Grin logo
de en es fr
Shop
GRIN Website
Texte veröffentlichen, Rundum-Service genießen
Zur Shop-Startseite › Informatik - Künstliche Intelligenz

The future of Education with AI. How AI affects Higher Education

Opportunities, Challenges, Academic Integrity and Future Implications. A Systematic Literature Review

Zusammenfassung Leseprobe Details

Generative and predictive artificial intelligence systems have moved from the edges of education to the heart of teaching, learning, research and administration in just a few academic years. In this article I offer a literature review of the chances and dangers that artificial intelligence, especially generative AI brings to universities. I focus on integrity the way teachers are changing how students learn and think critically digital equity, digital divide and institutional governance. I rely on peer‑reviewed studies, systematic reviews, meta‑analyses and international policy guidance that were mostly published from 2022 to 2026. The review pulls together evidence on tutoring systems and personalized learning on how reliable and fair AI detection tools are, on how often and how artificial intelligence causes academic misconduct and on how assessment redesign is occurring in learning environments full of artificial intelligence. The review shows that data on learning outcomes is very mixed. Long research on tutoring systems shows steady moderate gains in achievement. But recent studies on AI are split: some say students become more engaged others say exam scores drop students rely too much on artificial intelligence and independent thinking weakens. Likewise artificial intelligence‑detection technologies that aim to protect integrity are not reliable and they show bias against non‑native English writers highlighting algorithmic bias and due‑process concerns that current institutional policy has not fixed. The review points out a gap between fast technology roll‑out and slow high‑quality, long‑term research that covers many institutions. It argues that careful use of intelligence in higher education demands changes in teaching design, fairness of access, data governance, AI governance and keeping human academic judgement. The article ends with recommendations that are based on evidence for universities, teachers, students, researchers, policymakers and technology makers. It also suggests a framework that links artificial intelligence adoption to learning results via the middle steps of institutional governance and responsible implementation.

Leseprobe


Table of Contents

1. Introduction

2. Conceptual and Theoretical Background

3. Literature Review

4. Methodology

5. Opportunities of AI, in Higher Education

6. Risks

7. Academic Integrity

8. The Changing Role of Educators

9. AI, Students and Learning

10. Ethical and Governance Considerations

11. Discussion

12. Recommendations

13. Future Research Directions

14. Limitations

15. Conclusion

References

Abstract

Generative and predictive artificial intelligence systems have moved from the edges of education to the heart of teaching, learning, research and administration in just a few academic years. In this article I offer a literature review of the chances and dangers that artificial intelligence, especially generative AI brings to universities. I focus on integrity the way teachers are changing how students learn and think critically digital equity, digital divide and institutional governance. I rely on peer‑reviewed studies, systematic reviews, meta‑analyses and international policy guidance that were mostly published from 2022 to 2026. The review pulls together evidence on tutoring systems and personalized learning on how reliable and fair AI detection tools are, on how often and how artificial intelligence causes academic misconduct and on how assessment redesign is occurring in learning environments full of artificial intelligence. The review shows that data on learning outcomes is very mixed. Long research on tutoring systems shows steady moderate gains in achievement. But recent studies on AI are split: some say students become more engaged others say exam scores drop students rely too much on artificial intelligence and independent thinking weakens. Likewise artificial intelligence‑detection technologies that aim to protect integrity are not reliable and they show bias against non‑native English writers highlighting algorithmic bias and due‑process concerns that current institutional policy has not fixed. The review points out a gap between fast technology roll‑out and slow high‑quality, long‑term research that covers many institutions. It argues that careful use of intelligence in higher education demands changes in teaching design, fairness of access, data governance, AI governance and keeping human academic judgement. The article ends with recommendations that are based on evidence for universities, teachers, students, researchers, policymakers and technology makers. It also suggests a framework that links artificial intelligence adoption to learning results via the middle steps of institutional governance and responsible implementation.

Keywords: Intelligence; Generative AI; Higher education; Academic integrity; Academic misconduct; AI detection; Personalized learning; Intelligent tutoring systems; Algorithmic bias; Digital divide; AI governance; Assessment redesign

1. Introduction

1.1 Background

Higher education institutions have experienced waves of technology changes in the last thirty years. These changes started with learning management systems and massive open online courses. Then came learning analytics and adaptive courseware. Now the new kind of change is from language model-based generative artificial intelligence tools. The most visible of these tools is ChatGPT from OpenAI. It became public in November 2022. This new change is different from the ones. Before the technologies mainly helped deliver organize or track content created by people. Now these AI tools can create text, code and analysis that's fluent, structured and persuasive. This happens on demand. In two years there were many studies looking at how generative AI is used in teaching, learning, assessment and research in higher education. There were also surveys showing that most university students used such tools to help with their studies.

This fast growth happened without the term theory-based evidence that usually comes before big changes in teaching. Universities, which usually take a time to test new ideas had to make new rules change how they test students and rethink rules about honesty in work. UNESCOs own guidance on AI in education published in September 2023 said that there were no national rules. This left user privacy. Schools not ready to check the tools. This article starts with this situation. Of thinking of AI in higher education as either a great improvement or a serious danger it looks at what the studies show. It describes what is clear what is still in doubt and where there is not information to say much.

1.2 The Emergence of AI in Higher Education

Artificial intelligence has been in education in small ways for a long time. Intelligent tutoring systems, which change the lessons based on what a student knows have been around since the 1970s. Studies with experiments have found that these systems help students do better than regular teaching. Predictive analytics have been used for over ten years to find students who might leave school and guide advice. What is new since 2022 is three things. First there are general-purpose AI tools that any student or teacher can use without training. Second these tools are used by students quickly. Surveys show that most students use them every week. Third it is hard to tell the difference between work done with AI help and work done by AI. This makes old ideas about testing and honesty less clear.

1.3 Problem Statement

The main problem this review looks at is the difference between how fast schools are using AI and how ready the evidence is to guide this change. Schools are asked to get the parts of AI in learning protect how tests work and keep the value of degrees protect students data and help students and teachers get ready for the job market and knowledge economy. These goals are not easy to do at the time. Personalizing learning needs a lot of student data, which can cause privacy issues. Encouraging real AI use makes it harder to catch cheating with AI because the work looks the same. Tools that schools use to catch AI cheating are not good. They have errors. Can be unfair to students who are not native English speakers.

1.4 Research Gap

Three gaps make this review important. First even though many schools have studied how students and teachers see AI there is not evidence on how using AI over a whole school career affects learning, skills and thinking. Most studies look at one semester. Second the research on AI detection tools has results. Some say the tools are good. Others say they have problems with -native English speakers. This can be a problem for students who are accused of cheating because of wrong results. Most of the rules schools have about AI were made quickly after the ChatGPT issue. There is not study on if these rules or the ways schools change tests actually work. This review looks at these three gaps. Says they are connected. Uncertainty about how AI affects learning bad detection tools and quick rules explain why schools keep changing their policies.

1.5 Significance of the Study

This review helps university leaders, teachers, people who make tests and those who make rules. It helps them make decisions when they are not sure about AI. It also helps researchers find where more studies are needed. By bringing together studies on teaching, learning, honesty and rules the review gives a picture than any one part of the fast-changing AI in education research. It also checks all the claims made about AI in education. Only claims with research are included. Ideas that are guesses or, for promotion are not included.

1.6 Research Objectives

RO1: To understand how artificial intelligence and generative AI is being used in teaching, learning, research and administration in colleges and universities.

RO2: To examine the educational and institutional benefits of using AI, such as personal learning, smart tutoring, easier access and better efficiency in managing school work.

RO3: To look at the risks involved with using AI like unfair algorithms, privacy problems, differences in access to technology and issues with keeping academic honesty.

RO4: To look at how AI affects academic honesty whether tests are fair and how the job of teachers is changing.

RO5: To find ways and methods that help schools use AI in an long-lasting way.

