Publications

1. Introduction

This report reviews the publicly available guidance documents of leading Western universities regarding the integration of generative artificial intelligence into the teaching process. The research draws on both the policies of individual institutions and the sectoral framework documents on which these institutions rely.

Despite differences in regulatory environments, funding models, and institutional cultures, the world's leading universities share a substantially similar policy architecture: (1) the use of artificial intelligence is permitted subject to transparency and ethical norms; (2) specific rules for use/restrictions are delegated to the levels of schools or lecturers; (3) AI literacy is recognized as a fundamental graduate skill; (4) rethinking assessment methods, rather than bans or total control, is considered a necessary long-term condition.

The main distinguishing factors among universities are the degree of centralized policy, as well as the standards for AI-use guidelines and the legal frameworks (GDPR, FERPA, Privacy Act, FIPPA) that define data protection and security standards.

For universities in Georgia that are at the initial stage of developing institutional AI policy, international experience offers a unified guide and set of practices on how to integrate artificial intelligence into the teaching process ethically and transparently. Chapter 5 of the research report provides general recommendations for universities.

2. Methodology and Scope

In the “Times Higher Education” 2026 World University Rankings, the University of Oxford ranks first for the tenth consecutive year. The upper portion of the top fifty is again dominated by Western institutions (the U.S., United Kingdom, continental Europe, Canada, and Australia), alongside Asian institutions (mainland China, Hong Kong, Singapore, Japan). This report focuses on the Western universities included in this top 50.

Given the large number of institutions and the volume of policy documents, the report combines three categories of complementary sources:

• A direct review of AI/GenAI guidance documents published by more than 25 leading Western universities in the United Kingdom, the U.S., Switzerland, the Netherlands, Germany, Canada, and Australia;
• National framework documents (where such exist) which universities rely on;
• Comparative academic and peer-reviewed studies analyzing the institutional policies of dozens of universities.

3. Regional Analysis

3.1 United Kingdom – The Russell Group Model

The UK’s Russell Group, which unites 24 research universities, including Oxford, Cambridge, UCL, Edinburgh, Manchester, King's College, LSE, and Bristol University, published five joint principles in July 2023:

1. Spreading and supporting AI literacy;
2. Equipping staff with the relevant knowledge to help students use generative AI tools effectively and appropriately;
3. Adapting teaching and assessment to integrate the ethical use of generative artificial intelligence and ensure equal access;
4. Strengthening the protection of academic ethics and integrity;
5. Inter-university collaboration to share best practices in parallel with technological development.

Individual universities implement these principles differently in practice:

• Oxford’s approach is as follows: AI may be used to support learning, to generate new ideas, and for technical tasks; however, when used, the student is obligated to disclose it.
• At the University of Cambridge, the use of AI is permitted for personal learning and research but is restricted for summative assessment without permission. The university openly states that AI detectors do not work, which is why it refuses to implement a centralized detector.
• Imperial College London introduced “vivas” (oral examinations), in which a student may be asked to orally defend a submitted work to confirm that it reflects their own knowledge.
• King’s College London (KCL) created one of the most structured frameworks, organized on three levels: macro (university-wide), meso (department/program), and micro (individual lecturer and assessment design). Interestingly, KCL does not ask students to cite generative AI as a source in the bibliography (since it does not recognize it as a scholarly source); however, when used, it requires disclosure.
• The University of Edinburgh published the most detailed list of prohibited actions, which includes the use of AI in translating a work. Such use of the tool is interpreted as pseudo-authorship; the use of AI agents and browsers within the university’s virtual learning environment (VLE) is also restricted.

3.2 United States - Decentralized, Lecturer-Oriented Governance

The U.S.’s elite universities (Harvard, MIT, Stanford, Princeton, etc.) share a common architecture: general guidance on institutional principles (focused mainly on the honor code, data security, and ethics) and broad authority delegated to individual schools, departments, and lecturers. Almost every American university’s guidance repeats the phrase: “Follow your lecturer’s instructions, and if the syllabus is silent, assume that the use of artificial intelligence is prohibited”.

U.S. universities differ from British ones in that their central guidance documents emphasize information security and data classification more than pedagogy. For example, Harvard's central IT guidance prohibits providing confidential data to artificial intelligence tools (research data, HR, finances, student records). Similarly, MIT prohibits entering “high-risk” institutional data when using any AI tool, and in the case of “low-risk” and “medium-risk” data, it strongly recommends using only those AI tools that have an official licensing agreement with the university.

Unlike Harvard and MIT, at Stanford the policy varies by program level. For example, in undergraduate programs, the use of artificial intelligence is disallowed a priori unless the lecturer explicitly permits its use. However, in the graduate business program, for instance, lecturers are not permitted to prohibit students from using AI on take-home assignments or exams, though they retain control over in-classroom assessments.

Princeton University requires students to obtain the lecturer's permission before using AI for any assignment, as well as to disclose its use.

