SOP02 | Preserve what is in front of you first

English translation produced by the platform. It has not been reviewed by a native speaker and is not a certified translation. The Chinese version is the source.

Version 0.1 | 2026-09-24 | draft walkthrough. Goal: when you see something that looks like AI use, preserve the screen, the address, the time and how you got there — first.

Read this first

Preserve the facts in front of you before you organise files. Capturing what is on screen applies to model output, and equally to training clues you can lawfully see, such as a dataset listing or a company's own public statement.

A screenshot is not automatically worthless, and it does not on its own prove everything. Keep screen recordings and screenshots to material you are entitled to view. Do not get around a login or an access control, and do not record unrelated people or private audio.

Time, cost and tools

Plan 3–5 minutes for the urgent capture and about 10 minutes for the full note. These are estimates, not measured figures.

• Your system file manager, a notes app or the writing software you already use: nothing new to pay for; save to the device itself where possible; if cloud sync is already switched on, files may end up in your cloud account.
• The built-in screenshot function: included, nothing to pay for; screenshots may sync automatically with your system photo library or screenshots folder.

Before you start

• Preserve anything in front of you that might disappear before doing anything else. Do not sign out or refresh first.
• Do not crop original files or screenshots, do not repair them with AI, and do not fabricate metadata.
• Record who actually did the work if someone helps you. Never give a helper your password or a verification code.

Step by step

Step 1 | Stop doing anything that could change what is on screen

Where to click: Do not refresh, do not delete the conversation, do not edit the original prompt, do not regenerate, and do not file a takedown yet. Leave what is on screen where it is.

What you should see: The result you are looking at has not been overwritten.

If you can't: If it has already gone, write down when you last saw it and what you did. Do not substitute a regenerated result for the first one.

Step 2 | On a computer, take a screenshot that includes the product's own interface

Where to click: On Mac press Shift+Command+3. On Windows press Windows+Shift+S, choose full screen or window, then open the notification or the screenshot tool and use Save As to save a PNG. Do not turn on AI auto-crop or enhancement.

What you should see: An image file exists, showing the product interface and the content concerned.

If you can't: If you cannot find where it saved, check the desktop, your Pictures or Screenshots folder, or the tool's own save settings. If you are permitted and not otherwise restricted, you can photograph the screen with another device — say that it is a photograph of a screen.

Step 3 | On a phone, take a screenshot first

Where to click: On an iPhone with Face ID, briefly press the side button and volume up together. On an iPhone with a Home button, use Home and the side or top button. On most Android phones, briefly press power and volume down. Release — do not hold, or you will get the power-off screen.

What you should see: The current screen appears in your photo library.

If you can't: Models differ, so check your device's own screenshot menu. Where an app blocks screenshots, do not get around the restriction. Record the address, the text and the reason it is restricted instead.

Step 4 | If the content is long, capture it screen by screen, in order

Where to click: Start from the beginning of the relevant conversation and work down, leaving a little of the previous screen visible each time so the sequence joins up. Keep them in order. Do not crop down to the one similar sentence.

What you should see: Consecutive images join up, and the relationship between input and output is visible.

If you can't: If long-screenshot capture is unsupported, several ordinary screenshots are fine. If a section cannot be captured, record it as missing rather than stitching something together.

Step 5 | Copy the web address, or record the session identifier

Where to click: In a browser, click the address bar, copy, and paste it into the "Incident record" template at the foot of this page. On a phone, press and hold the address bar to copy. If the app has no address bar, record the app name and the page or session title.

What you should see: Either the address exactly as it was, or an explicit note that there is no visible address inside the app.

If you can't: Do not press "share publicly" just to obtain a link. A private link you already have may contain a token — keep it in your private notes only, and do not publish it.

Step 6 | Record the company, product and model you could see at the time

Where to click: Write the names visible on screen into the incident record. Photograph the visible model label if you can. If no version is shown, write unknown.

What you should see: The brand and the model information are each recorded separately.

If you can't: With an aggregator app you may not know the underlying company. Record the aggregator, and do not guess at OpenAI or any other supplier.

Step 7 | Write down the time and the time zone

Where to click: Record the date, the time and the time zone as your device showed them. If you cannot fill in the time zone, write the city or region you were in and "as shown by device", and leave the conversion for later. If you are writing this from memory now, say that it is from memory.

What you should see: The time you observed it, the time you are recording it, and where that information came from are kept apart.

If you can't: Only remembering the month is fine. Do not back-fill a precise moment you do not have. If your computer's clock has not been checked, say so.