1.7 Research Questions

RQ1: How is artificial intelligence, generative AI being used in teaching, learning, research and school work in colleges and universities?

RQ2: What does the research say about the benefits and opportunities from using AI in schools and how similar are these findings across different places and subjects?

RQ3: What problems and dangers does using AI cause in schools and how good are the ways people are trying to fix them like tools to find AI use?

RQ4: How is generative AI changing the way schools handle honesty testing fairness and the job of teachers?

RQ5: What ways, at the school and bigger levels help schools use AI in a better and more lasting way?

1.8 Scope of the Study

The study looks at artificial intelligence in the specific area of higher education, which includes undergraduate and graduate schools. It focuses on generative AI tools that became easy to use after 2022. The study uses articles from 2022 to 2026 and also includes older research on smart tutoring systems and technology use when it is still important today. The study does not explain how big language models work and it does not look at AI in secondary schools except when findings are similar. The study includes research from North America, Europe, Asia, the Middle East and Africa. Shows where findings change depending on the place.

1.9 Organization of the Article

Section 2 gives the background and ideas about intelligence and how it is used in schools. Section 3 looks at the existing research on AI in education. Section 4 describes how the research was done. Sections 5 to 10 discuss the analysis: the chances, the problems and dangers academic honesty, the changing role of teachers the effect on students and the ethical and management issues. Section 11 has a discussion Section 12 gives recommendations Section 13 talks about future research Section 14 says what the study could not do and Section 15 ends with the conclusion.

2. Conceptual and Theoretical Background

2.1 Defining Artificial Intelligence

Artificial intelligence is when computers can do tasks that people usually need to use their brains for. These tasks include seeing, thinking, learning and making choices when things are not clear. In schools the word has been used for different types of technology. These include systems that follow rules systems that use learning to predict things and help students, tools that change to fit the student and now systems that make new content using language models. It is important to see these as technologies, not just one. Systems that help with keeping students in school have problems with unfairness and getting permission (Miao & Holmes 2023); smart tutoring systems have questions about how they work (Ma et al., 2014); and systems that create new content have problems with who wrote the work if it is original and if it is correct which were not a problem before.

2.2 Generative Artificial Intelligence

Generative AI is when systems, like the GPT family, Gemini and others are trained on a lot of text to make new content that looks like what they were taught. Unlike systems that only did one thing, like checking spelling or finding plagiarism these systems can write essays, code, summaries and answer questions in a way that is hard to tell from human work (Walters & Wilder 2023). This ability to create is what makes it helpful to teachers offering help but also a problem with honesty as it can also do homework for students. These two uses are often the same in the work.

2.3 AI in Educational Environments

Studies on AI in schools usually put AI uses into categories: (a) helping teachers with content, translation and chatbots; (b) helping students with tutoring, practice and explanations; (c) grading and feedback; (d) helping with research like finding articles, coding and analyzing data; and (e) using AI in school jobs like admissions, advice and managing resources (Labadze et al., 2023; Yusuf, 2024). This way of grouping is used in the study in Section 5 to organize the discussion about chances.

2.4 Theoretical Perspectives

2.4.1 Cognitive Load and Cognitive Offloading Theory

Cognitive load theory and the idea of cognitive offloading helps explain the mixed results of using GenAI and learning. Cognitive offloading is when people use tools of doing things in their head. If done right like using a calculator for math it helps learning.. If it replaces the hard work that helps learning it can hurt. Studies on using GenAI for research found that it made things easier but also made thinking less good which shows it is replacing of helping.

2.4.2 Self-Regulated Learning

Self-regulated learning theory says that good learning is a cycle of setting goals using strategies checking progress and thinking about it. This theory is often used in studying how GenAI helps students learn on their own. Because GenAI can give answers without students planning or checking their understanding some studies say using it a lot leads to less thinking about learning called "metacognitive laziness."

2.4.3 Validity Theory in Assessment

Messicks theory about testing says that the value of a test is about what the test results mean, not the test itself. This idea is used when looking at tests that can be done with GenAI. A test is not good if the grade does not show what the student knows. This theory discussed more in Sections 7 and 12 helps change how tests are made by asking, "Would using a tool change what the grade means?" of "Can this test be done with a tool?"

2.4.4 Diffusion of Innovation and Institutional Adoption

Rogerss idea about how new ideas spread, along with the model about how people accept new tools helps study how teachers and students use GenAI. Things like seeing if the tool helps how risky it. How good people think they are at using it predict if they will use it. Some studies show that feeling good about using the tool makes people want to use it but thinking it is risky stops them meaning using it is not, about being excited or worried.

2.5 Conceptual Framework

Combining the theories discussed earlier with the findings from sections 5 to 10 this article presents a framework. In this framework the use of intelligence in higher education does not lead directly to educational results. Instead it is influenced by two layers. The first layer includes the opportunities and risks created by the technology. The second layer includes how well the institution manages the technology.

The framework is shown in Figure 1. Explained in more detail in section 11. It suggests the following chain of cause and effect: the use of AI creates both opportunities such as personalization, better access and more efficiency and risks such as threats to honesty, bias, overreliance and unfairness. The balance between these opportunities and risks depends on how the institution manages the technology.

How well the institution manages the technology affects how responsible the use of AI is in practice. It is the use of AI, not just the use of AI itself that influences the results. These results include learning, more fairness and the trustworthiness of academic qualifications.

This framework suggests that comparing institutions or studies based on whether they use AI or not is not likely to give clear results. The reason is that how well the institution manages AI can be just as important, as the technology itself.

Figure No.1

Illustrations are not included in the reading sample

3. Literature Review

3.1 Evolution of AI in Higher Education

Research on AI in education can be split into two main periods: a pre‑2022 period and a post‑2022 period. The pre‑2022 literature is dominated by tutoring systems, adaptive learning platforms and predictive analytics. This literature is relatively mature with many meta‑analyses covering decades of controlled studies. Those meta‑analyses show that intelligent tutoring produces positive effects on achievement compared with conventional instruction though the effects are smaller than those seen with expert human one‑to‑one tutoring (Ma et al., 2014; VanLehn, 2011). The post‑2022 literature grew rapidly after ChatGPT was released to the public. Systematic mapping reviews have identified hundred publications on generative AI in education within about eighteen months of ChatGPT’s release alone (Yusuf, 2024). Most of these publications are small‑sample surveys of student or faculty perceptions not controlled outcome studies.

3.2 AI‑Supported Teaching

Studies of AI‑supported teaching describe applications that help instructors. These include automated creation of lecture materials quiz items and grading rubrics; language and grammar help for faculty who teach in languages; and chatbots that answer routine student questions outside of class hours. Faculty in countries report that these tools save preparation time and in some cases help students who need extra support (Labadze et al. 2023). At the time faculty surveys consistently show concern about the reliability of AI‑generated content and about the risk that relying on AI‑produced materials could reduce the depth of subject‑matter engagement that instructors are expected to provide.

3.3 AI‑Supported Learning and Personalized Education

The strongest and longest‑standing evidence for AI‑supported learning comes from tutoring systems. These systems have been evaluated in least twenty‑two different types and in dozens of controlled studies in higher education (Ma et al., 2014). The meta‑analyses report effect sizes (Hedges’ g of about 0.3 to 0.4 relative to conventional instruction). Intelligent tutoring systems generally outperform forms of computer‑assisted instruction but are still weaker than expert human one‑to‑one tutoring (VanLehn, 2011). Personalisation in the GenAI era extends this tradition by using interfaces rather than structured problem sets to adapt explanations to individual learners. However because GenAI systems were not designed as teaching agents the evidence for their effectiveness is thinner and more mixed than for purpose‑built intelligent tutoring systems. This difference is often overlooked in popular discussions.

