Declare it where a reader’s judgement of the work depends on a decision you did not make. Spellcheck moves no judgement, so nobody names it; a generated paragraph under your byline moves several. Between those, ask whether the reader would assess the work differently knowing, and whether you can still vouch for every claim in it.
Flow diagram of 5 steps: Academic and scientific work, Journalism and anything with a masthead, Competitions, grants and marketplaces, Commissioned and client work, Synthetic media of real people or places.
Nobody has ever written "spellcheck was used in the preparation of this article", and no reader has ever wanted them to. That is worth sitting with, because it rules out the intuition most disclosure arguments start from — that software touching your text is the thing being declared. Software has been touching text for forty years. Something else changed.
What changed is where the judgements live. A spellchecker flags a string it does not recognise and you decide; it makes no claim, holds no opinion, and cannot be wrong about anything you were relying on. A model that drafts your third paragraph has made a series of decisions — what to assert, how sure to sound, which detail to include — and those decisions arrive already made, in your voice, under your name. The reader assumes a person weighed them. Disclosure is how you correct that assumption when it is false.
So the workable test is not "did I use a tool" but "would a reader assess this differently if they knew, and can I still vouch for every claim in it". That is the same test that governs disclosing a conflict of interest, a paid placement or a composite character, and it has the useful property of returning sensible answers at both ends. It says nothing needs declaring when a model told you your opening was unclear and you rewrote it yourself. It says a great deal needs declaring when a model produced the statistic in your second paragraph and you did not check it.
Blanket notices — "this content was created with AI assistance", appended to everything an organisation publishes — read as diligence and function as camouflage. They cover the piece where a model wrote the conclusion and the piece where somebody ran a grammar check, which means they tell a reader nothing about either, and after a few dozen sightings the label stops being read at all. A disclosure that cannot distinguish between the cases it covers is not a disclosure. It is a legal reflex with a transparency costume on.
The practical half of this has nothing to do with ethics. Putting your name on a piece is a warranty: that the facts were checked, the quotes are real, the argument is one you hold. Generated prose enters a draft without that warranty attached, and the sentences that are wrong look exactly like the ones that are right, because fluency is uniform. Either you do the work to make the warranty true — check every claim, rewrite what you cannot verify — or you say plainly which parts it does not cover.
Laid out as a spectrum, most of the argument dissolves. The uses at the bottom move no judgement and carry no claim; the ones at the top move both. The interesting cases are in the middle, and what separates them is consistently whether the output arrived as a description you acted on or as text you accepted.
Common uses and whether they need declaring
| What the model did | Declare it? | Why |
|---|---|---|
| Spelling and autocorrect | No | It flags, you decide. No claim is made and no judgement of yours is replaced |
| Grammar suggestions accepted one at a time | No | Each one passed through your approval, which is what keeps the sentence yours |
| Editorial notes on a draft you then rewrite | No | The notes created work rather than removing it. Every sentence is still one you wrote |
| Transcribing an interview you recorded | No, but check it against the audio | The words are the speaker’s. Misheard ones become your misquote, not the tool’s |
| Translating your own text | Usually yes where nuance carries weight | Meaning shifts in translation, and a reader may reasonably want to know the original was in another language |
| A drafted section you edited afterwards | Yes wherever the byline is a warranty | You are vouching for decisions taken by a process you cannot inspect |
| Research summaries you did not verify | Yes — better still, check them and cite the source | The risk is a confident false claim published in your name, which no disclosure fully repairs |
| Synthetic images, voice or video of real people or places | Yes, always, and often as a matter of law | Readers treat photographs and recordings as records of something that happened |
| Wholly generated text under your byline | Yes | Without it, the byline is itself the false claim |
Prose has always been constructed, so a reader brings no assumption that any particular sentence came from a specific moment. A photograph carries exactly that assumption — that a camera was pointed at something real — and a synthetic image trades on it whether or not you intended to. This is why the same person can reasonably decline to declare a lightly assisted paragraph and must declare a generated street scene of a town they are writing about. The undisclosed synthetic image is not a smaller version of the undisclosed synthetic paragraph. It is a different kind of claim.
Everything above is a principle, and a principle is what you use when nothing binds you. Several contexts do bind you, and in those the sensible move is to read the actual rule rather than to reason from first principles about what seems fair.
Institutions and journals publish explicit policies, and they vary enough that only your own is worth reading. The one point of near-consensus across scientific publishing bodies is that a model cannot be listed as an author, on the reasoning that authorship carries accountability a tool cannot hold — which usually means use gets described in the methods or acknowledgements instead. Undeclared use is generally treated as a misconduct question rather than a style one.
Newsrooms that have published policies tend to require a named human who signed off and a note when a model produced any published text. The asymmetry is worth understanding before you decide to skip it: the byline is yours, and the cost of a fabricated detail lands on the publication, which is why the rule exists and why it is enforced more firmly than it is explained.
