← Back to blog

Is It Safe to Upload Grant Documents to AI Tools?

How ChatGPT, Claude, Gemini and Copilot handle uploaded files by account type, which funder rules limit what you may upload, and a safer workflow for grant documents.

A hand pausing with a plain envelope pushed halfway into a street mailbox slot

In this article (10)

Is it safe to upload grant documents to AI tools? It can be, once three checks pass: the account you are signed in to, the kind of data in the file, and the rules of the funder whose application it is. The same product name can hide opposite defaults, since OpenAI may train on content from its services for individuals but by default does not train on ChatGPT Business, ChatGPT Enterprise or the API (OpenAI Help Center, read September 2026).

A grant file also carries more than prose. Budgets list salaries, letters of support name partners, and research applications can hold unpublished results or data covered by a use agreement. Each of those meets a different rule.

Quick Answer: Is It Safe to Upload Grant Documents to AI Tools?

Uploading grant documents to AI is safe only when the account, the data and the funder all allow it. Personal ChatGPT, Claude, Gemini and Copilot accounts can feed model training, depending on settings; the same providers' business plans restrict training by default. Some data, such as NIH controlled-access genomic data, may never go into public AI tools.

Choosing grant software? Compare GrantCue's plans with the alternatives. See plans and pricing

What Happens to a Grant File After You Upload It

What happens to an uploaded file depends on the provider and the account type. Some accounts may use it for training, providers keep content for set periods even after you delete it, and deleting a chat does not always delete the file.

Provider and accountTraining on your contentHow long content is keptSource
ChatGPT, individual accounts such as PlusMay be used unless you turn off "Improve the model for everyone"Deleted chats scheduled for permanent deletion within 30 days; uploaded files stay in Library until deleted manuallyOpenAI Help Center, read September 2026
ChatGPT Business and Enterprise, OpenAI APINot by defaultChatGPT Business: workspace admins control retention, and deleted or unsaved conversations are removed within 30 days unless law or harm prevention requires longerOpenAI Enterprise privacy page, January 8, 2026
Claude Free, Pro and MaxThe user chooses; if allowed, de-identified data may be kept up to 5 years in training pipelinesDeleted conversations removed from back-end storage within 30 daysAnthropic Privacy Center, July 1, 2026
Claude for Work and the Anthropic APINot by defaultDeleted conversations removed within 30 days; API inputs and outputs deleted within 30 days, with listed exceptionsAnthropic Privacy Center, training page read September 2026; retention page July 1, 2026
Gemini Apps, personal Google accountUsed for training, with human reviewers, when Keep Activity is onChats saved 72 hours even with Keep Activity off; human-reviewed chats kept up to three yearsGemini Apps Privacy Hub, July 15, 2026
Gemini with a qualifying Google Workspace editionNot used for training outside your domain without permissionSet by Workspace termsGoogle Workspace Privacy Hub, August 14, 2026
Copilot, consumer Microsoft AccountUsed for training unless you opt out or fall in an excluded groupConversation activity 18 months by default; uploaded files no longer than 18 monthsMicrosoft Copilot Privacy FAQ, August 31, 2026
Microsoft Copilot, work Entra ID accountPrompts and responses not used to train foundation modelsCovered by Microsoft's Data Protection Addendum, with Microsoft as data processorMicrosoft Learn, August 18, 2026

Deleting a chat is not the same as deleting a file. On ChatGPT plans with Library, uploaded files "are saved to your account until you delete them manually," and "Deleting a chat containing files does not delete those files saved to Library" (OpenAI Help Center, File storage and Library in ChatGPT, read September 2026).

A deleted ChatGPT chat is "removed from your account immediately and scheduled for permanent deletion from OpenAI systems within 30 days," unless it was already de-identified or "OpenAI must retain it longer for security or legal obligations" (OpenAI Help Center, Chat and File Retention Policies, read September 2026).

Legal obligations are not hypothetical. OpenAI's page on the New York Times litigation, updated October 22, 2025, says "we are no longer under a legal order to retain consumer ChatGPT and API content indefinitely" and that its obligations under the earlier order ended on September 26, 2025, while it still stores limited historical user data from April to September 2025 under legal hold (OpenAI, response to New York Times data demands, updated October 22, 2025).

Personal and Business AI Accounts Follow Different Rules

The account you sign in with decides which policy applies, and product names no longer make that obvious. Check the sign-in before you upload anything.