3.4. Feedback

The literature on AI and assessment splits into two strands. The first strand looks at AI used to generate feedback and grading. Evidence suggests that GenAI can give individualized formative feedback on writing and code at a scale that human instructors alone cannot achieve. Nonetheless concerns remain about the accuracy and the teaching value of that feedback. The second strand deals with the disruption that AI causes to the validity of existing assessment formats. A growing body of work argues that conventional take‑home essays and unsupervised written assignments no longer reliably show that a submitted grade reflects a student’s understanding. This argument has sparked proposals to redesign assessment discussed in depth in Section 7 (Gonsalves, 2025; Roe et al., 2025).

3.5 Academic Research

GenAI tools are widely used by students and faculty to support literature searching, summarisation, drafting and coding in the research process. This use case carries a risk that is well documented in the scientometric literature: GenAI systems frequently create plausible‑looking bibliographic citations, a phenomenon called hallucination. Controlled evaluations found that early GPT‑3.5 fabricated than half of the citations it generated for academic literature reviews and that even the more capable GPT‑4 fabricated close to one‑fifth (Walters & Wilder 2023). Separate case studies in scientific domains report fabrication and error rates of a similar magnitude (Alkaissi & McFarlane 2023). The risk of hallucination is discussed in detail in Section 6.

3.6 Student Engagement and Faculty Experiences

Large cross‑sectional surveys of students report a favourable overall attitude toward GenAI. Students frequently value GenAI for brainstorming writing assistance and rapid access to explanations. At the time students express concern about factual accuracy the erosion of independent skill development and the fairness of institutional policy (Chan & Hu 2023). Faculty surveys by contrast tend to show ambivalent or divided sentiment. Studies analysing samples of faculty free‑text responses find substantial proportions of neutral or negative sentiment even when majorities acknowledge some productivity benefit. Faculty consistently rank integrity as their most salient concern (Cotton et al. 2023).

3.7 Institutional Adoption

At the level AI adoption follows a clear trajectory. An initial reactive phase ( 2022–2023) was dominated by prohibition‑oriented policy and a proliferation of AI‑detection tool procurement. A subsequent phase saw institutions publish formal guidance distinguishing permitted from prohibited uses. In a recent phase assessment redesign, AI‑literacy curricula and structured disclosure requirements have become the dominant institutional response gradually replacing pure prohibition or pure detection‑based enforcement.

3.8 Contradictions and Unresolved Issues in the Literature

genuine contradictions recur across the literature reviewed above and are treated critically throughout this review rather than smoothed over.

First findings on GenAI and academic performance directly contradict each other: some studies report that GenAI‑assisted test preparation is linked to subsequent exam scores while others find no significant difference in grades between AI‑assisted and unassisted writing and still others report performance gains, especially in programming and language‑learning contexts.

Second findings on the reliability of AI‑detection tools are sharply inconsistent. Some evaluations report near‑zero rates for leading commercial detectors while others find substantial systematic bias, against non‑native English writers (Liang et al. 2023).

Third, faculty and student perceptions of the technology frequently diverge. Faculty emphasize integrity risk whereas students emphasize inconsistent enforcement. This divergence itself creates a governance challenge of the technology’s actual capabilities.

Table 1. Major Applications of Artificial Intelligence in Higher Education

Illustrations are not included in the reading sample

4. Methodology

4.1 Research Design

This article uses a narrative-synthesis literature review design instead of a formal Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) systematic review. This choice reflects the scope of the topic, which covers pedagogy, assessment, ethics, governance and technology across a very diverse and rapidly expanding literature. It also reflects the limits of a single-author narrative synthesis, which cannot include dual-reviewer screening, pre-registration or exhaustive multi-database Boolean searches that a formal systematic review or meta-analysis would need. When the review references studies that used PRISMA-guided review or meta-analytic methods (for example Ma et al., 2014; Yusuf, 2024) the review makes this clear. It reports the effect sizes and sample characteristics of those studies instead of accepting their conclusions as unmediated fact.

4.2 Information Sources and Search Approach

Literature was found through iterative web-based search that targeted peer‑reviewed journal articles, systematic reviews, meta-analyses and major institutional and international policy reports. The search focused on sources indexed by. Cross‑listed with major publishers and repositories such as Springer Nature, Elsevier (ScienceDirect) Taylor & Francis SAGE, MDPI, Nature Portfolio, Wiley, ERIC and arXiv. Official publications from UNESCO and national education authorities were also included. Search terms combined core concepts (“generative AI " "ChatGPT," "large language models”) with domain‑specific terms (e.g. "higher education," "academic integrity," "assessment," "intelligent tutoring," "AI detection," "digital divide," "algorithmic bias," "AI literacy " "academic governance”).

4.3 Publication Period

The main search window was January 2022 to September 2026 covering the period before and after the public launch of ChatGPT. Foundational studies published before this window were kept when they remain the most cited empirical basis for a specific claim discussed in the review, especially the meta‑analytic literature on intelligent tutoring systems (Ma et al., 2014; VanLehn, 2011).

4.4 Inclusion Criteria

- Peer‑reviewed journal articles, systematic reviews or meta‑analyses addressing AI or generative AI in the context of education published in English.
- Major reports or guidance documents issued by organisations (e.g., UNESCO) or national education authorities that directly address AI in higher or tertiary education.
- Working papers and preprints (e.g., arXiv) that report original empirical data are recent and are not contradicted by subsequently published peer‑reviewed work identified in the search.
- Studies reporting quantitative effect sizes, prevalence rates or comparative outcome data relevant to the review’s research questions.

4.5 Exclusion Criteria

- Sources that focus on primary or secondary (K‑12) education without findings clearly transferable to the higher‑education context.
- Non‑peer‑reviewed promotional or vendor material making unverifiable claims about product performance.
- Sources for which bibliographic details (author, venue or year) could not be confirmed through the search process in line with the review’s commitment not to present unverifiable material as an academic source.
- Opinion. Blog commentary that do not report original data except where cited narrowly for a specific documented event (e.g. a reported institutional policy change) rather than for an empirical claim.

4.6 Screening Process and Study Selection

Because this review does not follow a systematic‑review protocol no dual‑reviewer PRISMA flow diagram is reported. Instead sources retrieved through the search process described above were screened by the author for relevance credibility of venue and where empirical claims were extracted, for the presence of verifiable bibliographic detail (author names, publication venue and where available digital object identifiers). Sources that could not be corroborated in this way were excluded from the basis of the review rather than cited on the strength of a secondary or unverified mention.

4.7 Data Extraction and Thematic Synthesis

For each retained source the author extracted the research question or purpose, method and sample (where and principal quantitative or qualitative findings relevant to the themes of this review: opportunities, risks, academic integrity, the role of educators, student learning and governance. Extracted findings were then organised thematically according to the structure set out in Sections 5 through 10 following the logic of narrative thematic synthesis rather than statistical meta‑analytic pooling given the heterogeneity of outcome measures across the underlying studies.

4.8 Quality Assessment

Sources were informally appraised for transparency (clarity of sample, design and analysis) venue credibility (peer review status, publisher reputation or official institutional/international authorship) and for quantitative claims whether an effect size, sample size or comparable statistic was reported rather than an unsupported summary judgement. Meta‑analyses and systematic reviews (e.g. Ma et al., 2014) were treated as carrying evidentiary weight than single cross‑sectional surveys and this hierarchy is reflected in how findings are qualified throughout the review (for example through language distinguishing "consistent" findings across multiple meta‑analyses from "emerging" or "preliminary" findings from single studies).

4.9 Ethical Considerations

As a literature review synthesising material this study did not involve the collection of new primary data from human participants and therefore did not require institutional ethical approval. Ethical care was nonetheless exercised in two respects: first, in representing the findings, limitations and uncertainty of the underlying sources rather than overstating their generalisability; and second in explicitly declining to reproduce, invent or embellish bibliographic details statistics or quotations beyond what could be verified consistent with the broader scholarly obligation to avoid citation fabrication—a phenomenon this review itself documents as a significant risk of uncritical AI‑assisted scholarship (Walters & Wilder 2023).

