One-size-fits-all AI policies could fail universities and students

Generative AI is now part of everyday student work, but universities risk weakening academic integrity and student learning if they treat all AI use as the same behavior, according to a new analysis by researchers from the Universidad Nacional de Colombia.

The entry, titled Discipline-Sensitive Generative AI in Higher Education, was published in Encyclopedia and it argues that generative AI in higher education should be understood as a situated educational practice shaped by discipline, assessment design, AI literacy, study habits and academic integrity norms.

One-size-fits-all AI rules miss how disciplines define learning

The analysis addresses a fast-moving challenge for universities as students increasingly use AI systems to generate text, code, explanations, feedback-like responses, images, summaries and study support. Students now turn to these tools to revise drafts, translate passages, test code, brainstorm topics, compare examples, prepare assignments and rehearse explanations before class. The authors argue that these practices cannot be judged properly through a single institutional rule that treats AI use as either acceptable or unacceptable in all contexts.

The same AI-supported action can have different academic meanings depending on the course and discipline, the study claims. Asking for a simpler explanation before class is not the same as submitting a generated answer in a closed assessment. Using AI to test code is not the same as outsourcing the core logic of a programming assignment. Brainstorming visual ideas in a design course is not the same as concealing machine-generated work as original creative production.

The authors reject both broad alarm and unchecked enthusiasm. Generative AI can help students when it supports explanation, practice, feedback, comparison and revision. It can also damage learning when it replaces reading, reasoning, evidence-checking, disciplinary judgment or independent academic performance. The paper therefore calls for a shift from generic AI rules to discipline-sensitive guidance that connects tool use to the purpose of each academic task.

The debate, as the analysis suggests, hinges heavily on disciplinary culture. Academic fields differ in what they value, how they define evidence, how students demonstrate knowledge and what forms of originality matter. These differences shape whether AI assistance supports learning or weakens the evidence of student contribution.

In humanities and many social sciences, writing is often part of the thinking process itself. Students are expected to read, interpret, weigh evidence, build arguments and develop a voice. AI may help them compare explanations or organize early ideas, but it can also interfere with source-based reasoning and the slow formation of interpretation. The integrity issue is not only whether AI produced sentences. It is whether the final work reflects the student's own engagement with evidence.

In computer science and engineering, generative AI can write code, suggest fixes, explain errors and compare approaches. That creates a different challenge because professional computing already involves documentation, libraries, debugging tools and collaboration. The authors argue that students still need to understand algorithms, design choices, testing logic and code behavior. Assessments in these fields may need to focus less on final code alone and more on explanation, testing, revision history and the ability to defend a solution.

Health sciences raise sharper concerns because plausible AI output can still be unsafe. A system may produce a fluent clinical explanation while omitting risk, misrepresenting evidence or ignoring context. For students in medicine, nursing, pharmacy and allied health, responsible AI use must be tied to accuracy, confidentiality, patient safety and professional accountability. AI may support concept review or simulated communication, but it cannot replace judgment.

Business and applied professional fields face another issue. Students may use AI to draft reports, prepare market examples, generate presentations or shape case analysis. Since many workplaces are already experimenting with these tools, universities cannot simply ignore them. But assignments still need to show that students can analyse information, justify decisions and take responsibility for claims. A polished AI-assisted report can hide weak reasoning if the task does not require visible evidence of judgment.

Creative and design fields face questions of authorship, ownership and originality. AI can support ideation, variation and prototyping, but it can also narrow creative diversity or blur the line between human choice and machine suggestion. The authors argue that process documentation, critique and reflection are especially important in these fields because the value of creative work often lies in choices made over time, not only in the final product.

Academic integrity depends on task purpose, not detection alone

The paper argues that academic integrity in the AI era should not be reduced to catching misconduct. Detection tools can misclassify both human and AI-generated writing, creating risks for due process and student trust. They also do not teach students how to use AI responsibly.

The authors distinguish between support that helps students learn and support that creates a misleading picture of learning. AI use may be legitimate when it helps students understand material, practise concepts, receive feedback or improve clarity while leaving their own reasoning visible. It becomes a problem when students submit AI-generated work as their own reasoning, fabricate sources, conceal prohibited help or use AI where independent performance is being assessed.

Disclosure matters, but disclosure alone is not enough. A student who says AI was used still needs to show how it was used and whether the work remains accountable to the task. Instructors must explain whether AI may be used for brainstorming, language support, feedback, source preparation, code testing, outlining or final drafting. Without such detail, students are left to guess where support ends and misrepresentation begins.

