03 · What You Need to Know
The Difference Depends on What AI Contributed
AI-Assisted and AI-Generated Answer Different Questions
The easiest way to understand the distinction is to notice that the two terms describe different things.
AI-assisted
Describes how the research was conducted: AI supported a human-led task or workflow.
AI-generated
Describes the origin of an output: substantive material was generated by an AI system.
This means they are not mutually exclusive categories.
A researcher might conduct AI-assisted research in which an AI system generates code. The workflow is AI-assisted; the initial code is AI-generated. If the researcher inspects, tests, corrects, and uses that code in an analysis, human evaluation has entered the workflow, but the original source of the generated code has not magically changed.
Simple Editing Is Different From Generating Substantive Content
Consider two interactions.
In the first, you write: “The results indicates that student engagement were associated with persistence.” You ask an AI system to correct the grammar. It changes “indicates” to “indicate” and “were” to “was.”
In the second, you provide your results table and ask: “Write a 500-word discussion explaining these findings in relation to self-determination theory.” The system generates the argument, connections, transitions, and much of the language.
Both are AI-assisted activities in the broad sense, but the second clearly contains a much more substantial AI-generated contribution.
The distinction matters because generating new scholarly content introduces different questions about accuracy, source grounding, authorship, disclosure, originality, and intellectual contribution than correcting surface-level language in researcher-authored text.
Generation Is Not Limited to Prose
Researchers sometimes hear “AI-generated” and think only of manuscript text. Generative systems can produce many kinds of research-relevant outputs.
| Output |
Possible AI-assisted use |
When generation becomes substantive |
| Text |
Correcting grammar in researcher-written prose |
Producing new paragraphs, arguments, interpretations, summaries, or explanations |
| Code |
Explaining an error in researcher-written code |
Generating functions, scripts, models, or analytical pipelines |
| Images |
Helping modify presentation formatting |
Creating synthetic figures, illustrations, or other visual content |
| Search material |
Suggesting alternative keywords |
Generating purported citations, literature summaries, or descriptions of studies |
| Qualitative analysis |
Formatting researcher-developed codes |
Generating candidate codes, classifications, themes, or interpretations |
| Research design |
Critiquing a researcher-developed procedure |
Generating instrument items, hypotheses, procedures, or methodological proposals |
Not every generated output has the same importance. A generated manuscript title and a generated interpretation of primary data may both be AI-generated, but the latter has much greater potential to affect the intellectual substance of the study.
Generation Can Be Substantial Even When the Researcher Supplied the Facts
A common assumption is that content is not really AI-generated if the researcher supplied all the source material.
That is too simple.
Suppose you provide five verified findings and ask an AI system to turn them into a discussion section. The factual ingredients came from you, but the system may still generate the argumentative structure, causal language, theoretical connections, emphasis, qualifications, and prose.
Some of those generated connections may go beyond the evidence you supplied. Others may subtly change the strength of a claim.
Source-grounded generation can reduce some problems, particularly fabrication arising from missing context, but it does not eliminate the need to evaluate what the system generated.
Editing AI-Generated Content Does Not Necessarily Erase Its Origin
Imagine that an AI system generates a 1,000-word discussion section. You spend two hours correcting it, rewriting sentences, removing unsupported claims, adding citations, and reorganizing paragraphs.
At what exact moment does the text stop being AI-generated and become human-written?
There is no universally accepted percentage or word-count threshold that answers this question across research contexts. The process is better understood as a continuum of human and machine contribution.
Researcher-created, AI-polished
The researcher supplies the substantive ideas and text; AI makes limited linguistic or presentational changes.
Collaboratively transformed
The researcher supplies substantial content, AI restructures or expands it, and the researcher substantially evaluates and revises the result.
AI-generated, researcher-edited
AI produces the initial substantive output; the researcher subsequently checks, modifies, or curates it.
These labels are descriptive rather than universal policy categories. They are useful because they expose what a binary “AI or no AI” label hides.
The More Intellectual Work AI Performs, the More Consequential the Distinction Becomes
Changing punctuation is not equivalent to generating an interpretation. Formatting references is not equivalent to generating references. Suggesting synonyms is not equivalent to producing a literature synthesis.
