University AI Policy: What Schools and Journals Actually Allow

On 2 June 2026, the University of Sydney's page on responsible AI use still read the way it has since the university's default flipped in 2025: AI is allowed in open assessments unless a lecturer says otherwise, the reverse of where the rule stood two years earlier. A university AI policy is not a fixed document. It is a live position that moves as fast as the technology it tries to govern, which is why reading one page from one institution, then assuming it speaks for the sector, is the single most common mistake a student or a researcher makes.

This piece maps the whole landscape instead of one corner of it: what universities require for coursework, what they require for research, what journal publishers require for submission, and what the Committee on Publication Ethics says about authorship itself. Every policy quoted below is drawn from the institution's own page, read in August 2026, with the page's own revision date noted wherever it gave one.

There's no consensus to summarize here, and pretending otherwise would misrepresent the sources. Toronto bars generative AI by default. Sydney allows it by default. Both call themselves an academic integrity policy, and both are current as of 2026. This piece keeps the disagreement visible instead of editing it away. Smoothing that gap into one imagined rule would be easier to write and less accurate.

What Does a University AI Policy Actually Say?

This depends entirely on which university, and often on which assessment inside that university. The table below checks five institutions across five countries: what each one permits by default, and exactly what a student has to disclose once AI use is allowed. The disclose column is usually the more demanding one, even at the institutions that allow AI freely.

Institution	Country	Default stance	What you disclose
University of Sydney	Australia	AI allowed in open, unsupervised assessments since 2025; barred in supervised assessments unless the unit coordinator says otherwise	The AI tool's name and version, its publisher, the URL, and a description of how you used it
University of Toronto	Canada	AI barred by default in every course; using it without permission is an academic offence	Nothing by default; once an instructor grants permission, whatever that instructor specifies
Imperial College London	United Kingdom	AI allowed for coursework unless the assignment brief says otherwise, provided the submitted work is in the student's own words	The tool, its publisher, the URL, and a description of its contribution, cited like any other source
Columbia University	United States	Set centrally by the Provost, then applied school by school; Columbia Business School requires disclosure, and researchers may not enter unpublished data into a generative AI tool	Which platform was used and how, reported to the relevant school or instructor
National University of Singapore	Singapore	AI permitted if acknowledged; the Code of Student Conduct names undisclosed AI-generated content as a form of academic dishonesty	Which tool contributed, following the department's own acknowledgment guidance

Two things hold across all five, even though the default does not. Disclosure survives every policy that allows AI at all: no institution above lets a student use AI silently, only openly or not at all. And the penalty for the same mistake, using AI without saying so, is treated as a variant of an old problem rather than a new one: submitting help you did not do yourself and did not name.

University AI Policies Compared: Oxford, Cambridge, Harvard, MIT and Monash

Every row below answers the same three questions, because those are the three that actually matter once you are sitting in front of an assignment: what the default allows, what you have to declare, and who set that rule in the first place.

Institution	Permitted by default	What must be declared	Who actually sets the rule
Oxford	Banned in summative assessment unless the department, faculty or assessment-setter explicitly allows it	AI use, following the department or faculty's own instructions, whenever it is permitted	The department, faculty or individual assessment-setter, not the central governance page
Cambridge	Allowed for personal study, research and formative work; unacknowledged AI content in a summative assessment is misconduct unless the brief says otherwise	Any AI-generated content used in a summative assessment, unless the brief states otherwise	The individual assessment brief, which can override the university-wide default
Harvard (HBS MBA cited)	No blanket rule; each course decides whether tools like ChatGPT are allowed at all	AI tool use, cited the way the MBA Honor Code requires wherever it is permitted	The individual course and its instructor
MIT (Sloan cited)	No institute-wide rule on using generative AI for assignments	Whatever the individual instructor's own course policy requires	Each school and instructor, separately; MIT's only central rule concerns data handling, not assignments
Monash	Set unit by unit; the Chief Examiner specifies where and how AI may be used	AI use, always, in line with what the Chief Examiner specified for that unit	The Chief Examiner of each individual unit

Read down the last column and the shape of the whole piece is right there. Not one of these five puts the decision at the level most students assume it sits: the university-wide policy page. Oxford and Cambridge push it down to the department, the faculty or the assessment brief. Harvard and MIT push it down to the course. Monash names a single role, the Chief Examiner, and gives that person the call for every unit they run. A central page can tell you the default. It can't tell you what your specific module allows.

Why Doesn't a Central Policy Page Settle What You Can Do?

Because the people writing a university-wide policy are not the people setting an individual assessment. A registrar or an academic-integrity office can state a default and a disclosure expectation, but marking, invigilation and assessment design happen at department or module level, so the actual permission a student needs sits with whoever wrote that specific brief. A joint-honours student can sit two different AI rules in the same term, one for each half of their degree, and both are correctly described as "the university's policy" even though they say different things.

That is why the same institution can say two true things that look contradictory from the outside. Cambridge's central page permits AI for formative work while also stating that unacknowledged use in a summative assessment is misconduct by default. Both sentences are accurate at the same time, because they describe two different tiers of work, and the assessment brief for any specific piece is the document that actually decides which tier applies. A formative essay set in week three and a summative dissertation submitted in June can sit under the same institutional page and land on opposite sides of that line, and nothing on the page itself flags which tier a given piece of work belongs to. That has to come from the brief.