Blanket bans may be justified in some settings, especially where independent performance is central to the assessment. But broad bans can drive use underground and prevent students from learning how to work responsibly with tools they may encounter in professional life. Unrestricted use creates the opposite risk, allowing students to bypass the learning process that an assessment is designed to measure.

The authors argue for a balanced approach that distinguishes prohibited, restricted, permitted and required AI use. This approach links rules directly to learning outcomes. If an assignment is meant to assess independent interpretation, AI limits should protect that purpose. If the goal is to teach students how to critique AI-generated responses, then AI may be built directly into the task.

Assessment design becomes imperative. Assignments that reward only polished final products are more vulnerable to hidden AI outsourcing. Tasks that include drafts, source trails, oral explanations, code walk-throughs, version histories, reflective notes or critique of AI outputs make student contribution more visible. Such designs do not remove all risk, but they make it harder for students to substitute output for learning.

Authorship is also a key concern. AI systems cannot take responsibility for claims, respond to criticism or be held accountable for errors. Students may acknowledge AI assistance under institutional rules, but responsibility for the final submission remains with them. That principle is especially important when AI systems generate false claims, fabricated references or biased assumptions in confident language.

The analysis also links academic integrity to study habits. Students often first encounter AI as a study aid, not an assessment shortcut. They may use it for summaries, examples, explanations, practice questions and feedback. These uses can support self-regulated learning when they help students plan, monitor understanding and revise their work, but AI can also make weak habits easier to sustain.

A student who has not read the assigned material may rely on a summary and feel prepared. A student under time pressure may generate a draft rather than work through the assignment. A student struggling with programming may accept a generated answer without understanding it. The authors caution that these problems are linked not only to AI, but also to workload, anxiety, procrastination, uneven preparation and poor assessment incentives.

Dependence on AI should hence be treated as an educational concern, not simply a moral label. Some students may rely on AI because it reduces anxiety, provides immediate feedback, supports language development or helps them navigate unfamiliar academic discourse. Students with disabilities may also use AI in ways that support access. The challenge for universities is to preserve agency, judgment and effort while recognizing legitimate forms of support.

Implications for universities, instructors and students

Universities need shared AI principles that are flexible enough to be interpreted by each discipline. Broad institutional rules on transparency, privacy, authorship and fairness remain necessary, but they will work only if academic programs translate them into course-level guidance that reflects how learning is actually assessed.

For instructors, AI policy must be tied directly to assignment design. Teachers need to explain when AI can support brainstorming, feedback, language help, source preparation or code testing, and when it crosses into misrepresentation. Assessment should make student reasoning visible through drafts, source trails, oral explanation, version histories, reflective notes or task-specific disclosure.

For students, responsible AI use depends on judgment, not just access to a tool. Students remain accountable for the accuracy, evidence and integrity of submitted work. AI can support study, revision and practice, but it cannot replace reading, disciplinary reasoning, source verification or independent academic performance.

Further, the authors argue that AI literacy should not be reduced to prompt writing. Students need to understand hallucinations, bias, privacy risks, false references, source verification, disclosure and human accountability. They also need disciplinary knowledge to decide whether an AI output is valid.

A student may write a strong prompt and still fail to detect a weak legal argument, unsafe clinical answer, flawed statistical explanation, false historical claim or broken code. Prompting can improve the relevance of a response, but it cannot replace expertise. AI literacy must therefore be taught as both a digital skill and a disciplinary skill.

Equity is another concern. Students differ in access to paid tools, digital confidence, language proficiency, disability-related needs and prior academic preparation. AI can narrow or widen these differences depending on whether universities provide clear guidance, fair access and practical training. Rules that assume all students have the same tools, skills or confidence will not serve students equally.

The paper also calls for faculty support. Instructors need time and institutional help to test tools, redesign assessments, communicate expectations and align AI use with disciplinary standards. Academic integrity staff, librarians, writing centers, accessibility specialists, educational developers and students all have roles to play in turning broad AI principles into daily teaching practice.

The analysis identifies several research gaps. More comparative research is needed across disciplines, institutions and national contexts. Longitudinal studies are needed to examine whether AI-supported study improves durable learning or encourages dependence over time. Assessment validity also requires further study as AI makes it easier to produce polished work that may not reflect genuine understanding.

The authors also point to learning analytics as a future issue. AI-supported study and assessment may create new records of revision history, process documentation and interaction data. These traces could help instructors understand learning processes, but they raise concerns about privacy, consent, bias and whether analytics can capture discipline-specific reasoning.