AI-assisted writing is text a person drafted and shaped, with AI helping on some part of the work: fixing grammar, suggesting an outline, rewording a sentence. AI-generated writing is text a model produced from a prompt, with the person doing little more than asking. There is no sharp line between them. It is a spectrum, and the terms are used differently by detectors, schools and publishers.
Key takeaways
- The two terms describe how much of the wording and thinking came from the model. They do not describe quality or honesty.
- A detector sees only the final text. It cannot see a brainstorming chat, an outline or a prompt, and it cannot tell you what a writer actually did.
- Policies decide what is allowed. Many schools and publishers treat grammar polishing, idea generation and sentence generation differently, so read the rule for your own context.
- Detector labels for “assisted” text are the least settled part of the field. Pangram, whose API powers this site, says itself that co-authorship is a spectrum.
What is the difference between AI-assisted and AI-generated writing?
The difference is who did the composing. In AI-generated writing, the model composes: someone supplies a prompt and gets back a draft, a paragraph or a whole document. In AI-assisted writing, the person composes and the AI helps with a step along the way, or revises what the person wrote.
Most explainers, including those from content-marketing vendors, define the terms in one line each and stop. That is fine as a start but it leaves out the middle, which is where almost every real case sits. A student who writes an essay, asks a chatbot to tighten two paragraphs and then edits the result has done something that neither word describes well.
A more useful way to think about it is two separate questions. Where in the process was AI used? (before writing, during writing, or after writing). How much of the final wording is the model’s? (none, some, most). The first is about process and cannot be seen in the text. The second is about the text and is what a detector can look at.
Where does each kind of AI use sit on the spectrum?
Here are the common uses, in rough order of how much of the final wording comes from the model. The order is a guide, not a rule: asking for “ten counterarguments” can shape an essay more than a grammar pass does.
| Use | What it looks like | Whose wording is in the final text |
|---|---|---|
| Brainstorming | Asking a chatbot for topic ideas, angles or counterarguments, then writing alone | The person’s |
| Outlining | Asking for a structure, then writing every sentence yourself | The person’s |
| Grammar and spelling correction | A tool fixes errors in a finished draft | Almost entirely the person’s |
| Autocomplete | Suggested next words or sentences accepted while typing | Mixed, depends on how often accepted |
| Rewriting or paraphrasing | A tool restates your paragraph in a smoother, different tone or style | Largely the model’s wording |
| Human editing of an AI draft | The model drafts, the person revises, cuts, adds and checks | Mixed, depends on how much changed |
| Full generation | A prompt in, a finished text out, used as it is or nearly so | The model’s |
Two of these deserve a closer look because they are often misread.
Brainstorming and outlining leave no mark on the text. If you ask a model for angles and then write the piece yourself, the words on the page are yours. A detector looking at that text has nothing to find, and a clean result says nothing about whether you used AI earlier in the process. Whether that earlier use was allowed is a policy question.
Editing an AI draft is not the same as writing. Changing a few words in a generated draft leaves most of the model’s phrasing, structure and ideas in place. Rewriting it thoroughly, adding your own examples and checking each claim moves it toward being your work. Where it lands depends on how much changed, and nobody, including a detector, can measure that exactly.
How do AI detectors treat assisted writing?
Detectors differ, and many offer only a single “AI” score. Some now report an in-between category. Pangram, the detector behind this site, is one of them, and its documentation is specific enough to quote.
According to Pangram’s December 2025 announcement of its AI assistance detection, light assistance “typically indicates surface-level changes that do not affect the underlying ideas, structure, or content of the text,” with examples such as spelling and grammar fixes, updated phrasing, translation and readability changes. Moderate assistance “typically indicates changes where AI may have rewritten significant portions of the text or added content of its own,” with examples such as adding details, adjusting tone, restructuring or rewriting in a different style. Its model card for Pangram 3.2 describes the document-level “AI-Assisted” outcome as one returned where text shows signs of AI assistance rather than direct generation, including documents that mix assisted and human-written content. The model card also says the detector is accurate to a resolution of roughly 50 words, so a short human passage inside an AI-edited text can be mislabeled.
Three points follow from those sources.
- The label describes how the wording looks, not what the person did. The model has no access to chat history or edits. It sees a passage and decides which category it resembles.
- The boundaries are soft. Pangram’s own announcement describes co-authorship as a spectrum and acknowledges that the line between light and moderate assistance is hard to fix. Treat any “assisted” label as an approximate region, not a measurement.
- Not every edit shows up. Wiki Education, in its FAQ on Pangram reports, says that simple grammar fixes in Grammarly will not typically trigger Pangram, but use of writing suggestions may. That is one editor’s guidance and not a guarantee, but it matches the pattern above: the more a tool changes the wording, the more likely a detector reacts.
On this site, Pangram’s AI-Assisted passages appear as a “Mixed signal” in the highlights and are included in the share of text flagged as AI-like. The verdicts read Mostly human-like, Mixed signals or Likely AI-assisted, and the checker never returns a reason for a label. Our methodology page lists what each part of the result does and does not show, and what a detection percentage means explains how to read the figure.
Other detectors handle this differently. Turnitin, for instance, says its model also looks for AI-generated text that was later run through a paraphraser or “bypasser” tool (English only), and that its detector is not tuned to flag ordinary spelling, grammar and punctuation changes from grammar checkers. In its tests on fully human-written documents, those changes were mostly not flagged. It adds that text produced by a grammar tool’s generative features, such as drafting or paraphrasing, will likely be flagged (Turnitin FAQ). That is the vendor describing its own tests, not independent evidence.
How do schools and publishers treat AI assistance?