Do Oxford and Cambridge Actually Disagree?

Not on the structure, only on where the default starts. Oxford's policy on AI use in summative assessment states plainly: "use of AI in assessments is only allowed when explicitly permitted and must be declared following department or faculty instructions; unauthorised use is considered academic misconduct." The default is off. It switches on only when a department or faculty says so, in writing, for that assessment. In practice, that permission usually shows up as a single line in a module handbook or an assignment brief rather than a separate document of its own, which is exactly why searching the central governance page for a yes-or-no answer wastes the time it takes to find the actual brief.

Cambridge's policy opens from a more permissive default and narrows from there: "Students are permitted to make appropriate use of artificial intelligence tools to support their personal study, research and formative work." For summative assessment specifically, the university's own wording flips the burden onto disclosure: "A student using any unacknowledged content generated by artificial intelligence within a summative assessment as though it is their own work constitutes academic misconduct, unless explicitly stated otherwise in the assessment brief." Oxford starts closed and opens by permission. Cambridge starts open and closes by omission. Both hand the actual decision to a document smaller than the policy page itself.

Harvard and MIT: Why Does the Course Decide, Not the Institute?

Because neither university's central guidance sets a usable rule for a specific class, so the practical answer comes from whoever teaches it. Harvard's central guidance leaves the permitted-or-not call to each course; Harvard Business School's MBA Program Handbook spells out what that looks like at the school that has actually written it down: "Faculty may allow the use of ChatGPT and similar technologies in some circumstances, but students must ensure that such use is permitted and must make sure they understand any limitations on such use before using this technology." Where it is permitted, citation is not optional: "students must cite their use of these AI tools appropriately. Not doing so violates the MBA Honor Code." Harvard's university-wide Provost guidance, first published 13 July 2023, sets the same expectation at a higher level: it leaves the permitted-or-not call to each course, rather than setting one campus-wide default, which is why the HBS-specific wording above is the more useful document for an actual student to read.

MIT's Sloan MBA Program Handbook states the same devolved model in so many words: "MIT Sloan does not have a definitive policy on using generative AI tools (e.g., ChatGPT) to complete individual or team assignments or prepare for a class discussion. Rather, faculty will set their own policies pertaining to the use of AI tools in their courses, including permitting or prohibiting some or all uses of such tools." The one rule MIT does set centrally has nothing to do with academic integrity at all: "MIT work should be conducted using MIT-licensed AI tools whenever Institute data is involved." That is a data-handling requirement, not a permission slip for an assignment. A student could be fully within a course's own AI policy for an assignment and still breach the data rule separately, for instance by pasting confidential Institute data into a public tool with no academic-integrity angle at all. The two rules run on entirely different tracks.

Who Sets Monash's Rule: the University or the Chief Examiner?

The Chief Examiner, by name, for every unit they run. Monash's policy makes Chief Examiners responsible for the assessment regime in a unit and requires them to specify where and how AI may be used in that unit's assessments, which puts the rule at unit level rather than university level. Once a Chief Examiner has specified a use, the acknowledgement duty is constant rather than conditional: students have to acknowledge AI use clearly and openly, in line with what the Chief Examiner set out, and the policy puts the requirement in absolute terms, so generative AI used to produce work submitted for assessment always has to be acknowledged. Two students in the same course can therefore work under identical rules on paper and still land in different places, if one follows what their Chief Examiner actually specified and the other assumes a general Monash-wide answer that the policy never promised.

Why Do Coursework Rules and Research Rules Differ at the Same University?

Because they answer to different people. A coursework policy protects the grade: it exists so a mark reflects the student's own understanding, and it usually sits with a registrar, a dean of students, or an academic integrity office. A research policy protects the record: it exists so a published or submitted finding can be trusted, and it usually sits with a research ethics board, an IRB, or the office that signs off on responsible conduct of research. The same university can run both at once, worded differently, enforced by different offices, and rarely cross-referenced on the page a student actually finds.

The Columbia example is clean because the division is visible on their own webpages. The Business School has a student-facing policy that requests students to report their use of the platform and how they used it for classwork. There is also a researcher-facing policy that prohibits posting unpublished data or confidential information into any type of generative AI tool whatsoever. Finally, Teachers College's institutional review board requires researchers to disclose the use of AI as part of human subjects research approval. Three rules, one university, three different offices, and a student who only reads the undergraduate page never sees the other two.

What Do Journal and Publisher AI Policies Require?

They require disclosure almost everywhere, and they agree, without exception, that an AI system cannot be listed as an author. Where they genuinely differ is the boundary of what counts as a disclosable use, and how strict the ban is on AI-generated images and figures.

