Does a virtual data room train AI on my documents?
Do the bytes leave the vendor's cloud, is the prompt kept after the call, and is any of it used to improve a model.
A data room can run a model on your files without training on them if those three answers are no.
An MCP connector to ChatGPT or Claude differs: bytes leave, and that vendor's plan decides training.
Consumer ChatGPT is not that bar.
A homepage slogan is not that bar either.
Get the three answers in writing, on a legal URL.
No-train data room (definition). A room whose model call does not keep the prompt and does not use the files to improve any model.
"We use AI" is not an answer.
A homepage slogan is not an answer.
An MCP connector is not the room's model call.
The three questions above are the answer.
Does a data room train on my files?
Only if the answer to one of the three questions is yes. A model can read a file for Ask or search and still not train on it. The bytes can stay in the vendor's cloud. The prompt can be discarded when the request finishes. None of that is automatic. Get each answer on a public legal URL, not a feature page.
Is an MCP connector the same as the room's own model?
No. Model Context Protocol (MCP) is a bridge that lets ChatGPT, Claude, Copilot, or another assistant read files that still sit in the room. The room vendor's no-train clause does not cover that call. The bytes leave to the assistant's vendor, and that vendor's plan decides training. Rooms that ship an MCP connector say the same thing in their own FAQ: check the assistant's plan, not the room's policy. Pasting the packet into consumer ChatGPT or Gemini is the same leave, without the bridge. See whether you should connect ChatGPT to a data room.
Which setups train, keep, or leave
| Setup | Bytes leave the room? | Prompt retained after the call? | Used to train a model? | Use it for diligence? |
|---|---|---|---|---|
| Paste into consumer ChatGPT / Gemini | Yes. They leave. | Retained under consumer rules. | May train, unless you opt out of Improve the model for everyone on OpenAI's consumer page (retrieved 18 September 2026). | No. |
| MCP connector on a consumer ChatGPT or Claude plan | Yes. They leave to OpenAI or Anthropic. | Retained under consumer rules. | ChatGPT may train unless you opt out. Claude chats may train if Model Improvement is on; Anthropic's consumer page (16 March 2026, retrieved 9 October 2026) excludes raw MCP connector content, not text copied into the chat. The room's no-train clause covers neither. | No. |
| ChatGPT Business / Enterprise / API, pasted or over MCP | Yes. They leave to OpenAI. | See OpenAI's business data page (retrieved 18 September 2026). Default API abuse logs can last 30 days unless zero data retention is approved. | No by default on ChatGPT Business, Enterprise, and the API. | Only if you are willing to send the packet to OpenAI. |
| Claude for Work / Anthropic API, pasted or over MCP | Yes. They leave to Anthropic. | Set by Anthropic's commercial terms. | No by default, per Anthropic's commercial training page (18 August 2026, retrieved 9 October 2026), unless you send feedback or opt in. | Only if you are willing to send the packet to Anthropic. |
| In-room model with a published no-train + no-keep policy | No, if the vendor's legal page says so. | No, if zero data retention or the equivalent is on that page. | No. | Yes, if the three answers are on a public legal URL. |
| Room where an admin turns AI off per project | No model call while it is off. | n/a | n/a | Yes, if you will not run a model at all. |
| No-AI overlay (files stay in Drive) | n/a | n/a | n/a | Only if you accept Drive's own sharing holes. |
| Cashpian | Files stay in the room. The model is Amazon Bedrock. | Privacy Policy: Bedrock Zero Data Retention. The prompt is not retained after the request. | Privacy, Terms of Service§5.3, and the homepage: not used to train any model, ours or a third party's. Amazon Bedrock (retrieved 9 October 2026): inputs and outputs are not used to train base models. | Yes, for the packet. Do not also connect consumer ChatGPT. |
Consumer ChatGPT loses the diligence column, pasted or through an MCP connector. A room where an admin can turn AI off per project wins if you will not run a model at all; Cashpian has no such toggle. A no-AI overlay can win too, if you accept Drive's sharing holes. ChatGPT Business or the API still sends the packet to OpenAI, and Claude for Work still sends it to Anthropic.
The Cashpian homepage already publishes these lines: Never used to improve a model - ours or anyone else's. The model keeps nothing it saw. Your files stay in your room. We don't mine them. The assurance strip says Zero data retention. Read the Privacy Policy for the legal wording, including Bedrock Zero Data Retention.
Terms §5.3: Cashpian will not use Room Content to train, fine-tune, or improve any AI or machine learning model, whether ours or a third party's. Terms §9.1: the model is Claude on Amazon Bedrock under a Zero Data Retention configuration and a signed Business Associate Agreement, as described in the Privacy Policy.
No-keep and no-train are different guarantees here too. Retention is switched off on the Bedrock account, so Bedrock refuses to run any model that would keep the prompt. No-training rests on AWS's Bedrock terms and Terms §5.3. See What is Cashpian?, which data room has AI that organizes documents, and pricing.
Can a feature page stand in for a legal URL?
No. Bank-grade, secure by design, and a redaction screenshot are not answers. The three questions are the answer, on a public legal page you can cite: privacy policy, terms, or a dated FAQ that names leave, keep, and train. If the only no-train sentence lives on a product marketing URL, get it restated on a legal URL before you put a live packet in that room.
Can a data room stop a phone camera?
No. Nobody can stop a phone camera. That is why every page is watermarked to the viewer and every open is tracked. A no-train policy does not cover a photograph of the screen.
FAQ
If the room has Ask or search, does that mean it trains?
No.
Running a model on a file is not the same as training on it.
Ask and search send a prompt for that request.
Training would keep the files to improve a later model.
Get the three answers in writing.
Does an MCP connector to ChatGPT or Claude train on the files?
The room vendor's no-train clause does not cover that call.
The files leave to OpenAI or Anthropic, and that vendor's plan decides training.
ChatGPT Business, Enterprise, the API, Claude for Work, and the Anthropic API do not train by default.
Consumer plans can.
Is "we do not train" the same as "the model keeps nothing"?
No.
No-train means the files are not used to improve a model.
No-keep means the prompt is not retained after the call.
A vendor can say no-train and still keep logs for 30 days.
Ask both.
Can I turn the model off and still use the room?
Some rooms let an admin turn AI features off per project.
Cashpian has no admin toggle that turns the model off.
The model runs on Ask, classify, and gap-detection, and that call is not a second copy of the files.
The room still watermarks pages and logs opens.
Sources
- Cashpian homepage security band and assurance strip, retrieved 18 September 2026.
- Anthropic, Is my data used for model training? (commercial), 18 August 2026, retrieved 9 October 2026. Claude for Work and the API do not train by default.
- Anthropic, Is my data used for model training? (consumer), 16 March 2026, retrieved 9 October 2026. Chats train if you allow model improvement. Raw content from MCP connectors is excluded; text copied into the chat is not.
- Amazon Web Services, Amazon Bedrock security, privacy, and responsible AI, retrieved 9 October 2026. Inputs and outputs are not shared with model providers or used to train base models.
- Cashpian Privacy Policy, section 4, and Terms of Service §5.3 and §9.1.
- OpenAI, Business data privacy, security, and compliance, retrieved 18 September 2026.
- OpenAI, How your data is used to improve model performance, retrieved 18 September 2026.
If the files are already in a folder, grade which row they are on. Cashpian is a gated room for that packet. It is not a listing portal. It does not sell the company or run the deal.