The Ethics of AI in Research Writing: What Universities Should Expect From an AI Research Assistant (2026)

We build Litrevu, so we have a stake in this question, which is a reason to be precise about where the line sits.

What does ethical AI use mean in research writing?

Ethical AI use in research writing means a human author keeps full responsibility for every claim, citation and sentence in the work, uses AI only in ways the institution permits, and discloses that use openly. The Committee on Publication Ethics (COPE) states the core of it plainly: "AI tools cannot be listed as an author of a paper", because they "cannot take responsibility for the submitted work". Responsibility is the test, not the tool.

That test rules out more than ghost-writing. An invented citation, a summary that misreads a paper, or an argument the student could not defend in a viva all fail it, because nobody responsible checked them. Searching, reading, comparing and checking sources all pass it, because the author stays responsible for the result.

Source: COPE position statement on authorship and AI tools.

What do journals and publishing bodies require when researchers use AI?

Publishing bodies require three things: AI cannot be an author, authors must disclose AI use, and humans remain responsible for accuracy and attribution. COPE asks authors to disclose "in the Materials and Methods (or similar section) of the paper" how and which tool they used. The International Committee of Medical Journal Editors (ICMJE) goes further on where disclosure belongs, asking authors to describe AI use "in both the cover letter and the submitted work", and warns that failing to disclose "may be construed as misconduct in some circumstances".

ICMJE also names the risk that matters most for literature reviews: "Humans must ensure there is appropriate attribution of all quoted material, including full citations", and AI output should be reviewed because it can be incorrect, incomplete or biased. Examiners hold a thesis chapter to the same standard in spirit, and treat a fabricated citation as an integrity problem rather than a typo.

Sources: COPE, ICMJE recommendations on AI use by authors.

What do universities allow students to do with AI?

Universities differ, and the honest summary is that the module or assessment brief decides. Some institutions start from permission with conditions, and others start from prohibition with exceptions. Here is how four published positions compare, as of October 2026.

The pattern across all four is that AI literacy and transparency are expected, and that detection is not treated as the answer. That puts the weight on how a tool is designed and how a student uses it, which is where an assistant can help or harm.

Sources: Russell Group principles, University of Birmingham generative AI guidance, UCT AI in Education Framework announcement, UNESCO guidance.

How many students already use AI for assessed work?

Nearly all of them, at least in the UK. The Higher Education Policy Institute’s Student Generative AI Survey 2026, published with Kortext on 12 March 2026 and based on 1,054 full-time undergraduates, found that "95% of students now use AI in some form" and "94% use generative AI to support assessed work". The same release reports that "fewer than half feel supported by their teaching staff".

Those are UK undergraduates, not South African postgraduates, and we would not stretch the figures further. But when use is close to universal, a ban mostly decides whether AI use is visible, not whether it happens. The useful questions become which tools students use, what those tools encourage, and whether students know how to declare what they did.

Source: HEPI Student Generative AI Survey 2026.

Where is the line between an AI research assistant and AI writing the paper?

The line is whether the AI helps the student understand and check sources, or produces the substance the student submits as their own. The University of Birmingham’s Code of Practice on Academic Integrity draws it in unusually concrete terms, and it is a useful model even for students elsewhere. Submitting "work and Assessments created by someone or something else, as if it was your own, is plagiarism", and the code says that includes "Artificial Intelligence (AI)-generated content".

Birmingham’s editing rules, in clause A.1.6 of the same code, show how fine the line is: generative AI "may be permitted only where it is limited to technical or presentational support", and a student "must not use generative AI or other editorial assistance to" change the substance.

Other universities set the line elsewhere, so a student should read their own module brief. Litrevu’s own site says the same thing in its FAQ: "Many allow AI as a study aid for reading, searching and summarising your sources, which is what Litrevu is for", while "Many do not allow AI-written text in work you submit, even after you edit it."

Sources: University of Birmingham generative AI guidance and its Code of Practice on Academic Integrity (2026-27), clauses 1.5 and A.1.6; Litrevu FAQ.

How does Litrevu try to keep the student as the author?

