate, produce confident errors, and can be misused for academic dishonesty. Krause, Panchal, and Ubhe assess the transformative impact of generative AI on higher education and document how institutions are wrestling with questions of assessment design, skill development, and the changing role of the lecturer. If an AI can draft an essay in seconds, universities must rethink what written assignments are meant to measure and how to evaluate the process of thinking rather than only the final product. The collaborative paradigm implies that assessment itself must evolve, rewarding students for how effectively they direct, question, and verify AI output, a skill the authors implicitly treat as a new form of literacy.

Zhu, Wang, and colleagues extend the discussion to K-12 education, surveying current trends and future prospects for large-scale foundation models in schools. Their inclusion in the editorial’s reference frame signals that the paradigm shift begins long before university. Students arriving on campus will already be accustomed to AI-assisted learning, and higher education institutions that fail to adapt risk offering an experience that feels outdated to their own applicants. The technical trajectory supports this expectation: foundation models are becoming multimodal, capable of processing text, images, and diagrams, and increasingly efficient enough to run in settings with limited computing infrastructure. Wu and Chen’s argument is that universities should lead this integration deliberately, shaping tools to pedagogical goals, rather than reacting after the fact to technologies students have already adopted on their own.

What emerges from the editorial is a vision of higher education in which the lecturer’s role is elevated rather than diminished. Freed from repetitive explanation and routine grading support, instructors can focus on mentorship, critical thinking, ethical reasoning, and the design of learning experiences that machines cannot supply. The AI handles scale; the human handles meaning. Wu and Chen describe this as a paradigm shift in the strict sense, a change in the underlying assumptions of the field rather than an incremental tool upgrade. Their position as editors of the journal, disclosed and excluded from the peer-review of their own piece, underscores how seriously the academic publishing community is taking the question. As generative models continue to improve, the institutions that thrive will likely be those that treat the human-AI partnership not as a compromise to be managed but as the defining educational relationship of the coming decades.

Subject of Research: Human-AI collaboration and large language models in higher education

Article Title: Collaboration of Generative AI and Human: Paradigm Shift for Higher Education

Article References: Wu, F., & Chen, J. (2025). Collaboration of Generative AI and Human: Paradigm Shift for Higher Education. Frontiers of Digital Education, 2(2), Article 24. https://doi.org/10.1007/s44366-025-0061-z

Image Credits: AI Generated

DOI: 10.1007/s44366-025-0061-z

Keywords: generative AI, large language models, higher education, DeepSeek, personalized learning, cognitive diagnosis, educational technology, paradigm shift, open-source AI, assessment, Generation Z, Zhejiang University