Is it safe to upload files or PDFs to an AI? Training, storage, and third-party platforms

You can drop a long PDF into a chat box and ask “what is the deadline” or “summarize the conclusion in five points.” That is convenient, but “convenient” and “this file is appropriate to upload” are not the same question. What you need to separate first is: whether the content may be used to train a model, whether it will be stored, whether it is still there after you delete it, and — if you are using a third-party platform — whether the file first passes through the operator’s servers.

This article adds risks and a way to judge, not another feature tutorial. If you only want to know “how to ask a PDF a question,” the site already has an operational guide: How to use AI to chat with any PDF. What follows is the data path that article barely covers.

One clear conclusion: no responsible account should be written as “absolutely nothing is stored” or “it is definitely safe after you upload.” What you can do is first see the difference between training and storage, then decide whether this file should leave your device.

Separate two things first: training, and storage

Many people treat “will it be used for training” as the only question. That matters, but it is not the whole story.

Training means your conversation or file may be used to improve a model. Once it enters a training process, deleting the original conversation later is not the same as being able to “dig the information out of an already trained model and erase it.” Official products usually offer an option to turn training use off, but the default, the scope, and the exceptions (for example, you clicked feedback, or content was flagged by a safety system) follow each vendor’s settings at the time.

Storage means that, so you can continue a chat later, see history, handle abuse, or meet legal obligations, the operator or model provider keeps a copy of inputs, outputs, or files in its own systems. Even if content is not used for training at all, storage can still exist. Chat history itself is a form of storage. The Trymodel pricing page lists “unlimited chat history” as a plan capability, which means conversation records are kept so you can review them; it does not, by that fact, promise that content is not saved.

So four common combinations appear: no training, but still storage; both training and storage; you turned training off, but safety or legal reasons still require a retention period; you clicked delete in the interface, but the backend still has a retention window. Calling any one of these “fine, because training is off” is incomplete.

Why something may still remain for a while after you delete it

In official consumer products, “it disappeared from the list” usually does not mean “it disappeared from every system immediately.” Official notes checked on 11 November 2026 are roughly as follows (policies are updated; use the original pages):

• OpenAI / ChatGPT: A deleted conversation is removed from your account view immediately, and is planned to be permanently deleted from systems within 30 days; but if the content was previously de-identified and unlinked from the account, or must be retained for safety or legal reasons, that timeline may not apply. Temporary chats in the account are not used to improve models by default, and may still be retained for up to about 30 days for safety purposes. See Data controls in ChatGPT, How OpenAI handles data in consumer services, and Deleting and archiving chats in ChatGPT. In a personal workspace, “Improve the model for everyone” is on by default and can be turned off in settings; after it is off, new conversations are generally no longer used to improve models, but existing chats are not deleted automatically. See How your data is used to improve model performance and the OpenAI privacy policy.
• Anthropic / Claude: The privacy policy states that inputs and outputs in consumer services may be used to train and improve models unless you opt out in account settings; even after you opt out, content flagged in a safety review, or feedback you submit yourself, may still be used for improvement. After you delete a single conversation, it is removed from conversation history immediately and deleted from the backend within 30 days. See the Anthropic privacy policy (effective 10 September 2026) and the Privacy Center note on how long data is stored.
• Google, xAI, and others: Each has its own privacy policy and retention rules. The 30-day figures from the two vendors above cannot be substituted for them. Read the current policy directly before you use the product, for example the xAI privacy policy.

These rules only govern your relationship with the official product. They do not automatically apply to a third-party platform, and they should not be read as “every AI product is 30 days.”

Third-party platforms: the file usually takes an extra hop

If you are using an aggregator, a mirror, or a relay, rather than an official app or official website, the data path is usually:

Your browser → that platform operator’s servers → (possibly then) the model provider.

In other words, there is at least one additional visible processing party. The operator needs to receive the file in order to hand it to the model; it may also keep a copy in order to provide “history” or “continue on another device.” On the model-provider side, the agreement between that provider and this platform applies — and that agreement is usually not the consumer data-control set you tick for your own account on the official site.

Several facts should therefore be stated clearly:

1. You cannot turn off the “used to improve the model” switch in official ChatGPT or Claude settings from inside a third-party platform, because you are not operating in that official account.
2. The 30-day deletion cycle written in official docs does not automatically become a third-party platform’s commitment.
3. Even if a platform “connects to an official API,” that does not mean the official vendor will give this upload the same controls and after-sales support as the official client.

