The short answer: it depends on your plan, and "safe to upload" is the wrong question
If you are asking whether it is safe to upload company documents to ChatGPT, the honest answer is that it depends on your plan, your settings, and the type of document. Consumer ChatGPT, Team, and Enterprise or API each treat your data differently, so there is no single yes or no. The more useful reframe is this: the question your employer and your IT team actually care about is not "is ChatGPT private," it is "are you allowed to transfer this file to an outside company at all." For a document covered by an NDA, a data processing agreement, or a sector rule, the act of sending it to a third-party service can be the problem on its own, whether or not anyone ever leaks it.
What actually happens when you upload a file to ChatGPT
When you upload a document to a cloud AI service, the file does not stay on your computer. Here is the real data path:
- The file leaves your device and travels over the internet to the provider's servers.
- It is processed there, and the content may be stored for a period of time.
- Depending on your plan and settings, the content may also be used to help train or improve models.
This is true of any cloud-based assistant, not just ChatGPT. The point is that "uploading" is a transfer of custody. Your confidential document now exists, at least temporarily, on infrastructure you do not control and cannot inspect.
Training opt-out is real, but it has limits
OpenAI does let you turn off model training in some plans, and that setting is real. It is important, though, not to overstate what it does. Turning off training is not the same as "my file never leaves" or "my file is never stored."
- Opt-out stops training use. It tells the provider not to use your content to improve models.
- It does not stop transfer. Your document still travels to external servers to be processed.
- It does not always stop retention. Conversations may still be retained for a period, for example to monitor abuse or to meet legal obligations.
So "I turned training off" answers one narrow concern and leaves the transfer-and-retention concern untouched. For confidential files, that distinction matters.
How the plans differ: consumer vs Team vs Enterprise and API
The accurate picture is plan-dependent. Treat the following as a map of how terms tend to vary, and confirm it against your current contract and settings.
- Consumer ChatGPT (Free, Plus, Pro): Content may be used to train models unless you opt out. Retention and controls are aimed at individuals, not regulated business data.
- ChatGPT Team: Business-oriented terms. Business data is generally not used to train models by default, with admin-level controls.
- ChatGPT Enterprise and the API: Business terms state that business data is not used to train models by default, typically with stronger retention controls and contractual commitments suitable for organizations.
In every case the document still leaves your machine and is processed on external infrastructure. The plan changes the contractual and training terms, not the basic fact of transfer.
The question your IT team actually cares about: who is liable
Most articles frame this as privacy. The sharper, employee-level question is accountability. If a document is governed by an agreement, sending it to an outside processor can itself be a breach.
- NDAs often restrict disclosure of confidential information to third parties. A cloud AI provider is a third party.
- Data processing agreements (DPAs) specify exactly which sub-processors are permitted. An unapproved AI tool is usually not on that list.
- Sector rules add their own constraints. Legal, accounting and tax, finance, public sector, and healthcare and research all handle information where unauthorized external transfer is a serious matter.
The key insight is that the trigger is the transfer, not the leak. You can get into trouble for uploading a confidential file to an unsanctioned tool even if that file is never exposed, because the unauthorized transfer is itself the violation. This is why on-device processing is a different category. It removes the transfer entirely.
A practical checklist before you upload anything
Run through this before pasting or uploading any work document:
- Is it confidential? Customer data, contracts, financials, source code, patient or personal records, unreleased plans. If yes, slow down.
- Is it covered by an agreement? An NDA, DPA, employment policy, or client contract may prohibit third-party transfer outright.
- Is there a sanctioned tool? If your organization provides an approved Enterprise or on-device option, use that and nothing else.
- If in doubt, do not upload. When you cannot confirm that it is allowed, treat the file as not uploadable and find an alternative.
The safer alternative: on-device document AI
On-device (local) document AI means the AI processes your files directly on your own computer, so the document content is not transmitted to an external server or cloud. This is the category that removes the external transfer at the root, because there is no outbound trip for your document to take.
It fits anyone who handles material they cannot or should not put in the cloud: legal, accounting and tax, finance, public sector, healthcare and research, and corporate security. Instead of choosing between doing the work and breaking a rule, you do the work locally and the rule is never tested.
Comparison table: cloud ChatGPT vs on-device document AI
| Aspect | On-device document AI (Apioni) | Cloud ChatGPT |
|---|---|---|
| Where document data goes | Stays on your machine | External servers |
| Outbound network for documents | None for document content | Required to function |
| Retention of content | Local files only, on your disk | May be retained for a period |
| Training use of your content | Not applicable, nothing is sent | Varies by plan and settings |
| Liability and transfer | No external transfer to authorize | Transfer to a third party occurs |
| Offline use | Works without internet | Requires a connection |
| Pricing model | One-time purchase | Subscription |
How Apioni keeps document content on your machine
Apioni is an on-device document AI desktop app. It runs on your own computer and works without an internet connection. It is built as a Tauri desktop app, signed and notarized on macOS and available for Windows, and it runs a local LLM (Ollama with a Qwen3-family model) together with a local embedding model (bge-m3) directly on your machine.
Here is how it works in practice. You connect a folder of documents, and Apioni indexes them locally. It then answers questions, summarizes, and drafts documents while citing the source passages it used. Retrieval is a hybrid search, which means it combines keyword search (BM25) with semantic search and fuses the results with RRF, so you get both exact-term matches and meaning-based matches.
The network behavior is the part that matters most for confidential work, so here it is precisely:
- All document processing goes only to the local engine on your machine (localhost). Your document content never leaves your computer.
- The single connection Apioni makes to the internet is a version check to apioni.com when the app launches. It carries no document content and only checks whether a newer version exists.
- There is no usage telemetry.
- You do not have to take this on faith. You can verify it yourself with the in-app network self-check, or with standard operating system tools such as
lsof.
Apioni is a one-time purchase of KRW 39,000 with no subscription. The license is company-wide with no device limit, and there is a 14-day free trial that requires no card or payment details. You can download Apioni and try it before you decide.
Frequently asked questions
Does ChatGPT train on the documents I upload?
It depends on the plan and your settings. On consumer plans, uploaded content may be used to train models unless you opt out. ChatGPT Enterprise and the API state that business data is not used to train models by default. In all cases, opting out of training does not stop your document from being transferred to and processed on external servers.
Is ChatGPT Enterprise safe enough for confidential files?
Enterprise terms are meaningfully stronger than consumer plans: business data is not used for training by default, and retention and contractual controls are built for organizations. The document still leaves your machine and is processed on external infrastructure, however, so whether it is appropriate depends on your NDAs, DPAs, and sector rules. Enterprise reduces several risks, but it does not remove the external transfer.
Can I get in trouble for uploading company files to ChatGPT?
Yes, this is possible even if no leak ever occurs. If a document is covered by an NDA, a data processing agreement, or a sector regulation, transferring it to an unapproved third-party service can be the violation by itself. The trigger is the unauthorized external transfer, not whether the file is later exposed.
What is on-device document AI and how is it different?
On-device document AI processes your files on your own computer rather than sending them to a provider's servers. The difference is that the document content is not transmitted externally, which removes the transfer that creates most compliance and confidentiality concerns. Apioni is an example: it runs a local LLM and embedding model and keeps document content on your machine.
How can I verify that a local AI tool is not sending my files anywhere?
Check its outbound network activity. With Apioni you can use the built-in network self-check, or run standard operating system tools such as lsof to observe connections. You should see document processing staying on localhost, with the only outbound connection being a version check that carries no document content. If you want to confirm before committing, start with the 14-day free trial, which needs no card.