4.10 Methodological Limitations

This review’s methodology has acknowledged limitations, which are revisited in full in Section 14. It was conducted by a reviewer without independent dual screening, which introduces a risk of selection bias in which sources are included or emphasised. It relies on academic search rather than direct, licensed access to proprietary bibliographic databases such as Scopus or Web of Science meaning that database coverage, while substantial cannot be guaranteed to be exhaustive. It does not compute effect sizes across studies and therefore cannot offer the same quantitative precision as a formal meta‑analysis. Finally because the underlying field is evolving rapidly with new large‑scale surveys and institutional policy changes appearing monthly any synthesis of this kind carries an inherent risk of being partially superseded by the time of publication.

5. Opportunities of AI in Higher Education

5.1 Personalized Learning and Intelligent Tutoring

The methodologically robust opportunity documented in the literature is the capacity of intelligent tutoring systems to improve learning outcomes relative to conventional non‑adaptive instruction. A meta‑analysis of thirty‑nine studies across twenty‑two intelligent tutoring systems used in higher education found that these systems produced a moderate positive effect on academic achievement (Hedges’ g in the range of 0.32–0.37) outperforming other instructional technologies and traditional classroom instruction alone though remaining less effective than expert human one‑to‑one tutoring (Ma et al., 2014). An earlier meta‑analysis of fifty controlled evaluations similarly found an improvement of 0.66 standard deviations on locally developed assessments though this advantage shrank considerably to roughly 0.13 standard deviations, when outcomes were measured using standardised tests rather than instructor‑designed assessments—an important qualification suggesting that the benefits of intelligent tutoring are strongest when the assessment is closely aligned with the system’s own instructional content (VanLehn, 2011). This body of evidence accumulated over than a decade and across multiple independent meta‑analyses represents the single strongest empirical foundation for claims about AI’s positive educational impact and it should be distinguished clearly from the far thinner and more mixed evidence base concerning general‑purpose generative AI chatbots discussed in Section 9.

5.2 Accessibility and Student Support

A growing and increasingly rigorous strand of research documents benefits of generative AI for students with disabilities. Survey research with students who identified conditions including ADHD, dyslexia, dyspraxia and autism found that these students used AI chatbots primarily to summarise dense reading, structure essays and sustain attention across long academic writing tasks and expressed a strong desire to be involved in institutional AI policymaking rather than have policy imposed on them without consultation. The same body of work however identifies serious limitations: students reported concerns about the accuracy of AI‑generated content the risk that AI use could itself trigger academic integrity investigations and inequitable access created by the cost of premium AI subscriptions relative to the free tiers available to better‑resourced peers (University of Virginia Teaching Hub, 2024). Institutions applying a Universal Design for Learning framework have begun using AI to help instructors proactively build multiple means of representation and expression into course materials rather than relying solely on individualised reactive accommodation requests a shift with potential benefits extending to the broader student population and not only to students with disclosed disabilities.

5.3 Automated Feedback and Teaching Assistance

Instructors report that generative AI tools reduce the time required to draft feedback comments generate practice questions and rubrics and produce first‑pass explanations of difficult concepts, particularly in large‑enrolment courses where individualised feedback would otherwise be infeasible at scale (Labadze et al. 2023). Because feedback speed and specificity are established correlates of learning gains in the educational‑technology literature this capability is among the more plausible mechanisms by which AI could improve learning outcomes independent of any change to the underlying instructional content. This opportunity is qualified however by the accuracy concerns discussed in Section 6: feedback that's fluent but factually or pedagogically unsound may be worse, than no automated feedback at all particularly if students are unable to independently judge its correctness.

5.4 Research Assistance

For research generative AI offers clear efficiency gains in searching literature, summarising, drafting first drafts and coding. Generative AI has been widely adopted by researchers at all career stages for these tasks. I have seen many colleagues use AI for these tasks. However this opportunity also brings the serious and well‑documented factual reliability risk found in this review: generative AI systems often fabricate plausible but non‑existent citations. This problem is discussed in full in Section 6.4 (Walters & Wilder, 2023; Alkaissi .  McFarlane, 2023). I worry about this risk. The overall impact of AI on research productivity therefore depends on whether researchers verify the content produced by generative AI against primary sources. Empirical audits suggest that this verification step is often skipped.

5.5 Administrative Efficiency and Data‑Informed Decision‑Making

Beyond the classroom universities use AI‑enabled chatbots and predictive analytics to streamline admissions address financial aid questions, schedule courses and spot students who might withdraw early. For example Georgia State University uses a chatbot to reduce "summer melt" among admitted students who do not enrol. This case is cited repeatedly in the literature showing that AI‑supported outreach can noticeably improve enrolment and retention. I have read reports that highlight this. These applications raise concerns about assessment validity, which are the focus of Sections 6 and 7. Instead they bring their governance issues, such as data privacy, informed consent and algorithmic bias in predictive models that label students as "at risk." These issues are discussed in Sections 6.6 and 10.

5.6 Curriculum Development and Potential Improvements in Educational Outcomes

At the curriculum level AI tools are increasingly used to help design learning objectives assessments with outcomes and identify gaps in the curriculum by analysing student performance data. I have noticed that teachers are excited about these possibilities. The potential for AI to improve educational outcomes at scale is real but it should be stated carefully and specifically not in broad terms. The strongest and most consistent evidence comes from structured purpose‑built tutoring systems, within well‑defined domains (Ma et al., 2014). Claims that general‑purpose AI can transform learning outcomes are far less certain. This distinction is explained further in Section 9.

Table 2. Opportunities and Potential Benefits of AI in Higher Education

Illustrations are not included in the reading sample

6. Challenges and Risks

6.1 Hallucinations and Inaccurate Information

Generative AI systems often produce content that sounds convincing but is factually wrong. This problem, known as hallucination is one of the well-documented risks in current research. A controlled study tested GPT-3.5. Gpt-4 by asking them to write short academic literature reviews on forty-two different topics. Researchers then checked all 636 citations these models generated. The results showed that 55% of the citations from GPT-3.5 were completely made up. For GPT-4 this number dropped to 18%. Among the citations that were not completely fabricated many still contained errors at the reference level (Walters & Wilder 2023). Another study focused on stem cell research found that out of 86 AI-generated references over 15% were entirely false and another 9% had inaccuracies (Alkaissi & McFarlane 2023). These results point to a truth: hallucinations are not rare quirks. They are common and systematic in today’s AI models. This has implications for academic work, including literature reviews, grant proposals and even legal or medical documents. Other studies testing chatbots beyond ChatGPT found similar variation in error rates showing that this issue is not limited to one product. It is a feature of current language models.

6.2 Bias and Overreliance

Bias in AI used in education appears in two main forms, often mixed up in discussions. The first is bias in the content that AI produces. This bias comes from the data the models were trained on. May leave out or distort perspectives from non-Western non-English or minority groups. The second form of bias is in the tools used to detect AI-generated content. These tools have been shown to flag the writing of non-native English speakers as AI-generated (Liang et al., 2023). This kind of bias in detection systems can lead to treatment. Alongside bias overreliance is a growing concern. Numerous studies, both qualitative and experimental show that students who frequently use AI for tasks like writing or problem-solving tend to rely less on their own thinking. They show independent reasoning and are less likely to keep trying when faced with tough analytical challenges. In lab studies brain scans revealed reduced activity in areas linked to creativity and cognitive control when students used AI while writing. These students also remembered less of the material they produced. The risk of overreliance is further explored in Section 9 in relation to critical thinking and cognitive development.

6.3 Reduced Critical Thinking

One specific concern within the issue of overreliance is a decline in critical thinking. Interviews with university students show that using AI leads to a kind of offloading. Students report thinking deeply and being less willing to stick with difficult thinking tasks. This matches earlier findings that overreliance on AI dialogue systems is linked to decision-making, lower critical thinking and poorer analytical skills. Experiments comparing AI-assisted research with web-based research show that AI helps reduce the mental effort required.. At the same time the quality of arguments produced with AI assistance is lower. The depth of the reasoning is less. This suggests that the reduction in effort many students report is not always a thing. It may come at the cost of understanding and better thinking. The trade-off between effort and poorer reasoning is discussed further in Section 9.