The terms and conditions of Trymodel state that the platform is operated by TUCANTELA SL, that it provides access to multiple third-party AI models through one interface, that third-party models are provided by the respective providers, and that you should not “submit confidential, regulated, or highly sensitive information without authorization.” That is not the same as “the file never leaves your computer.”

What the Trymodel privacy policy actually says, and what it does not say

The excerpts below come from the Trymodel privacy policy as of 11 November 2026 (effective date: 7 May 2026). They only restate what is on the page, and do not add guarantees the page does not make.

The policy states that it collects:

• Account information (name, email address, and other details);
• Usage data (features used, frequency, and interaction patterns);
• Device information;
• Cookies;
• Payment information is handled by third-party payment processors, and full card numbers are not stored on the servers.

Purposes of use include: providing, maintaining, and improving the service; processing transactions; sending technical notices and support information; and detecting and preventing fraud and abuse.

On sharing: the policy states that it does not sell personal information; sharing occurs only in limited cases, including with vendors that help operate the service under a confidentiality agreement, to comply with legal obligations or respond to lawful requests, and to protect the rights, property, and safety of the platform or of users.

On security: it uses “industry-standard security measures, including encryption in transit, access controls, and regular reviews,” and it states clearly that “transmission over the internet cannot be 100% secure.” Compared with the more colloquial “state-of-the-art encryption protocols” on the pricing-page FAQ, the privacy policy’s wording should be treated as authoritative.

On rights: you have the right to access, update, export, or delete personal information, through account settings or the contact method provided on the site.

Important matters the site has not stated in public include:

• Whether conversation content, prompts, and uploaded PDFs / files are stored, for how long, and where;
• Whether that content will be used by Trymodel or by a model provider for training, or for model training as part of “improving the service”;
• After you delete a conversation or a file in the product, how long it is retained in the backend and in backups;
• Which model providers content is sent to, and which data-processing agreement applies;
• A contact email is not written directly in the body of the privacy policy; the policy only says to use the email provided on the site or the contact page.

Because these points are not stated clearly, this article will not describe Trymodel as “not storing user files” or “not used for training.” A more accurate statement is: it is an independent third-party platform; requests enter its service and may be shared with model vendors; when sensitive content is involved, follow the limits in the terms, and narrow what you upload yourself.

Which files not to upload

Whether official or third-party, the following should not be uploaded by default, and should not be pasted into a chat box:

• Identity and money: ID cards, passports, driver’s licenses, bank card numbers, account passwords, verification codes, payment QR codes.
• Access credentials: API keys, database passwords, SSH private keys, unpublished internal addresses, temporary login links.
• Regulated or highly sensitive personal data: medical records, complete financial statements, information about minors, biometric data, and customer or employee files you are not authorized to handle.
• Material you do not have redistribution rights for: contracts someone else allowed only you to review, unpublished merger documents, secrets bundled inside customer source code, unpublished financial reports.
• Anything that, if leaked, would force you to change passwords or call the police immediately. If it appearing on a public web page would make you uneasy, do not upload it “just to let the AI read it first.”

The note in Trymodel’s terms (do not submit confidential, regulated, or highly sensitive information without authorization) is consistent with that judgment. The platform may functionally allow you to drop in a PDF; being allowed to upload is not the same as this file being appropriate to upload.

How to redact when you must have an AI read a document

If the work really requires a summary or an extracted clause, narrow the scope first, then decide whether to upload:

1. Ask whether you can paste only the relevant passages. An 80-page contract may need only the “termination” and “payment” sections. If you can copy excerpts, do not upload the full text.
2. Remove direct identifiers. Replace real names, company names, addresses, amounts, account numbers, phone numbers, and project codes. Keep the structure that is useful for the question.
3. Remove hidden information in attachments. PDFs and office documents may carry revision history, author names, hidden worksheets, or cropped layers that can still be recovered. If needed, export to plain text and check again.
4. Do not put secrets and a document in the same package. Source repositories, config bundles, and “readme” files are often packed together; split them before uploading.
5. After you finish, check on that platform and on your own device whether you can still delete it. Also remember: deletion may still have a retention period, and should not be treated as already gone.
6. Check the output yourself. The AI may repeat information you failed to notice before redacting, or it may summarize incorrectly. Do not send the answer directly to a client or a court.