Entry terms frequently prohibit generated work outright or require it to be declared at submission. This is contract rather than ethics, and it does not soften with good intentions — an award can be withdrawn after the fact, and the terms you agreed to are the ones that decide, not what you considered reasonable at the time.
A contract that assigns copyright usually asks you to warrant that the work is original and yours to assign. Purely machine-generated output may not attract copyright protection at all in several major jurisdictions, since human authorship is generally required, and the position is still moving. If you cannot make that warranty truthfully, that is a conversation to have before signing rather than after delivery.
Labelling duties for AI-generated and manipulated media are being written into law in a number of jurisdictions and are changing quickly. Check what currently applies where you publish rather than relying on any summary of it, including this one — the specific obligations, deadlines and exemptions are exactly the details that have moved most since anybody last wrote them down.
Most disclosures fail not because they are absent but because they are empty. "Produced with AI assistance" leaves every question a reader might have unanswered: did it write the argument, or fix the commas? Are the quotes real? Did anybody check the figures? A note that prompts all three questions and answers none has spent the reader’s attention without buying them anything.
A useful disclosure is one sentence, uses verbs, and is placed where the person deciding whether to trust the work will actually meet it. Compare "this article was created using AI tools" with "a model drafted the summary at the end; the interviews, quotes and figures are my own and I checked each figure against the source". The second is barely longer and settles the question. It also does something quietly valuable for you: writing it forces you to notice which claims you have not actually verified, which is usually the moment the disclosure stops being the main problem.
Write "drafted the closing section" or "suggested a reordering of the middle" instead of "assisted with". Assistance is not a thing a reader can picture, and vagueness in a disclosure reads, correctly, as an attempt to satisfy a norm without submitting to it.
In reported work the reader’s actual worry is the quotes and the numbers. One clause confirming those are yours and checked does more than three sentences about your general approach to technology.
There is a real difference between reading generated text and finding it plausible, and verifying each claim against a source. Both are defensible; only one of them is checking, and a reader is entitled to know which one happened.
A note at the foot of a long page is not reaching the person who read the headline claim and stopped. If a specific passage is the assisted one, the honest place for the note is next to that passage.
One line is almost always enough, and for tasks that carry no claim the right length is zero. Declaring a grammar check alongside a drafted conclusion flattens the difference between them, which is the same failure as declaring neither.
Concretely, and so this page is not mistaken for a policy: there is no AI-disclosure field on the submission form. It asks for a category and format, a title, a description, then shows you a review step. Nothing here obliges you to declare anything about how you wrote your idea, and the guidance above is general rather than a house rule.
What the model on this site does is read your submission and return written notes and scores across clarity, originality, social impact, engagement potential and feasibility. It does not write the submission, does not offer you a rewritten version to accept, and blocks nothing — community votes decide what gets scheduled. That design is also why the disclosure question mostly does not arise at this stage: notes are not text you have to account for.
We do not run a detector on submissions, and would not act on one if we did. Automated detection flags careful, formal and non-native English as machine-written and misses anything lightly rewritten, so acting on it would penalise particular writers for how they write rather than for what they did. Any platform claiming it can reliably tell is overstating what is currently possible.
Where this genuinely matters here is later than submission, and it is worth saying plainly. If an idea is ever produced, the material inside it has to hold up — a claim about a place must be true, a person who appears must have agreed to. A detail that entered a proposal because it sounded right, and stayed because nobody checked it, becomes a problem at exactly the point it is expensive to fix, and whether it was declared makes no difference to that. Kind Channel is new, the community is small and nothing has aired yet, so that is a statement about where the risk sits rather than a lesson drawn from cases we have had.
In almost every context, no. Feedback describes a problem and leaves the repair to you, so the sentences in the finished piece are ones you wrote and can account for, and a reader learns nothing useful from knowing what prompted a revision. Nobody declares the friend who said the opening was confusing. The exception is a formal setting — some academic and competition rules are written broadly enough to cover any use at all, in which case the rule decides rather than the principle.
No, and treating any single rule as universal is how people get caught out. There is a principle that travels well — declare it where a reader would judge the work differently knowing, and where you cannot personally vouch for a claim in it. But academic policies, newsroom standards, competition terms, client contracts and laws on synthetic media each impose their own requirements, they differ from one another, and several are changing. Where something binds you, read that rather than reasoning from what seems fair.
Name the task in a verb and say what was not touched. "A model drafted the closing section; the quotes and figures are mine and I checked each against the source" settles the reader’s real questions in one sentence. "Created with AI assistance" raises three questions and answers none, which is why blanket labels stop being read after a few sightings. Put the note where somebody deciding whether to trust the work will meet it, not at the foot of a long page.
It can, and this is the part most likely to matter commercially. Copyright in several major jurisdictions generally requires human authorship, so purely machine-generated output may attract no protection, and the position is still being worked out in courts and legislatures. That becomes concrete when a contract asks you to warrant that the work is original and yours to assign. If a meaningful portion was generated, raise it before signing rather than after delivery, and check the current position where you are publishing.