Publisher	AI as a listed author?	What must be disclosed	One notable restriction
Elsevier	Never; authorship duties "can only be attributed to and performed by humans"	A separate declaration naming the tool and its purpose, unless the use was a basic grammar or spelling check	Generative AI cannot create or alter artwork, figures, or primary research images
Springer Nature	Never; an LLM "cannot take accountability and do not meet authorship criteria"	Acknowledged in the introduction, preface, or acknowledgements when AI helped generate text, analysis, or content	Copy editing done only for readability, grammar, or formatting does not need declaring
Wiley	Never, by name	Disclosed in the methods section, a dedicated disclosure statement, or the acknowledgements, except for spelling and grammar tools	Manuscripts cannot contain unverified AI-written sections, including an abstract or a conclusion
IEEE	Not framed as an authorship question at all; the policy is written purely as a disclosure rule	Named in the acknowledgments, with the specific AI system identified and the affected sections explained	A reviewer may not run a manuscript under review through a public AI platform
Taylor & Francis	Never; authorship carries “uniquely human responsibilities” an AI tool cannot undertake	Tool name, version, purpose and method, placed in the methods or acknowledgements, or at the proposal stage for a book	Generative AI cannot create or manipulate images, figures, or research data

The one genuinely new split in this set is IEEE, which never frames its rule as a question of authorship at all. Every other publisher above states outright that AI cannot be an author; IEEE's policy simply never raises the question, because it is written entirely as a disclosure requirement for content, whoever or whatever produced it. That is a real difference in how a policy is built, not just in what it says.

Can AI Be Listed as an Author?

No, according to every publisher named above and to the Committee on Publication Ethics, whose position on the question has stood since 13 February 2023. COPE's stated reasoning isn't really about writing quality. Authorship is partly a legal and ethical status, not just a description of who typed the sentences. An AI tool can't take responsibility for a submitted paper, cannot hold copyright, and cannot declare or deny a conflict of interest. COPE states the disclosure requirement plainly: authors who use AI tools in preparing a manuscript must disclose, in the methods section or its equivalent, how the tool was used and which tool it was. ICMJE, WAME, and the JAMA Network hold the same line, which is part of why the rule reads as settled rather than as one publisher's house style. Disclosure is the part the guidance turns on. It does not matter if the writer of the text is a person or an AI program. All authors have responsibilities under the rule.

What Does an AI Disclosure Statement Actually Include?

Every major policy, across both universities and publishers, converges on the same four pieces of information, even when they disagree on almost everything else. A usable disclosure statement names:

• The specific tool and version, not just "AI" as a category
• What it was used for: drafting, summarizing, editing, translating, or generating code or analysis
• Confirmation that a human reviewed and verified the output against the original sources
• Where the disclosure lives: a methods section, an acknowledgements line, or a dedicated statement, depending on what the target journal or course asks for

That short list matters more than the yes-or-no stance a policy takes, because a policy that permits AI but accepts a vague disclosure protects nobody, and a policy that bans AI but never checks for it protects nobody either. If your discipline expects the tool cited rather than just named in a sentence, a citation generator for AI tools formats that entry the way a reference list expects. The four items above are what a reader, an editor, or an examiner can actually verify regardless of the exact format.

Is Using an AI Humanizer Against Your University's Policy?

Not if you are editing an argument you actually made. That's what every policy in the tables above covers: language editing, restructuring your own sentences, adjusting tone, varying wording. That's exactly what supervisors and copyeditors have done forever. Editing your own drafted argument with a humanizer falls into the same category, but make sure to declare it the way your institution or journal asks. TextPulse's own position on the ethical use of AI in academic writing sets out that distinction in full: editing your own work and submitting an argument you never made are different acts, and only one of them is misconduct.

The tables above make the boundary concrete in another way too. Imperial's rule that submitted work has to be in the student's own words, and Sydney's requirement to name the tool and describe how you used it, both assume the same thing: you are allowed to end up sounding like yourself, not like an unedited model output. A humanizer built for student coursework is one route to that, and the mechanics of doing it section by section, since a methodology reads differently from a discussion, are covered in a separate piece on humanizing AI text in academic writing.

What Happens If You Do Not Disclose AI Use?

It varies by institution and publisher, but it looks like this: A university that detects undisclosed AI use in submitted coursework classifies it as an academic integrity issue (just like undisclosed help from a person). The specific policies vary; for example, the penalty could be a request for a resubmission or formal proceedings for misconduct. A journal that detects undisclosed AI use in a manuscript after accepting it has the option to correct, retract, or ask for a revised submission. An editor who suspects fabricated citations verifies them against the original sources and decides whether to accept or reject the paper. If you haven't picked a target journal yet, a journal finder that matches your manuscript to journals already publishing similar work is worth trying before you check that journal's specific policy.

The policy questions underneath this each get their own treatment. How to disclose AI use in a paper and a ready disclosure statement template are set out one at a time, and so is the question of whether a model can be listed as an author. On the enforcement side there are pieces on what happens if you are caught and on whether humanizers themselves breach policy. The prior question, whether using AI to write an essay counts as cheating at all, is answered on its own page, and journal policies are compared in a separate piece.

None of this settles into a single rule you can memorize and reuse, and that is the point rather than a gap in the research. Check the specific page for your own institution and your own target journal before you submit anything, because the version that mattered was updated more recently than this article was.