Step 8 | Write down how you got it and who was operating

Where to click: Record the type of device, the app or browser name, whether it was you or a helper, how many screenshots you took, whether anything malfunctioned, and whether you refreshed or regenerated. Do not record passwords, verification codes or cookies.

What you should see: Someone else could follow your note and understand how the material came to exist.

If you can't: If you do not know how to find the system version, write "don't know". Ordinary preservation does not require serial numbers or other sensitive details.

Step 9 | Record training clues separately

Where to click: If you are looking at a dataset page, save the visible entry ID, the dataset name and version, and the official source. Fill in the "Training clue" record, keeping "this entry contains the work" apart from "this model used that version of the dataset".

What you should see: The source, the version, the connecting evidence and the gaps are recorded separately.

If you can't: Not finding something does not prove it was never trained on. Finding only a title or a link does not prove a model actually trained on it either. Do not download leaked material and do not get around a paywall or a permission control.

Step 10 | Save a copy, then fill in the whole session

Where to click: Put the screenshots and the note into a private incident folder, and follow SOP03 to save the surrounding context. If you try again later, start a new incident (I002) rather than overwriting the first one (I001).

What you should see: The first observation can still be reconstructed on its own.

If you can't: If all you manage is the screenshots, save those and take part anyway. The rest can follow.

Check when done

• A retried result has not been passed off as the first observation.
• The address, product, time and capture method are recorded at least as far as you know them.
• You have recorded whether you supplied the original work yourself — see SOP03.

What you end up keeping

• The original sequence of screenshots, or another lawful record of what was on screen
• The incident record
• The training clue record, where it applies

If something is missing

• You cannot take screenshots: keep the selectable text, a description of the page, and the reason you were blocked.
• You only have someone else's account of it: mark it "not observed by me", keep the original link and the source of the account, and do not fabricate a screenshot.

Where your data goes, and sharing

Files on your device are not uploaded by this process. Your photo library, iCloud, Google Drive, OneDrive or your editor may already be syncing, so check where something is saved before you save it. An account export communicates with the original service. A full account export is not handed to this project by default.

Keep originals somewhere private. Make a separate copy before sending anything to anyone outside, and check it first for unrelated personal material and for contract restrictions. Share only when the recipient, the purpose and the scope are clear. Nothing is sent automatically in this phase.

When to get professional help

If evidence is about to disappear, if you have received court papers or a deadline, if the chain of rights is unclear, if trade secrets or a minor is involved, or if a formal certificate is required, have a qualified professional in the relevant place check it promptly rather than waiting for this research to finish.

One particular warning: do not start by trying to "prove the model memorised my work"

Preserve what already exists first. Any test you run afterwards can be affected by the text you have just typed in, by retrieval, by attachments, or by earlier turns in the same conversation, and has to be recorded separately. Taking part here never requires you to coax a result out of a model repeatedly.

Basis and testing status

Source identifiers: S03, S05, S07, S08, S20, S21, S22, S23. Official URLs and locators are listed on the law pages. The actions in this walkthrough are preservation suggestions, not a claim that the law requires every person to complete every step.

Document checks have been done. Real-device testing on Mac, Windows, iPhone and Android, testing with five non-technical users, and lawyer sign-off have not.

Incident record

Private. Keep what you observed apart from what you infer.

• Incident number
• Related work number
• Seen first-hand, or reported by someone else
• Company or product you saw
• Model or version visible (write unknown if you don't know)
• Web address or session ID (may contain an access token — keep private)
• Time you saw it, time zone, and how precise
• Time came from: the device / the service display / memory afterwards
• Time you are writing this note
• Which file holds the full text you typed in
• Had you supplied the original work or a similar fragment before: yes / no / unknown
• How you supplied it, which file or link, and where
• Visible retrieval, citations, attachments, custom settings and so on
• Which file holds the output, and whether it is a download, copied text or a screenshot
• Screenshot numbering and their order
• Device and app; who actually operated it; known faults or omissions
• Any retry or regeneration, and the separate incident number for it
• Facts I saw with my own eyes
• What I currently suspect
• What I cannot confirm, or is missing

Training clue

This is not a "training confirmed" label.

• Work or incident number
• Source: the company's own documentation / a lawfully accessible dataset / a public court filing / third-party reporting / the model's own claim / other
• Source URL, title, publisher, publication date, date you obtained it
• Dataset name, version, entry ID or locator
• What the entry actually contains: file content / a link / a title / a summary / don't know
• Why you think it corresponds to your specific version
• Do you know which model and version it went into: yes / no / unknown
• Independent material supporting the model-to-dataset connection
• Could this be only retrieval, prompt input or third-party republication
• How you obtained it