Sunda Cloak
When you attach a document to ChatGPT, the file is sent to the provider’s servers. A text extractor pulls the words out of the document’s internal structure, and the model reads that extracted text — not your formatting, not the page as you see it. The file may be retained and, depending on plan and settings, used to improve models; policies differ and change, so check the provider’s current data-controls page. For a consultant, though, the bigger practical risk is usually not training. It is the person on the other end getting a reusable version of your work in seconds — and that is the part you can do something about.
The pipeline: upload, extract, model
Every major chat tool handles an attached document in roughly the same three steps.
- Upload. The file leaves your computer and is stored, at least for the length of the conversation, on the provider’s infrastructure.
- Extract. The model does not open a .docx the way Word does. A separate text extractor — often a common open-source library — walks the document’s structure and produces plain text.
- Model. That extracted text is placed into the conversation alongside your prompt, and the model answers from it: a summary, a rewrite, a template.
The middle step is the one people forget. The extractor decides what the model can and cannot see.
What the model actually sees
- Text, not layout. Fonts, colours, column widths, headers and footers mostly vanish, and an embedded chart is skipped or reduced to alt text unless the tool runs a separate vision pass.
- Hidden text, often included. Word files have a hidden-text channel — formatting that tells Word not to display a run of text. Many extractors read it anyway, because they walk the structure rather than rendering the page. That asymmetry is what cloaking relies on.
- Upload and copy-paste are different paths. Pasting skips the extractor: the model gets what Word put on the clipboard, which leaves hidden text out. Shaping what an AI reads affects the upload path only.
Retention and training, in general terms
Retention and training defaults differ by provider, by plan, by region and by the settings on the account — and they change. There is no safe general answer; the only reliable source is the provider’s current data-controls page for the exact plan being used.
Retention windows, opt-out switches and training defaults differ by plan and region, and they change. We give no specific periods here because they go stale. Read the provider’s current data-controls page for the exact plan you are on.
Claude, Gemini and Copilot follow the same shape, though their extractors differ and change over time.
When it’s your deliverable being uploaded
Most advice online is about your data going into your AI tool. For a consultant, freelancer or agency the worry usually runs the other way: you send a proposal to a client, and someone on their side drops it into ChatGPT with “turn this into a template”. You do not control their plan or settings. You control two things: the file and the contract.
- The contract. A one-sentence clause about AI reuse of your work sets the norm. Starting language is in the AI-clause guide and Telling your client.
- Cloak the outgoing copy. Cloaking adds hidden text only the extractor reads — look-alike decoys beside each real number, name and link, plus decoy paragraphs some extractors prefer. Your client reads exactly what you wrote; a one-click upload gives a noisy read. It covers the upload path only, and its strength depends on the tool’s extractor: here is what it does and doesn’t do.
- Mark each copy. An invisible per-copy mark tells you which copy travelled. Sunda Cloak adds one automatically; a reference code in the footer is the low-tech version.
- Set a tripwire. An opt-in hidden reference link that records a time, network address and user-agent if an AI tool actually fetches it. Many tools never follow links in uploaded files, so a hit is a clear signal and silence is no information. More in Can you tell if someone used AI on your document?
Any contract language on this site is a starting point to review with a lawyer, never legal advice. Your jurisdiction, your engagement and your client’s own terms all change what is sensible.
When you upload client material
The reverse situation is the one that can end an engagement. If a client’s document is under NDA, uploading it to a consumer AI plan sends it to a third party and may count as disclosure. Practical habits:
- Check the NDA or engagement terms for anything about third-party processing before you upload.
- Prefer a business plan whose terms exclude training, turn on its opt-outs, and confirm on the current data-controls page that they are in effect.
- Where possible, paste the paragraph you need rather than the whole file, with names and figures stripped.
- Sunda Cloak’s optional upload protection is for the other direction — files you attach to approved sites (webmail, client portals) that you do not want machine-read. It does not make uploading client material to an AI tool safer: the real text still reaches the provider, and the AI’s answer to you gets worse.
What this guide cannot tell you
- We cannot tell you what any specific provider keeps, for how long, or trains on today. Only the provider’s current data-controls page can answer that.
- Nothing stops a person who has your file from uploading it. Cloaking changes what the extractor gets, not whether the upload happens.
- Copy-paste, screenshots, OCR and PDF export bypass cloaking entirely; extractors that strip hidden text recover the original; anyone who shows hidden text sees the decoys.
- The tripwire fires only when a tool actually follows the hidden link. Silence proves nothing.
- All of it is a deterrent against casual AI reuse — not a guarantee, and never a substitute for a contract. Contract language here is a starting point to review with a lawyer, not legal advice.
Frequently asked
It depends on the plan and the settings, and the defaults differ. Policies also change, so check the provider's current data-controls page for the plan you are on.
Not from the provider — we know of no AI service that offers a file's author a way to learn that someone else uploaded it, and provider policies differ by plan and change. What you can do is put a hidden reference link (a tripwire) inside each copy you send; if an AI tool actually fetches it you get a time, a network address and a user-agent. Many tools never follow links inside uploaded files, so a hit is a clear signal and silence is no information. See Can you tell if someone used AI on your document?
It may be. Uploading sends the document to a third party's servers, and whether that is permitted depends on your NDA, your engagement terms, the client's own policies and the plan you are using. Many agreements treat it as disclosure. This is not legal advice — if you are unsure, ask the client or a lawyer before you upload, or use a plan whose terms you and the client have both accepted.
No. Nothing stops the upload — the file is theirs to attach. Cloaking makes what the extractor gets noisy and unreliable, so a one-click summarise or templatise gives a poorer result. Copy-paste, screenshots, PDF export and extractors that strip hidden text bypass it. It is a deterrent, not a guarantee, and never a substitute for a contract.