Writing: what you can keep

For writing

Books, articles, journalism, fiction, poetry, scripts, newsletters and blog posts are all "writing" here. That is a content label so guidance reaches you — it does not create a separate group or a separate count.

• Manuscript files with their version history, and the dates of each version.
• Publication records: ISBN or DOI where one exists, the publisher's page, the issue or edition, the byline as printed, and any licence terms you agreed to.
• If a publisher or platform holds the rights, note who holds what. Your contract matters and it is between you and them; keep your copy, do not send it here.
• If your text appears in a public dataset index, save the listing page and the date, and note that a listing is a listing — not proof that a given product used it.

Before anything else

Keep your own material where you control it. This site does not accept uploads and will never ask for your files, contracts or cloud drive. Nothing below is legal advice.

1. Keep your originals. Working files, drafts, version history, export files, the camera or instrument files, the repository history. Do not overwrite an original with an edited copy — save the edit as a new file.
2. Write down where your work was published, when, and under what name or licence. Screenshots of your own published page, plus the URL and the date you took it.
3. Keep links you came across lawfully, with enough of the surrounding page to show the context, and the date you found it. Do not break a paywall, a login, or a site's terms to collect anything.
4. Keep the three kinds of material separate: your originals, screenshots or copies of what you saw, and your own explanation of why you think they are connected. Mixing them together makes all three weaker.
5. Do not rewrite your account to match anyone else's. Say plainly what you know, what you assume, and what you are unsure about.

If a model's output is part of what worries you

Save the whole exchange, not the striking part of it:

• the exact prompt you typed, in full;
• the complete answer, including anything that came before or after;
• the product and model version shown in the interface, and the date and time;
• whether web search, file upload or any other tool was switched on;
• whether you supplied the text or image yourself in that conversation. If you pasted your own work in first, record that honestly. A model repeating text you just gave it shows nothing about training.

What this kind of material does not settle

Being honest about the limits is more useful than overstating them:

• A title appearing in a public dataset listing does not by itself show that a particular product used it.
• A model saying it "was trained on" something is not evidence of that. Models state things that are not true about themselves.
• Similar style, tone or subject matter is not by itself infringement. Many things look alike without one having been copied.
• Screenshots without the surrounding context are easy to dispute.

When to talk to a lawyer

Formal preservation, whether you personally have a claim, and any time limit that might apply are all questions for a lawyer qualified in your own jurisdiction. Time limits vary by country and by the kind of claim, and this site does not publish a deadline it has not verified for you. Keeping your own records well does not depend on that answer, so it is worth doing now.