3.3 Continental Europe - Fairness, Transparency, and GDPR

Continental Europe’s leading universities (ETH Zurich, EPFL, the Technical University of Munich (TUM), the University of Amsterdam, and TU Delft) focus in their AI guidance documents on values such as responsibility, fairness, and transparency, with the EU’s General Data Protection Regulation (GDPR) serving as the legal framework. For example, in ETH Zurich's guidance, the protection of confidentiality and copyright falls under the definition of the “principle of fairness”. Moreover, the university has licensing agreements with several major providers (Microsoft Copilot, Google Gemini, NotebookLM) and advises students to use these protected, licensed tools in the learning process. In this way, the university protects users’ personal data to prevent its use for the further training of models. Similar to ETH Zurich, TU Delft's guidance devotes particular attention to protecting users’ personal data, which, beyond the confidentiality policy, also falls within the protection of the intellectual property principle.

3.4 Canada - Privacy First and the Strictest Exception

Canadian universities (Toronto, McGill, and UBC) are distinguished by the fact that, before deploying artificial intelligence systems, they assess their potential impact on privacy and personal data protection (PIA). In addition, the approach of Canadian universities is based on national legislative documents (FIPPA and the federal “FASTER” principles).

McGill University appears in this report as the most restrictive institution: according to its guideline, if a tool has not been officially reviewed and approved by the university, the use of any generative AI tool for purposes related to McGill University is prohibited.

The University of Toronto, on the contrary, uses a more standard, permissive approach with safety guardrails, under which the enterprise version of Microsoft Copilot is considered the recommended and protected tool.

3.5 Australia - Systemic Reform Coordinated by the Government

Australia represents the most centrally coordinated model. In June 2024, the national regulator (TEQSA) required all higher education institutions in the country to submit an institutional action plan aimed at reducing and managing the risks to academic integrity arising from the spread of generative AI.

By 2025, TEQSA’s approach had changed, moving from AI detection and control toward a longer-term strategic vision. According to the regulator, orienting toward AI development puts universities at a disadvantage, because in terms of timing, institutions’ responses are always reactive rather than proactive. Therefore, the agency developed three alternative reform models: (1) rethinking learning outcomes across an entire academic degree (bachelor’s/master’s) program; (2) embedding at least one “secured” assessment task at the subject/module level, meaning a supervised exam, an oral presentation, an in-class assignment, and so on; or (3) a hybrid of these two approaches.

In addition, the experience of Monash University is noteworthy: it created a unified, well-structured guide to AI citation formats (according to various citation styles, APA, MLA, Harvard), which other international universities, for example, Duke University in the U.S., cite as an additional resource.

4. Common Characteristics and Similarities

Despite different starting points, several core principles recur in nearly every university’s framework document:

Transition from banning to managing: complete bans have been replaced by conditional-permission models, because AI is already integrated into everyday applications such as Word, Excel, and Google Docs, which makes bans impossible.

Delegation of authority to the course/lecturer level: students are clearly required to check the policy of the specific course.

AI literacy as a primary goal: developing the capabilities of staff and students to use AI critically is considered a priority direction.

Disclosure rules: when AI is used, most universities require transparency or specific citation.

Skepticism toward AI-detector software: academic circles agree that detectors are unreliable and cannot serve as a protective mechanism.

Rethinking the assessment system: the long-term solution is to change assessment formats (oral exams, in-person assignments, portfolios) rather than maintaining old formats.

Data classification: a strict directive that confidential, personal, or unpublished research data must not end up in the databases of free AI tools.

Enterprise licensing: providing a secure and “protected” tool (often Microsoft Copilot), which simultaneously solves the problem of both data protection and student equity.

The necessity of continuous updating: in their guidelines, universities emphasize that the documents are subject to periodic review in line with technological development.

5. Key Findings and Recommendations for Georgian Universities

The experience of the West's leading universities offers a kind of trajectory for the responsible and ethical implementation of artificial intelligence in Georgian universities. Western experience is relevant in the context of applying best practices as measured by international rankings, taking into account both Georgia's constitutional and historical aspirations and other objective factors.

Over the past three years, Western universities have converged on several important recommendations that can be considered in the process of properly integrating artificial intelligence:

• Encouraging the use of artificial intelligence

As shown by the example of Western universities, total bans do not work in practice. Instead, the use of artificial intelligence is permissible on the condition of its ethical use; ultimately, however, the issue regarding a specific syllabus is decided by the lecturer.

• Changing assessment methodology and retraining staff

Universities unanimously agree that no reliable tool exists for detecting artificial intelligence. Accordingly, it is important that the approach be proactive rather than reactive. To this end, resources should be directed toward retraining academic staff, changing assessment criteria, and teaching oriented toward the development of critical thinking and skills.

• Drafting simple data-processing rules

It is necessary to develop clear guidance defining what type of data is impermissible to use in public, free artificial intelligence tools. This guidance should comply with Georgia's data protection legislation.

• Licensing a university-wide, protected artificial intelligence tool

Following the European and Canadian examples, reaching an enterprise agreement, for example, using university accounts on the Microsoft Copilot platform, would simultaneously address two objectives: ensuring equality of access, so that the ability to purchase a premium version does not turn into an academic advantage for individual students, and data protection.

• Prioritizing artificial intelligence literacy

AI literacy should be presented as a prerequisite for a graduate's competitiveness in the labor market.

• A mechanism for periodic review of the policy document

Any policy document should, from the outset, provide for a regular, for example, annual, review cycle, insofar as the technological and legal environment is subject to constant and rapid change.

For the complete document, including relevant sources, links, and explanations, please see the attached file.