The distinction becomes especially important when AI contributes to activities ordinarily associated with scholarly intellectual responsibility: framing a research problem, selecting or justifying methods, interpreting evidence, constructing arguments, determining conclusions, or representing the literature.
This does not automatically make such use prohibited. Policies differ. It does mean that researchers should pay closer attention to verification, transparency, authorship, and whether they are still making the intellectual contribution represented as theirs.
AI-Generated Does Not Automatically Mean Wrong or Unethical
The phrase “AI-generated” can acquire a moral meaning that it does not inherently possess.
An AI-generated piece of code may be correct. An AI-generated translation may accurately preserve the source. An AI-generated diagram may communicate a concept effectively. An AI-generated draft may become useful after rigorous verification and revision.
Whether the use is appropriate depends on the task, research context, applicable policies, validation, transparency, data handling, intellectual property, and the researcher's responsibility for the result.
Conversely, calling something merely “AI-assisted” does not make it responsible. A researcher could describe an entire AI-written literature review as “assisted” and thereby conceal the substantive role of the system.
Watch Out
Do not use “AI-assisted” as a euphemism for substantial AI generation. Accurate description matters more than choosing the label that sounds least consequential.
AI-Generated Content Is Not Automatically Plagiarism, but Other Integrity Questions Remain
Plagiarism traditionally concerns presenting another source's words, ideas, or work without appropriate attribution. Generative AI complicates this framework because machine-generated output is not simply equivalent to copying a passage from an identifiable human author.
That does not make undisclosed or inappropriate AI generation acceptable. Depending on the circumstances, it can raise separate questions about authorship, misrepresentation of contribution, attribution, intellectual property, compliance with policies, fabricated information, or other forms of academic or research-integrity concern.
Researchers should therefore avoid reducing every problematic use of generative AI to “plagiarism.” The more precise question is what scholarly norm or policy the use actually affects.
AI Generation Does Not Transfer Authorship or Responsibility to the AI
Even when AI generates substantial material, responsibility for scholarly outputs remains with human researchers.
UNESCO's guidance emphasizes human agency and accountability in the use of generative AI. The European Commission's guidance for researchers similarly emphasizes that researchers remain responsible for outputs produced with generative AI support and should critically assess generated material.
This creates an important asymmetry: AI can contribute content without being able to accept scholarly responsibility for that content.
If a generated paragraph contains a false claim, a fabricated citation, or a distorted interpretation, the researcher cannot resolve the problem by explaining that the sentence originated with the system.
The Boundary Should Be Described Functionally, Not by Word Count
Researchers sometimes look for a numerical threshold: 10% AI-generated text, 20%, 50%, perhaps some magical point at which a manuscript crosses categories.
There is no generally applicable scholarly rule of this kind.
A single AI-generated sentence could be highly consequential if it contains the study's central interpretation. Ten pages of AI-generated formatting instructions could be intellectually trivial. Counting words ignores what those words actually do.
A more useful description asks:
- What did the researcher create before AI became involved?
- What did the AI system generate or transform?
- How substantially did the generated output influence the final work?
- How was the output verified and revised?
- Would a reader need to know about that contribution to understand how the research was produced?
Disclosure Should Describe the Actual Contribution
Where disclosure is required, “AI was used in preparing this manuscript” may be technically true but uninformative.
Compare that with a description indicating that generative AI was used to produce candidate Python code for a specified analysis, that the code was subsequently inspected and tested by the authors, or that AI was used for language editing without generating scholarly claims.
The second form gives readers information that helps them understand the role of the system.
Exact disclosure requirements vary among publishers, journals, institutions, funders, and other authorities. Researchers should verify current rules rather than assume that one wording or threshold applies everywhere.
The Central Issue Is Human Intellectual Contribution
The distinction ultimately leads to a deeper question: what part of the research are the human researchers claiming as their scholarly contribution?
If AI helps researchers express an argument they developed and can defend, that is different from asking AI to create an argument they had not developed and then accepting it because it sounds plausible.
If AI generates code that researchers understand, test, correct, and validate, that differs from running generated code they cannot explain.
If AI proposes a possible interpretation that researchers independently investigate and ultimately justify from the evidence, that differs from allowing the system's interpretation to become the conclusion by default.
The question is not whether researchers touched the final output enough times. It is whether they performed the intellectual and evaluative work necessary to take responsibility for it.