Microsoft Learn notes: "Microsoft 365 Copilot is now named Microsoft Copilot, and Microsoft 365 Copilot Chat is now named Microsoft Copilot Chat" (Microsoft Learn, updated August 18, 2026). Microsoft's consumer privacy FAQ "does not apply to the use of Microsoft 365 Copilot when signed in with Entra ID" (Microsoft Copilot Privacy FAQ, August 31, 2026). The name on the screen can read Copilot either way.

Google draws the same line. "If you have a work or school Google Account, your use of Gemini Apps may be subject to different data handling terms," Google says (Gemini Apps Privacy Hub, updated July 15, 2026). For qualifying Workspace editions, "Your content is not human reviewed or otherwise used for Generative AI model training outside your domain without permission," and Google calls that commitment contractual under the "Training Restriction" section of its Workspace Service Specific Terms (Google Workspace Privacy Hub, updated August 14, 2026).

Anthropic's consumer terms change applies to Claude Free, Pro and Max, and the company states: "They do not apply to services under our Commercial Terms, including Claude for Work, Claude for Government, Claude for Education, or API use" (Anthropic, Updates to Consumer Terms, August 28, 2025). OpenAI's enterprise commitments cover "ChatGPT Business, ChatGPT Enterprise, ChatGPT for Healthcare, ChatGPT Edu, ChatGPT for Teachers and our API Platform" (OpenAI Enterprise privacy page, January 8, 2026).

No-training promises are about training, not storage: ChatGPT Business and Anthropic's commercial products still take up to 30 days to remove deleted conversations (OpenAI Enterprise privacy page, January 8, 2026; Anthropic Privacy Center, July 1, 2026). Microsoft's organizational statement says prompts and responses "aren't used to train foundation LLMs" (Microsoft Learn, updated August 18, 2026), which says nothing either way about how long they are kept. One commenter on Reddit put the caution bluntly: "At the large research university level we are still warned to avoid uploading grants, even with our proprietary license" (Reddit thread 1ubd86x, read September 2026).

Two similar lanyard badges, one blue and one grey, on a desk beside a closed laptop

Turning Off Training Does Not Erase What You Shared

A training opt-out covers future conversations only. It does not delete what is already stored, and a feedback click can put a whole conversation back into use.

For its services for individuals, OpenAI says "Once you opt out, new conversations will not be used to train our models," but if you give thumbs-up or thumbs-down feedback, "the entire conversation associated with that feedback may be used to train our models" (OpenAI Help Center, read September 2026). Anthropic says that when a Claude Free, Pro or Max user turns the setting off, it "will not use any new chats and coding sessions you have with Claude for future model training," though data "will still be included in model training that has already started and in models that have already been trained" (Anthropic Privacy Center, August 3, 2026). Feedback on those plans works differently: after a thumbs up or down, "we will store the entire related conversation, including any content, custom styles or conversation preferences, in our secured back-end for up to 5 years" (Anthropic Privacy Center, March 16, 2026).

On a personal Google account with Keep Activity off, future chats "won't be used to train our AI models, unless you choose to send Google feedback," but "Future chats are still saved for 72 hours" (Gemini Apps Privacy Hub, updated July 15, 2026). For consumer Copilot, opting out "will exclude your future conversations from being used for training these AI models," and "an opt-out of human review is not available" (Microsoft Copilot Privacy FAQ, August 31, 2026).

Temporary modes reduce training exposure without making a chat disappear:

ModeTrainingStorageSource
ChatGPT temporary chatNot used to improve models while it remains temporaryA copy may be kept up to 30 days for safety; saving it as a regular chat returns it to your normal settingOpenAI Help Center, read September 2026
Claude Incognito chat (consumer plans)Not used to improve Claude, even with Model Improvement onIf flagged for a Usage Policy violation, inputs and outputs kept up to 2 years whatever the settingAnthropic Privacy Center, March 16 and July 1, 2026
Gemini temporary chat (personal account)Not used to train Google's modelsKept with your account for 72 hoursGemini Apps Privacy Hub, July 15, 2026

What Funders Say About Uploading Grant Material

Funders speak to two audiences. Applicants get cautions and data limits; peer reviewers get prohibitions on uploading other people's applications.