6.4 Privacy, Surveillance and Cybersecurity

AI systems used in education whether for personalized tutoring predicting student retention or handling administrative tasks through chatbots rely on collecting and processing large amounts of student data. This data includes grades, behavior patterns, engagement levels and in some cases, biometric or emotional data from AI-enabled monitoring tools. UNESCO’s global guidance highlights this as a governance issue. It points out that most countries do not have laws to protect student data when generative AI is used. The guidance recommends that schools and governments put in place data protection rules and consider age limits on how students can use AI independently (Miao & Holmes 2023). This risk is made worse by incidents where student data shared with commercial tech vendors was used in ways not agreed upon. Data governance failures are not just theoretical. They happen. Cybersecurity risks grow as AI tools connect with core systems like student information systems and learning management platforms. These connections increase the number of entry points for hackers. Any data breach becomes more dangerous because of the amount and sensitivity of the information involved.

6.5 Digital Inequality and The AI Divide

The adoption of AI in higher education risks making existing educational inequalities worse. This is called the AI divide. It builds on the digital divide, which is about unequal access to devices and internet.. The AI divide adds another layer: unequal access to the skills, knowledge and support needed to use AI effectively. Research with faculty shows that differences in access go beyond devices or subscriptions. Differences also exist between disciplines. Humanities departments often express skepticism toward AI tools.. They also face growing demand for the interpretive and critical skills that AI cannot easily provide. Many advanced generative AI tools are offered through paid subscriptions. This makes cost a barrier to entry. Studies on students with disabilities have found that subscription costs can make it harder for students who might benefit most from AI to access it (University of Virginia Teaching Hub, 2024). Research focused on community colleges shows that AI-related digital divides worsen existing gaps. These gaps affect students chances to learn AI skills and later impact their job prospects.

6.6 Algorithmic Bias in Predictive and Administrative Systems

The risk of bias is not just about generative AI. Predictive analytics used in admissions, advising and early-warning systems for student retention carry this risk. These systems often learn from institutional data. That data may reflect inequalities. As a result the models can. Even amplify those biases. Students from marginalized communities can be unfairly disadvantaged by these systems. A major concern is the lack of transparency. These models are often opaque. Students and their advisors may not. Challenge a classification that labels a student as high risk. This lack of transparency creates due-process problems. It is very similar to the fairness issues seen in AI-detection tools that were discussed in Section 7. The ethical concerns raised by this risk are addressed in detail in Section 10.

6.7 Copyright and Intellectual Property

Using AI in academic writing and research brings up unresolved questions about copyright and intellectual property. Current institutional policies have not kept up with the technology. Large language models are trained on collections of text. Much of this text is. Often not credited. This raises the question: who owns the content generated by AI? Is it the student, the institution or the model provider?. How should AI-generated work be acknowledged? These issues are still. Vary by country. This uncertainty adds to the confusion around authorship and attribution which are discussed in Section 7. It also shows that policies are lagging behind both the technology and the legal environment.

6.8 Lack of Transparency and Technological Dependency

A major challenge is the lack of transparency in AI systems. Large language models produce outputs through processes that're hard to understand even for the people who built them. This makes it difficult to check for bias verify the source of information or explain why a model gave a result. This problem includes grading decisions and risk classifications. The lack of explainability creates a challenge in academic settings. It also increases the risk of dependency. Institutions are increasingly relying on AI for tutoring, grading and administrative tasks. This creates a dependency, on third-party providers. If a vendor changes pricing discontinues a service or updates a model institutions have control. This dependency can disrupt academic functions and create long-term risks.

Table 3. Risks and Challenges Associated with AI in Higher Education

Illustrations are not included in the reading sample

7. AI and Academic Integrity

7.1 AI-Generated Assignments and Contract Cheating

The biggest concern around AI in education is its ability to create full convincing assignment responses on demand. This acts like a instant version of contract cheating. A form of academic dishonesty where students get help from others to complete work. But without the usual money exchange that happens in traditional essay mills. Soon after ChatGPT became publicly available reports began to surface of students using it to cheat. These early cases led schools and universities to respond quickly often with strict bans. Unlike contract cheating, which leaves behind signs of a deal. Like payment records or communication. AI-assisted cheating leaves no such trail. The content generated by AI is not copied from any one source. It is new, but not original in the way we usually think. This makes it hard for standard plagiarism detection tools to catch because those tools look for matches not for new text that still reflects another’s input.

7.2 Plagiarism, Authorship and the Concept of "AIgiarism"

Since AI-generated content is technically original. It doesn’t copy any existing text. Its unacknowledged use challenges the idea of plagiarism. Plagiarism used to mean taking someone ’s words or ideas and presenting them as your own.. Now students can use AI to write a paper and still claim it as their own work even though they didn’t write it. This new kind of misconduct has been given a name: "AIgiarism." It refers to the act of submitting AI-generated material without saying it was made with AI help. This is misleading because it hides the source of the work. Some students might not even realise they’re doing anything. They may think that because the text isn’t copied it’s okay to use.. Most schools’ academic integrity policies see this as dishonesty. This confusion matters because it affects how institutions teach honesty and what kind of education they provide. The lack of understanding makes it harder to prevent this kind of behavior.

7.3 Disclosure and the Question of Legitimate Use

A shift in policy and thinking has moved away from simple bans on AI use. Instead institutions are now focusing on whether studentsre being honest about their use of AI tools. The idea is to encourage transparency than just try to catch cheating. This approach is similar to how assessment redesign works. Focusing on the process not the final product. Than trying to detect whether AI was used institutions are asking students to say how they used AI and what role it played in their work. This changes the focus from "Did you use AI?" to "Did you say you used AI and was it used in the way?" The goal is not to stop AI use but to make sure its used responsibly. Students are expected to explain their process show their drafts or give a reflection on how they worked. This helps keep the focus on learning, not on getting a good grade.

7.4 AI Detection: Capabilities and Limitations

Many institutions have turned to AI detection tools like Turnitin, GPTZero, Copyleaks and Originality.ai to help find cheating.. The science behind these tools is not solid. Research shows that they make mistakes. And often in ways that're unfair. One major study found that early AI detectors wrongly labelled than half of essays written by non-native English speakers as AI-generated. These essays were written by students but the detectors confused their language patterns. Lower vocabulary diversity, simpler sentence structures. With AI patterns. The study concluded that bias was built into the tools not because of cheating but because of differences in writing style. Other studies have seen the bias in formal writing like literature reviews, which often follow a predictable structure. Even human experts struggle to tell the difference. One small study found that teachers and students could only spot AI-generated text about 25% of the time. Better than random guessing.

Recent evaluations of newer tools, like Pangram and Copyleaks look better. Some studies report no positives on their test sets.. Experts caution that this doesn’t mean the tools are reliable. The samples were small. The results may not hold up in real-world use. The key point is this: no AI detection tool can be trusted as proof of cheating. Institutions must be careful. They should not rely on detection alone. If a tool flags a student’s work that’s a signal. Not a verdict. More evidence is needed. This is especially true for students who're not native English speakers or who write in formal structured ways. They are more likely to be accused. So institutions must be cautious and fair.

7.5 Assessment Redesign

Because detection is unreliable educators are looking at ways to redesign assessments so they are harder to cheat with AI. One popular idea is called "assessment twins." This means creating two tasks: one that’s easy to cheat on (like an essay at home) and another that’s harder to fake (like an in-class exam or a spoken presentation). The idea is to keep the benefits of assignments while still measuring real learning. Another approach is to design assessments that're resilient”. Meaning they can’t be easily handled by AI. Oral exams, viva voce and in-class presentations are examples. These require students to speak on the spot explain their thinking and defend their answers. AI can’t do that. A recent survey showed that student use of AI for coursework went from about 50% to 90% in just a year. That makes redesign urgent.