These steps cannot turn a third-party path into “the same as the file never leaving the computer.” They can only reduce the loss if a leak happens.

How to read this together with the existing “chat with a PDF” guide

How to use AI to chat with any PDF covers upload, asking questions, and checking against sources, and it notes that a model better suited to long text can be used for very long documents. That is a feature guide. When you read it together with this article, it helps to ask three more questions:

• Does this PDF contain any of the sensitive content listed above? If it does, redact first, or do not upload.
• Am I using an official client right now, or a page that will pass through a third-party operator?
• Do I accept that the conversation may enter history, and that after deletion it may not disappear from every system immediately?

If you cannot answer one of those three, do not upload the original file yet. If you continue to use Trymodel, you can start from the chat page, but practice the flow first with material you can discuss in public. When you choose a platform, you can also use How to choose a safer third-party AI aggregator. For legitimate access from China, see How to use Claude, ChatGPT, and Grok from China. For payment and whether charges renew automatically, see Trymodel pricing and billing.

FAQ

If I upload a PDF to an AI, is it safe as long as I turn off “used for training”?

No. Turning off training use only affects, in an official product, whether content is used to improve a model, and even then you have to follow that product’s settings and exceptions at the time. The file may still be stored, and there may still be a retention period after deletion. On a third-party platform, you usually also cannot use the switch inside an official account.

If I deleted the conversation, is the file gone immediately?

Not necessarily. In official consumer products, a common practice is to remove it from your view first, then delete it from the backend after a period (30 days is common in OpenAI’s and Anthropic’s public notes, with safety or legal exceptions). Trymodel’s privacy policy does not state a retention period after a conversation or file is deleted, so you should not infer 30 days.

Will a third-party platform send my file to the model company?

To let the model read the content, the relevant text or file usually has to be sent to the party that actually provides the model. Trymodel’s terms state that third-party models are provided by the respective providers, and the privacy policy also mentions possible sharing with vendors that help operate the service. The site has not published a list of each recipient or their retention rules.

Does Trymodel absolutely not store what I upload?

The site has not made an “absolutely not stored” promise, and this article will not make that promise on its behalf. The pricing page lists unlimited chat history as a capability, and the privacy policy says usage data is collected and used to provide and improve the service, but it does not separately state a retention period for file contents. The more careful practice is: decide whether to upload based on whether the content itself can leave your control.

What material is relatively more suitable for trying “chat with a PDF”?

Material you are authorized to handle, and whose leak would cause limited harm: for example, a paper already made public, a draft manual you wrote yourself, or sample text with names and account numbers removed. Do not use a real unsigned contract, a customer list, or an archive that contains secrets for a first trial.

Can I request deletion of personal data?

The Trymodel privacy policy states that you have the right to access, update, export, or delete personal information, through account settings or the contact channel. That is not the same as having published a precise erasure timeline for conversations and attachments. For official products, use each vendor’s own tools (for example, ChatGPT data controls or Claude privacy settings) and read their retention notes.

FAQPage structured data

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "Is it safe to upload a PDF to an AI?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Do not assume it is safe. Training and storage are not the same thing: turning off training use does not mean the file is not saved, and there may still be a retention period after deletion. On a third-party platform, the file usually passes through the operator’s servers first, and may then be sent to the model provider. ID documents, secrets, regulated data, and unauthorized confidential files should not be uploaded."
      }
    },
    {
      "@type": "Question",
      "name": "If I delete a conversation on a third-party AI platform, does the file disappear immediately?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Not necessarily. In official consumer products, a common practice after deletion is to remove the item from the account view first, then delete it from the backend after a period, which may be extended for safety or legal reasons. A third-party platform should be judged by its own privacy policy. The Trymodel privacy policy does not currently state a retention period after conversations and uploaded files are deleted."
      }
    },
    {
      "@type": "Question",
      "name": "Does Trymodel absolutely not store files users upload?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "That promise is not made. The Trymodel privacy policy describes how account information, usage data, device information, and payment information are handled, but it does not clearly state the retention period for conversation content and uploaded files, or whether they are used for training. The terms also note that you should not submit confidential, regulated, or highly sensitive information without authorization."
      }
    }
  ]
}

Facts were checked on 11 November 2026 (Asia/Shanghai). Official and Trymodel privacy terms may be updated; use the original policy pages on the day you check.