What If the Biggest AI Divide on Campus Has Nothing to Do With Access?

Generative AI has made sophisticated digital tools easier to reach, but access alone is not producing equal competence. Research from Nanjing University suggests that students can share the same technological environment while developing very different abilities to use, assess and question AI, depending partly on how their academic programmes are structured.

The study, "Mapping the AI Divide in Undergraduate Education: Community Detection in Disciplinary Networks and Survey Evidence," was authored by Xiaomin Ji, Wei Si, Liwen Zhang, Jiangli Zhu and Ke-Ke Shang and published in Big Data and Cognitive Computing. It combines curriculum network analysis with survey evidence from 301 undergraduates to examine how disciplinary environments relate to artificial intelligence-generated content, or AIGC, literacy.

The results point to a form of educational inequality that is less visible than the traditional digital divide. Students in science-dominant curricular networks reported stronger ability to use and evaluate AIGC than students in humanities and social sciences, while differences in motivation and perceived skills accounted for part of the gap. The pattern does not establish causation, but it raises a difficult question for universities: can equal access to AI coexist with unequal opportunities to become competent users of it?

The divide has moved from devices to capability

For much of the digital era, inequality was measured through access to computers, broadband and online services. Generative AI alters that framework because natural-language interfaces substantially reduce the technical knowledge required to begin using advanced systems.

The study argues that inequality can therefore shift from possession of technology to the quality of interaction with it. Two students may have access to the same AI platform, yet differ greatly in their ability to formulate useful requests, test outputs, identify weaknesses and decide when generated material can be trusted.

AI literacy is not simply familiarity with a tool. The researchers define AIGC literacy around two practical capacities: using AI to solve problems and evaluating generated outputs for quality, limitations, appropriateness and risk.

Such capabilities are likely to develop unevenly when disciplines expose students to technology in different ways. A programme that repeatedly integrates computational tools into assignments, problem-solving and assessment creates a different learning environment from one where AI is peripheral, optional or treated mainly as an object of debate.

Nanjing University provides a particularly revealing case because it already had substantial investment in AI education. AI-related courses were incorporated into undergraduate training plans from 2021, followed by a broader curriculum reform in 2023 that expanded foundational AI learning, interdisciplinary education and advanced courses.

The research examines inequality in a setting where basic technological access is comparatively strong. In such an environment, differences in exposure, curriculum and confidence become easier to distinguish from simple infrastructure gaps.

Curriculum structure is sorting students into different AI experiences

Instead of assigning students to broad categories such as STEM or humanities, the researchers mapped departments according to the courses they shared. The final network contained 29 departments and produced four curricular communities: science-dominant, science-peripheral, science and social sciences, and humanities and social sciences.

Survey results showed a clear gradient across those communities. Students in the science-dominant group averaged 4.30 for using AIGC and 4.25 for evaluating AIGC, while students in humanities and social sciences averaged 3.32 and 3.47. The other two communities fell between those groups.

Motivational measures followed a similar pattern. Science-dominant students recorded the highest averages for value efficacy, skill efficacy and usage efficacy, while humanities and social sciences students recorded the lowest. The differences were statistically significant across all three measures.

These results do not show that science disciplines inherently create better AI users. Students enter programmes with different backgrounds, interests and prior experience, while course structure captures only part of the educational environment. The study itself is cross-sectional, so it cannot determine whether curriculum produces the differences or whether existing differences also influence disciplinary choice.

The association is still important because it places curriculum architecture inside the debate over AI inequality. If repeated exposure to technical tools, interdisciplinary content and applied problem-solving is concentrated in certain parts of a university, AI capability may develop unevenly even when software access is widespread.

Humanities and social sciences also bring capabilities that technical training alone cannot supply, including interpretation, ethical reasoning, contextual judgement and scrutiny of social consequences. The problem arises when those strengths are developed without equivalent opportunities to build operational confidence with AI systems.

Skill confidence explains part of the gap

The researchers also examined motivational efficacy, divided into perceived value, skill efficacy and usage efficacy. This part of the analysis asks whether students' confidence and motivation help account for the relationship between curricular environment and AIGC literacy.

Skill efficacy produced the strongest and most consistent mediating effect. Students who felt more capable in information processing and technology-related tasks also tended to report stronger ability to use AIGC and assess its outputs.

The result shifts attention away from exposure alone. Universities can introduce AI tools into classrooms without necessarily giving students the confidence to work with them independently, troubleshoot problems or judge whether an output is reliable.

Usage efficacy produced a more complicated result. In the main model, it did not significantly mediate the relationship between curricular community and AI evaluation, but it became significant under alternative specifications that treated communities differently and accounted for clustering within schools.

This sensitivity is analytically useful because it prevents a simple conclusion that usage confidence has no role. It also shows why institutional data should be interpreted carefully when students are nested inside departments, programmes and disciplinary cultures.

Differences across admission cohorts add another layer. Students from the 2021 cohort reported higher scores for using and evaluating AIGC than students entering in 2024. The authors caution, however, that the groups experienced different curricular conditions, including the 2023 reform, so the pattern cannot be attributed only to spending more time at university.

The broader takeaway is that AI literacy may develop through repeated practice rather than one-off training. Students are more likely to become capable users when AI appears progressively across the learning cycle and connects to the kinds of problems their disciplines actually require them to solve.

Universities may need to redesign AI education by discipline

The study proposes interdisciplinary laboratories in which technically trained students contribute development expertise while humanities and social sciences students contribute scenario design and ethical assessment. It also recommends "AIGC + discipline" programmes that bring technical applications into non-technical fields while adding ethical and evaluative components to STEM pathways.

This approach differs from a standard university-wide AI course. A common introductory module can establish basic knowledge, but students also need to understand how AI changes the reasoning practices, professional standards and risks of their own fields.

A journalism student may need strong verification habits and an understanding of synthetic media. A business student may need to examine AI-assisted analysis and decision support. An engineering student may require deeper technical competence alongside stronger attention to ethics, accountability and the limits of automated systems.

For university leaders, conventional indicators of digital readiness may therefore be inadequate. Counts of licences, platforms, workshops or enrolled students say little about whether learners can detect unreliable outputs, apply AI to substantive problems or recognise when human judgement should take precedence.

The findings are also relevant to developing countries, but they should not be transferred mechanically. Nanjing University is a leading institution with substantial AI capacity. In universities where connectivity, devices, computing infrastructure and faculty expertise remain uneven, access inequalities may operate alongside the capability differences identified here.

The study covers one institution, some departments had small numbers of respondents, the measures rely on self-reported survey data and the cross-sectional design cannot establish causal relationships. The authors also note that access to AI-related information remains underdeveloped in their analytical framework.

Upcoming research will need to test the model across technical universities, liberal arts institutions, teaching-focused colleges and institutions with fewer digital resources. Longitudinal studies could also examine whether curriculum reforms actually change AI literacy over time, rather than simply capturing differences that already exist among students.