A growing body of research also supports "contextual assessment design." This means making assignments that are tied to real-life situations, local contexts or specific experiences. When assignments are grounded in unique situations they’re harder to fake. They’re also more meaningful. The idea is to make work that only the student could do because it’s connected to their life, views or experiences. Methods include requiring process documentation. Like drafts, outlines or reflections. And using peer review so that each person’s contribution can be checked. These kinds of changes help keep the focus on learning, not on getting a result.

7.6 Academic Misconduct Policies and Responsible Student Use

Policies around AI use are changing fast. Many schools used to say "Don’t use AI." Now many are shifting to nuanced rules. Some schools use a traffic-light system. Yellow red. To show how much AI help is allowed for different tasks. For example using AI to brainstorm might be green (okay). Using it to write a whole paper might be red (not allowed). The goal is to be clear not to block. Research shows that education. Teaching students why honesty matters. Works better than punishment alone. When schools combine training with technology tools students are more likely to follow the rules. Pilot programs using reflective personal tasks have helped reduce cheating because students are more engaged and less likely to take shortcuts.

When it comes to student use experts agree on a few key points. First students should always say if they used AI. Second AI should help them. Not do the work for them. It should be a tool that supports thinking not replaces it.. Third the final work must be something the student can explain. If asked they should be able to talk through their logic defend their choices and repeat their reasoning. This ensures that learning is real and that integrity is upheld. The same ideas apply to students, with disabilities. For them AI can be a support. As long as it’s used fairly and disclosed. The goal is not to stop technology but to use it wisely.

Table 4. Academic Integrity Risks and Possible Institutional Responses

Illustrations are not included in the reading sample

8. AI and the Changing Role of Educators

8.1 Teacher Responsibilities in an AI-Saturated Classroom

University teachers now have responsibilities that go beyond teaching subject matter. They need to talk about, use and change the AI policies that the institution has set. They have to make decisions on their own about which tasks in their classesre still good for traditional assessment and which need to be changed. They also have to deal with students who feel they are being unfairly accused of cheating because the tools to find AI use are not reliable. Studies of what teachers think show that this is confusing and adds a lot of work. Teachers at an university say that the quality of student writing is getting worse they have less control over how they test students and it is hard to decide if a student used AI without permission.

8.2 AI Literacy Among Educators

Faculty members have levels of knowledge about AI. Surveys across the world show that most teachers have used AI in some way. Most say they have not used it much. Some have not used it all. This difference is connected to the type of institution the subject they teach and whether they have training in AI. Studies show that teachers use AI to save time try teaching methods and grow professionally.. They worry about the quality of AI content less personal contact, security and not having support from their institution. These worries stay even if they like AI.

8.3 Professional Development and Institutional Support

Research shows that training for teachers is more important than having AI tools. Teachers who get help and training see AI as a way to help students and improve teaching. Those without training are more likely to feel worried or use AI in a way. This shows that supporting teachers with AI is needed before they can use it well. This point is important for how institutions manage AI.

8.4 Human-AI Collaboration and Pedagogical Decision-Making

of replacing teachers AI should work with them. AI can do tasks that're repetitive and not very important like answering questions or giving basic feedback. Teachers should still make decisions about learning goals, understanding student problems and handling cheating cases. Surveys show that students and teachers don't think AI can replace teachers. They both believe that human decisions are still very important in education.

8.5 Risks of Excessive Automation and Teacher Autonomy

A big worry is that using much AI can take away a teachers control over what they teach and how they test. If AI tools are forced on teachers without their input it can make teaching less varied. Teachers are worried that using the AI tools from a company might make all classes the same which is not good for teaching. This issue between using AI to save time and keeping teaching is still not solved and needs more research.

8.6 Future Educator Competencies

Looking at all the studies the skills needed for teachers when AI is part of the classroom include knowing what AI can and cannot do, changing how they test students talking about when AI should be used and knowing when to trust AI or not. These skills show that knowing about AI is very important for teachers and in the future.

9. AI, Students and Learning

9.1 Student Engagement and Perceptions

Many students like AI because it helps with ideas, writing and understanding course material.. Students are also worried about how accurate AI is and if they are depending too much on it. They also worry about how fair the rules about AI use're. Different regions have views on AI and how students see authority and technology can influence how they feel about using AI.

9.2 Independent Learning and Cognitive Offloading

The big question is whether using AI makes students learn better or worse. Some studies show that using AI makes students less involved in thinking and less willing to work through problems on their own.. Other studies show that when teachers plan well AI can help students learn. This means that how AI's used in class matters a lot.

9.3 Critical Thinking and Analytical Reasoning

Studies show that using AI a lot can make students worse at thinking on their own. They may not check if AI is right or wrong. Younger students are more at risk.. Some studies show that when teachers teach students to think critically about AI they can get better at spotting mistakes. This means that how teachers use AI is important.

9.4 Creativity, Writing Development and Problem-Solving

Some studies show that AI helps students write better and feel more confident.. Other studies show that when students use AI their brain activity changes and they don't remember as well. In math using AI to help with problems can be good if its used to support learning not replace it.

9.5 Accessibility and Differentiated Support

Students with disabilities find that AI helps them read, write and stay focused.. Other students don't always get the same benefits. This means that AI can help some students more than others and policies need to take this into account.

9.6 Distinguishing Productive from Inappropriate AI Use

There is a difference, between using AI in a way that helps learning and using it in a way that's not good. Productive use means AI helps students learn. They still do the work themselves. Inappropriate use means students use AI to do work that the teacher is trying to test. This helps teachers and students know how to use AI

10. Ethical and Governance Considerations

10.1 Explainability

Ethical governance frameworks for AI in education consistently identify transparency as a foundational principle encompassing both transparency about when and how AI is used within a course or institutional process and the more technically demanding requirement that AI-driven decisions (grading, risk flagging, admissions screening) be explicable to the students they affect. Because the underlying models are not fully interpretable to their developers full technical transparency is not currently achievable; institutional frameworks have therefore tended to substitute procedural transparency—clear disclosure that AI was used and clear channels for challenging AI-informed decisions—for the technical transparency that current systems cannot fully provide.

10.2 Accountability and Human Oversight

A universal principle across institutional and international AI governance frameworks is that a human, not an algorithm must remain the accountable decision-maker for consequential academic outcomes, including grades, admissions decisions and misconduct findings. This principle follows directly from the detection-reliability evidence discussed in Section 7.4: because AI-detection tools are demonstrably fallible and biased treating a detection score as dispositive than as one input into a human-adjudicated process would constitute an abdication of the human-oversight principle that virtually all major governance frameworks, including UNESCOs explicitly require (Miao & Holmes 2023).

10.3 Fairness

Fairness concerns in this domain operate on least three distinct levels: fairness in access to AI tools (Section 6.5) fairness in how AI systems themselves treat different groups of students most concretely illustrated by detector bias against non-native English writers (Liang et al., 2023) and fairness in how institutions apply AI-related policy across different courses, disciplines and student populations. Evidence that international and multilingual students may face an elevated risk of false misconduct accusations, purely as an artefact of detection-tool bias rather than any difference in actual academic behaviour represents one of the most concrete and ethically serious fairness failures documented anywhere in this review and one that current institutional practice has been slow to address systematically.

10.4 Privacy and Consent

Consistent with the risks outlined in Section 6.4 ethical governance frameworks emphasize consent as a precondition for the collection and use of student data by AI systems alongside clear limits on secondary use of that data by third-party vendors. UNESCOs guidance goes further recommending that institutions and national regulators consider age- limits on students unsupervised independent interaction with generative AI platforms reflecting a broader precautionary stance toward data collection from younger or more vulnerable student populations (Miao & Holmes 2023).