Litrevu is an AI research assistant that reads every paper a researcher adds, answers questions about them with a citation to the exact passage behind each claim, compares studies side by side, and checks the researcher’s own writing against their sources. The design choices that matter for integrity sit underneath that sentence. Litrevu works only from the student’s own library, and its site states that "Litrevu cites only papers in your library, and checks each citation against a passage it read before you see it." Anything Litrevu adds from general knowledge "carries no citation", so an untraceable claim looks untraceable rather than borrowing authority it has not earned.

Litrevu can draft cited text, and we will not pretend otherwise in a post about honesty. Under a code like Birmingham’s, a drafted section is AI-generated content however much it is edited, and using AI to develop an argument is ruled out, so a draft is no shortcut there. What we tell students is in our FAQ: "Litrevu can draft text, so if your policy rules that out, use it to read, compare and check, write your own text, and declare how you used it." We also ask them to treat every answer "as a lead to check rather than a fact". Litrevu also checks a student’s own paragraphs against their sources, which is the use we would point a supervisor to first, because it keeps the student writing and gives them a way to test what they wrote.

Here is the shape of that check, as a format illustration and not a real result. The tokens stand in for the student’s own paragraph and uploaded papers:

Your sentence 2: cites (Author, Year)
  Passage found: uploaded PDF 3, page n
  Read the passage. Does it support your sentence as written?

Your sentence 3: no citation
  Not traced to any paper in your library.
  Add a source you have read, or mark it as your own argument.

The limit is stated plainly on the site too: Litrevu checks that a citation points to a passage, and it cannot confirm that the passage fully supports the sentence. The student reads the passage. That step is the learning, and no tool should remove it.

How does an AI research assistant help students learn instead of outsource?

An AI research assistant helps students learn when it shortens the mechanical work around reading, finding, comparing and checking, and leaves the judgement and the writing with the student. The Russell Group’s first principle is that universities "will support students and staff to become AI-literate", and UNESCO’s 2023 guidance asks for "a human-agent" approach. Both describe a student who uses tools knowingly, not one who is kept away from them.

In practice that means asking questions across a stack of papers and getting answers that point to a passage, then reading the passage. It means a side-by-side comparison of methods across studies, the raw material for synthesis, with the synthesis still written by the student. And it means finding out before submission that a sentence has no source behind it.

Voice is where the difference shows. A student who has read the papers, compared them and tested each claim against a source has something to say, and the prose that comes out of that is theirs. A student who pastes in a generated paragraph has nothing to defend in a viva. If you are teaching students to put sources into their own words, how to paraphrase in academic writing covers the skill no tool replaces.

Sources: Russell Group principles, UNESCO guidance.

How should students declare AI use in a thesis or assignment?

Students should declare which tool they used, its version where known, what they used it for, and whether any AI-produced text, images, code or tables appear in the submitted work, in the format their university asks for. ICMJE’s expectation that disclosure sits "in the submitted work" and COPE’s preference for the methods section both point to a declaration that a reader can find, not a line buried in an acknowledgement.

Litrevu’s free AI disclosure statement generator produces a declaration in a chosen university’s format (South African, UK, Australian and Norwegian universities), an AI reference in APA, Harvard or MLA, and a list of records to keep. It needs no account, "uses templates, not an AI model", and says "Nothing you type leaves your browser, and nothing is saved." It also says it "cannot tell you whether a use was allowed", which is right: the brief decides that. For the wording itself, how to write an AI-use disclosure statement walks through examples.

What should a university ask before approving an AI research assistant?

A university should ask where student data goes, whether every citation can be traced to a real source, what the tool will not do, and how it supports disclosure. Those four questions separate a research assistant from a text generator faster than a feature list does. Here are the answers Litrevu publishes on its own site, as of October 2026.

Other tools answer different questions well, and we would rather a university chose the right one. Elicit is stronger at finding and screening papers across a large corpus, and Scite at showing whether citing papers support or contrast a claim. Litrevu is strongest once a student has chosen their papers and needs to read, compare and check against them. Whichever tool it is, one that cannot say where data goes, how citations are traced, what it will not do and how it supports disclosure is not ready for a postgraduate cohort. Why detection is giving way to process-based evaluation covers the assessment side of the same decision.

Frequently asked questions