They disagree, and many leave the call to the individual instructor or journal. The examples below are real, current policies. They show the range, not a consensus.
A traffic-light model at a US university. Georgia College & State University offers instructors three levels for assignments. Red means “AI use is not permitted at all for this assignment.” Yellow means some use is acceptable and some is not, with a statement spelling out which. Its sample yellow assignment allows AI “only during the brainstorming and outlining stage” and forbids “using AI to write any sentences that appear in your draft.” Green allows AI use without significant restriction. The same assignment therefore treats brainstorming as allowed and sentence generation as not, which is exactly the split a detector cannot see. (GCSU, AI policies for student assignments)
Default prohibition at a Canadian university. The University of Toronto Mississauga tells students they are “not allowed to use generative AI in a course unless the instructor explicitly permits it.” It also warns that paraphrasing apps and Grammarly’s AI writing partner are derived from generative AI, so some courses may not permit them, and advises checking with the instructor first. (UTM, Generative AI and academic integrity)
Acknowledgement in UK qualifications. The Joint Council for Qualifications’ guidance on AI in assessments says AI use must be referenced when it appears in submitted work, naming the tool and the date the content was generated, and keeping a copy of the content. It lists copying or paraphrasing AI-generated sections so the work is no longer the student’s own, and failing to acknowledge AI use, as AI misuse. (JCQ, AI use in assessments)
Disclosure rules in publishing. Elsevier lets authors use generative AI and AI-assisted tools in the writing process “only to improve the language and readability,” requires a declaration statement, and says basic checks of grammar, spelling and punctuation need no declaration. It also says AI tools must not be listed as an author. (Elsevier, generative AI policy for writing) The International Committee of Medical Journal Editors takes a similar line on authorship: chatbots and other AI-assisted tools should not be listed as authors because they cannot be responsible for accuracy, integrity and originality, and authors who use them should describe how in the cover letter and the work. (ICMJE, AI use by authors)
Notice that “AI-assisted” means something different in each. For Elsevier it covers language polishing. For Georgia College’s yellow level it can include brainstorming but not sentence generation. For UTM it may mean nothing is allowed. A text that a detector marks “AI-assisted” could be fully compliant in one setting and a violation in the next, and the detector has no way to know which.
What can a detector say about human, AI-assisted and AI-generated text?
This table is the short version. “What it usually means” describes how people tend to use the label; it is not a guarantee about any single text.
| Category | What it usually means | What a detector can say | What a detector cannot say |
|---|---|---|---|
| Human | A person composed the text, possibly with spell-check or light proofreading | Passages resemble patterns of human writing | That no AI was involved at any stage, or that a person wrote every word |
| AI-assisted | A person’s writing was changed or extended by AI, or human and AI passages are mixed | Some passages look edited or co-written by a model, with a confidence level | What tool, how much, at which step, or whether the use was allowed |
| AI-generated | A model produced the wording from a prompt | Passages closely resemble model output, with a confidence level | Who submitted it, whether it was disclosed, or whether a person then rewrote it heavily |
Two cautions apply to the whole table. First, a result is a label on passages, not a finding about a person. Second, every row can be wrong. Human writing can be labeled AI-like, and generated or heavily edited text can pass. Can AI detectors be wrong covers who is most affected, and how accurate are AI detectors explains why we publish no single accuracy figure.
How should you handle assisted writing in practice?
The honest workflow is about disclosure and process, not about the label.
If you are the writer: find the rule for your situation before you start, not after. If it is unclear, ask the instructor, editor or client. Keep drafts and version history, which show your process better than any detector can. If you used AI in a way the rule allows, say so in the form the rule asks for.
If you are the reader or reviewer: treat an “assisted” label as a prompt for a question, not an answer. Ask what the writer did and whether their account fits the flagged passages. Compare with their other work and look at drafts and notes. Do not penalize anyone on the label alone. The checker’s own methodology page states that a result is not proof of how a text was written and not grounds for a penalty by itself.
If you are an instructor or editor writing the rule: name the activities, not the tools. “AI may be used for brainstorming and outlining, not for writing sentences” is clearer than “AI-assisted work is allowed,” because the second phrase means something different to every reader. The traffic-light approach above is one workable way to do this.
Frequently asked questions
Is using Grammarly or a spellchecker AI-assisted writing? It depends on the feature and the rule. Basic spelling and grammar correction is usually treated as low-level help, and Elsevier, for example, says it needs no declaration. Rewrite and generative suggestions go further, and some institutions, such as UTM, say instructors may not permit them. Check the rule that applies to you.
Can a detector tell if I used AI to brainstorm? No. Brainstorming leaves no trace in text you wrote yourself. A detector sees the final words only, so it can neither confirm nor rule out earlier AI use.
If a detector labels my writing AI-assisted, have I broken a rule? Not necessarily. The label says some passages resemble AI-edited text. It does not say what you did or whether it was permitted. If you did not use AI, say so and show your drafts. Detectors can mislabel human writing.
Is AI-assisted writing better than AI-generated? Neither is a quality ranking. Assistance keeps the writer in charge of ideas, accuracy and voice, which is why many policies prefer it. Generated text can still be sound if it is checked, but the writer remains responsible for every claim in it.
Sources and check dates
All checked 5 October 2026 unless stated. Vendor statements are the vendor’s own.
- Pangram: announcement of AI assistance detection (11 December 2025) and Pangram 3.2 model card
- Wiki Education: FAQ on suspected AI edits and Pangram reports
- Georgia College & State University, UTM, JCQ, Elsevier and ICMJE pages linked above
If you already have the text, run it through our AI-generated text detector and review the flagged passages rather than relying on the overall score.