10.5 Institutional Governance and Policy Development

The governance literature converges on a number of structural recommendations: the establishment of a designated institutional body (an AI governance committee or equivalent) responsible for coordinating policy across academic, IT, legal and student-affairs functions; the development of graduated, task-specific rather than blanket AI-use policies; mandatory disclosure requirements for AI use in academic submissions; and ongoing iterative policy review given the pace of technological change. Comparative cross-national research on university AI governance frameworks finds variation in how far institutions have progressed along this path with a substantial gap remaining between the frameworks recommended in the academic and international-policy literature and the reactive often prohibition-oriented policies that many institutions had actually implemented as of the mid-2020s.

10.6 Responsible AI and Ethical Frameworks

Beyond institution- policy several broader ethical frameworks have been proposed or adapted for the higher-education context most drawing on the general principles of fairness, accountability, transparency and ethics (FATE) developed in the wider AI-ethics literature. UNESCOs guidance operationalises these principles for education by proposing a human-centred approach that treats AI as a tool intended to serve human educational agency, equity and cultural and linguistic diversity rather than a technology whose adoption is an end, in itself and by explicitly cautioning against allowing AI systems to be validated for educational use without a corresponding pedagogical and ethical assessment (Miao & Holmes 2023).

Table 5. Ethical Principles for Responsible AI Adoption in Higher Education

Illustrations are not included in the reading sample

11. Discussion

11.1 Interpreting the Major Findings

When you look at the information covered in Sections 5 to 10 it does not support an answer about whether AI is good or bad for higher education. This lack of an answer is itself an important discovery. The strongest and most consistent evidence about effects on learning comes from a type of AI called intelligent tutoring systems. These systems were developed before the popular AI tools and are different in technology. On the hand the evidence about the tools that are causing the most problems now like general-purpose AI chatbots is not as strong. It is newer and mixed in terms of learning results. However there is a lot of documentation about two risks: making up academic citations and problems with the tools used to detect such behavior. This difference between a established AI system and a new one along with the way people talk about AI explains why the debate about AI in higher education is often so heated. People are often talking about things but using the same words.

11.2 Comparing Contradictory Findings

Three contradictions found in Section 3.8 need explanation. First the difference between studies that show a drop in exam scores and those that show no change or a small improvement can be explained by how students used the AI. When students used AI of doing their own work it made their scores worse.. When they used AI to help explain things without replacing their own efforts the scores stayed the same or improved. Most studies do not mention how the AI was used, which makes it look like the results are conflicting when they may actually be measuring things. Second the differences in how AI detectors work from almost no false positives to many are partly because the tools have improved. However the review suggests that even if recent studies show results it does not mean the earlier problems with bias have been fixed. Third the difference between what faculty and students focus on—integrity risks versus enforcement—comes from their different experiences. Faculty see cases of cheating while students face the consequences of errors in detection tools. Both groups are responding rationally to their situations.

11.3 Theoretical Connections

The patterns seen in Sections 6 and 9 match the theories from Section 2.4. The idea of offloading explains the results: when students use AI they feel less effort but engage less deeply with the material. Whether this helps or hurts learning depends on the task. Messicks validity framework also explains why assessments are changing. Of asking if a task is AI-proof the focus is now on whether the grades still make sense if AI was used. This change is based on theory, not practice.

11.4 Broader Implications for Universities

For universities the main takeaway is that decisions about AI must consider the tools the tasks they are used for and the rules around their use. A university that uses AI broadly but does not change how it tests or trains staff will likely face problems.. One that uses AI and has good rules and training is better off.

11.5 Where the Evidence Remains Insufficient

Some areas lack data. There is little long-term evidence about how students develop over time with AI. Most studies look at one course or a short period. There is not comparison between different ways to change assessments. Also there is information on how policies affect students who are in between or unclear cases. These gaps are discussed in detail in Section 13.

12. Recommendations

The suggestions below come from the evidence and gaps in Sections 5 to 11. They are specific not general as the review suggests.

12.1 For Universities

- Create a group with departments to handle AI policy. This will help avoid enforcement.
- Do not use AI detection scores alone in cases of misconduct. Use evidence and have people decide.
- Give all students access to AI tools to reduce the gap for those who cannot afford them.
- Focus on changing how assessments are done not on detecting AI.

For Faculty

- Use rules about AI use in each class not just banning it.
- Change high-stakes tests to formats that require direct work not just written answers.
- Treat AI detection scores as one sign, not proof. Be careful with students who're not native English speakers.
- Get training on AI to use it better.

For Students

- Say when you use AI in your work. This helps avoid problems with detection tools.
- Use AI to help you understand not to do the work for you.
- Check any information from AI before using it in your work.

For Researchers

- Do studies that follow the students over time. This is missing now.
- Report how AI was used and not just the results. This can explain some of the conflicting findings.
- Test AI detectors for differences in performance, across groups.

For Policymakers

- Make rules to protect student data when using AI.
- Ask AI companies to show proof of how their tools work and if they are biased.
- Fund big studies that look across schools and countries.

For Educational Technology Developers

- Make it a priority to reduce the number of citations and incorrect information that is created and clearly tell institutional customers about the risk of incorrect information that still remains because there is a lot of evidence that shows that false information is a problem in academic settings (Walters & Wilder 2023).
- Make sure that AI detection tools are tested in an open way before and after they are used by institutions and share the details of how accurate these tools are because there have been serious problems with fairness that were noted in Section 7.4.
- Create pricing models that do not make it harder for some students to use AI than others in response to the problem of costs that stops some students with disabilities and students from low-income families from using AI (Section 6.5).

13. Future Research Directions

13.1 Longitudinal Studies

The important missing area found in this review is the lack of long-term studies that track how students improve in skills, thinking and writing over the course of an entire academic program when they use AI all the time. Future studies should follow groups of students for least two to three years and if possible collect data before they start using AI to see the real effects of using AI over time instead of just looking at one moment in time which is what most studies do now.

13.2 Cross-Country Comparisons

Most studies of how students and teachers think about AI are done in a few countries and the little cross-national information that exists shows that there are real differences in how people feel how they use AI and how schools manage AI in different parts of the world. Studies that compare countries using the same methods in many education systems would help to find out which effects of AI are the same everywhere and which are only found in certain places.

13.3 Experimental Studies

Most of the information about AI and how it affects learning is based on connections or descriptions not on experiments. Studies that use similar methods to compare how students learn when they use AI in a structured way versus when they use it without guidance or not at all and do this in the same tasks would help to understand better whether AI helps or replaces learning as discussed in Sections 9.2 and 9.6.

13.4 Student Perception Studies

More research should look into how students feel about fairness and proper procedures when they are accused of cheating because of AI detection tools. The current research shows more about the problems with these tools (Liang et al. 2023) than about how students actually experience being investigated because of them.

13.5 Faculty Studies

Future studies about teachers should go beyond just asking how they feel about AI and look at whether certain training programs change how they design tests and use detection tools. This review found that training for teachers not just having tools seems to be the factor that influences how teachers use AI (Section 8.3).

13.6 Policy Evaluations

Few of the policies about AI that were reviewed in Section 10.5 have been tested to see if they actually change how many students break rules how satisfied students and teachers are or how fair assessments are. Studies that compare policies before and after they are used or look at changes in how AI is used would help build a stronger evidence base for policy recommendations including the ones in Section 12.

13.7 AI Literacy Research

Because AI literacy is a factor in how teachers use AI (Section 8.2) and how students think critically (Section 9.3) future studies should create and check standardized ways to measure AI literacy for both students and teachers. This would help schools find gaps, in understanding and see how well training programs work, of just using informal methods.

13.8 Assessment Research

Finally studies that compare the quality, reliability and how students feel about ways of changing assessments like using similar assignments oral tests or tests that match the learning context would be a clear and doable research goal. These studies should compare the methods instead of comparing each one to an unclear standard.

Table 6. Research Gaps and Proposed Future Research Directions

Illustrations are not included in the reading sample

14. Limitations

This review has limitations that should be kept in mind when interpreting its conclusions. First as explained in Section 4.10 it is a single-author narrative synthesis, not a pre-registered dual-reviewer PRISMA systematic review. It does not calculate pooled effect sizes. When effect sizes are mentioned they come from the analyses cited in the review (such as Ma et al., 2014) not from independent calculations made by the author.

Second the review relies on web-based searches rather than direct access to comprehensive licensed databases like Scopus or Web of Science. Although multiple independent sources were used to check claims there is no guarantee that the coverage of the literature is complete.

Third the reference list for this review was intentionally limited to sources where core bibliographic details could be verified directly. This improves the reliability of each citation. Means the review includes a more limited pool of studies compared to a large-scale systematic review with full database access. Some relevant studies may have been missed.

Fourth the field is changing quickly. New large-scale surveys, policy updates and AI model releases happen every month. As such many of the findings—especially those about detector accuracy and adoption rates—reflect the state of the evidence at the time of writing and should not be seen as fixed or final.

Fifth the review is limited to English-language sources. This may overlook research and policy developments in higher education systems that operate in other languages.

Finally as a review article and not original empirical research this work cannot fill the evidentiary gaps identified in Section 11.5 and Section 13. It can only describe those gaps clearly as the available literature allows.

15. Conclusion

This review aimed to examine how artificial intelligence, generative AI is changing higher education. It also sought to weigh the benefits against the risks using peer-reviewed research and institutional evidence available as of 2026.

In response to the research question the evidence shows that AI is being used across nearly every part of the university—from personalized tutoring and automated feedback to research support and administrative tasks. For the research question the strongest and most consistent opportunity found in the literature is moderate learning gains from intelligent tutoring systems. This evidence comes from a established decade-long body of research (Ma et al., 2014; VanLehn, 2011). In contrast evidence on purpose generative AI tools is newer, less consistent and more mixed.

Regarding the research question the review highlights serious risks that are well documented. For example early versions of AI tools produced citation fabrication rates of more than 50% in some cases (Walters & Wilder 2023). Also detection tools have been shown to mislabel the work of many non-native English speakers as AI-generated (Liang et al., 2023). These errors have consequences, including due-process violations affecting real students.

For the research question generative AI has challenged the basic assumptions behind traditional assessment and academic integrity policies in ways that earlier educational technologies did not. As a result there is a growing shift from relying on detection tools toward policies that emphasize disclosure and a focus on assessment design that ensures validity (Gonsalves, 2025; Roe et al. 2025).

On the research question the evidence points to institutional governance as the key factor that determines whether AI adoption leads to positive or negative outcomes. This aligns with the framework discussed in Section 2.5. It is not the presence or absence of AI tools that matters most. How institutions govern their use.

The main takeaway from this review is that responsible AI adoption in education is not about choosing more or less technology. It is about finding and maintaining an approach between innovation and human judgment. It means balancing the benefits AI offers—like improved efficiency and access—with the real risks it presents—such as threats to academic integrity and fairness. It also means weighing the promise of personalization against ethical responsibilities around privacy, equity and human accountability.

No single policy, detection tool or assessment method can settle this balance permanently. Technology changes fast for that.. The evidence reviewed here supports a set of enduring principles: disclosure over concealment scaffolding over substitution human judgment over algorithmic determinism and continuous evaluation, over static policies. These principles can help institutions navigate technological change even as the tools themselves keep evolving.

Universities that embrace these principles—instead of blindly accepting AI or rejecting it entirely—are, based on current evidence best positioned to protect both the quality of education they deliver and the integrity of the credentials they award.

References

Alkaissi, H., & McFarlane, S. I. (2023). Artificial hallucinations in ChatGPT: Implications in scientific writing. Cureus, 15(2), e35179. https://doi.org/10.7759/cureus.35179

Baidoo-Anu, D., & Ansah, L. O. (2023). Education in the era of generative artificial intelligence (AI): Understanding the potential benefits of ChatGPT in promoting teaching and learning. Journal of AI, 7(1), 52–62.

Cardona, M. A., Rodríguez, R. J., & Ishmael, K. (2023). Artificial intelligence and the future of teaching and learning: Insights and recommendations. U.S. Department of Education, Office of Educational Technology.

Chan, C. K. Y., & Hu, W. (2023). Students’ voices on generative AI: Perceptions, benefits, and challenges in higher education. International Journal of Educational Technology in Higher Education, 20, Article 43. https://doi.org/10.1186/s41239-023-00411-8

Cotton, D. R. E., Cotton, P. A., & Shipway, J. R. (2023). Chatting and cheating: Ensuring academic integrity in the era of ChatGPT. Innovations in Education and Teaching International, 61(2), 228–239. https://doi.org/10.1080/14703297.2023.2190148

Dalalah, D., & Dalalah, O. M. A. (2023). The false positives and false negatives of generative AI detection tools in education and academic research: The case of ChatGPT. The International Journal of Management Education, 21(2), 100822.

Digital Education Council. (2025). Global AI student survey 2025.

Gonsalves, C. (2025). Contextual assessment design in the age of generative AI. Journal of Learning Development in Higher Education, (34). https://doi.org/10.47408/jldhe.vi34.1307

Kumar, R., & Mindzak, M. (2024). Who wrote this? Detecting artificial intelligence–generated text from human-written text. Brock University Learning and Teaching.

Labadze, L., Grigolia, M., & Machaidze, L. (2023). Role of AI chatbots in education: Systematic literature review. International Journal of Educational Technology in Higher Education, 20, Article 56. https://doi.org/10.1186/s41239-023-00426-1

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7).

Ma, W., Adesope, O. O., Nesbit, J. C., & Liu, Q. (2014). Intelligent tutoring systems and learning outcomes: A meta-analysis. Journal of Educational Psychology, 106(4), 901–918.

Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. https://doi.org/10.54675/EWZM9535

Roe, J., Perkins, M., & Giray, L. (2025). Assessment twins: A protocol for AI-vulnerable summative assessment (arXiv:2510.02929). arXiv. https://doi.org/10.48550/arXiv.2510.02929

University of Virginia Teaching Hub. (2024). Using generative AI to make learning more accessible in higher education.

VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221. https://doi.org/10.1080/00461520.2011.611369

Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, Article 14045. https://doi.org/10.1038/s41598-023-41032-5

Yusuf, N. M. (2024). Generative AI in education and research: A systematic mapping review. Review of Education, 12, e3489. https://doi.org/10.1002/rev3.3489

[...]

Ende der Leseprobe aus 51 Seiten  - nach oben

Jetzt kaufen

Titel: The future of Education with AI. How AI affects Higher Education

Rezension / Literaturbericht , 2026 , 51 Seiten , Note: N/A

Autor:in: Anonymous (Autor:in)

Informatik - Künstliche Intelligenz
Blick ins Buch

Details

Titel
The future of Education with AI. How AI affects Higher Education
Untertitel
Opportunities, Challenges, Academic Integrity and Future Implications. A Systematic Literature Review
Hochschule
The Islamia University of Bahawalpur  (The Islamia University Bahawalpur)
Veranstaltung
Computer Science
Note
N/A
Autor
Anonymous (Autor:in)
Erscheinungsjahr
2026
Seiten
51
Katalognummer
V1763890
ISBN (PDF)
9783389206799
Sprache
Englisch
Schlagworte
Intelligence Generative AI Higher education Academic integrity Academic misconduct AI detection Personalized learning Intelligent tutoring systems Algorithmic bias Digital divide AI governance Assessment redesign
Produktsicherheit
GRIN Publishing GmbH
Arbeit zitieren
Anonymous (Autor:in), 2026, The future of Education with AI. How AI affects Higher Education, München, GRIN Verlag, https://www.grin.com/document/1763890
Blick ins Buch
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
  • Wenn Sie diese Meldung sehen, konnt das Bild nicht geladen und dargestellt werden.
Leseprobe aus  51  Seiten
Grin logo
  • Grin.com
  • Versand
  • Kontakt
  • Datenschutz
  • AGB
  • Impressum
  • Vertrag widerrufen