Rules for Applicants

NASA's Science Mission Directorate warns proposers that anything given to a generative AI tool as a prompt "will often be stored and may be used by others," so they should keep in mind "that this information may aid a potential competitor" (NASA SARA FAQs, question 32, read October 2026). The competitor point applies with extra force to SBIR and STTR applicants, whose proposals describe technology a rival would value.

UKRI says "the confidentiality of information entered into generative AI tools is not guaranteed," and "Sensitive or personal data of others must never be input into a generative AI tool without formal consent from the individual" (UKRI generative AI policy, last updated July 6, 2026). Letters of support are the obvious case, since they carry other people's names and details. Wellcome requires applicants to "protect confidential information and personal data by not inputting it into generative AI tools" (Wellcome, April 8, 2026).

Other funders say similar things. Halifax Foundation NI, a Northern Ireland funder, tells applicants that "many of these free models retain data," that "we encourage you to avoid putting confidential or sensitive information into free software," and that anyone paying for a tool should check "their data protection and retention policies" (Halifax Foundation NI, page last updated July 2025).

Rules for Reviewers, a Separate Audience

NIH tells its reviewers that "uploading or sharing content or original concepts from an NIH grant application, contract proposal, or critique to online generative AI tools violates the NIH peer review confidentiality and integrity requirements" (NIH notice NOT-OD-23-149, released June 23, 2023). NSF prohibits reviewers from uploading proposal content "to non-approved generative AI tools" and says "Any information uploaded into generative AI tools not behind NSF's firewall is considered to be entering the public domain" (NSF notice, December 14, 2023). NEH "prohibits peer reviewers from uploading NEH applications into a cloud database" (NEH AI policy, October 23, 2024).

The Wenner-Gren Foundation closes the gap a no-training plan might seem to open: reviewers must never upload applications "even if its terms of use state that uploaded information will not be used elsewhere" (Wenner-Gren Foundation AI policy, March 19, 2025). Seven UK funders say in a joint statement that they "may use enterprise data protected (EDP) generative AI tools to process funding applications" but "will not input confidential funding applications into publicly available versions of generative AI tools" (UK funders' joint statement, revised June 2026). These rules govern the funder's side of review, and they show how funders treat an application once it leaves the applicant's hands.

A crumpled page in a wastebasket while an identical page stays pinned to a corkboard above

Data That Should Stay Out of Public AI Tools

Some rules turn on the data, not the plan, so a no-training account does not make every upload acceptable.

MaterialRuleSource
Controlled-access human genomic dataSharing it with public generative AI tools violates the Data Use CertificationNIH NOT-OD-25-081, March 28, 2025
ITAR or EAR material in a NASA science proposalMust not be uploaded to a website accessible to non-US personsNASA SARA FAQs, question 32, read October 2026
Other people's sensitive or personal dataNever without formal consent from the individualUKRI policy, July 6, 2026
Confidential business or personal documents in consumer ClaudeAnthropic advises being thoughtful about sharing themAnthropic Help Center, May 22, 2026
Anything you would not want a human reviewer to see, in personal GeminiGoogle asks you not to enter itGemini Apps Privacy Notice, June 29, 2026

The NIH rule is specific: "sharing controlled-access data with public generative AI tools (e.g., third party tools) via prompts or other user interfaces is in violation of the provision on Non-Transferability, and by extension, the DUC," meaning the Data Use Certification under NIH's Genomic Data Sharing Policy (NIH notice NOT-OD-25-081, released March 28, 2025).

Treat the NIH rule as a hard stop. A research team that holds controlled-access data should keep it out of every public AI tool, whatever plan the team pays for, and settle any question about a specific tool against the Data Use Certification it signed.

A Safer Workflow for Grant Documents

A safer workflow sorts each document before it goes anywhere, then matches the account and the redaction to the most sensitive thing in the file. The same steps apply when the AI sits inside a grants tool, and our guide to grant management software for nonprofits lists the AI features worth testing in one.

StepWhat to doWhy it matters
1. SortLabel each file public, internal or restrictedRestricted material stays out of every public tool
2. Check the funderRead the announcement's AI and data clausesFunder rules can be stricter than any provider's terms
3. Pick the accountUse your organization's business plan where one exists; otherwise a personal account with training turned offDefaults differ by account type
4. RedactRemove names, addresses, salaries and partner detailsPersonal data of others needs consent under some funder rules
5. Skip feedback buttonsDo not rate grant chats with thumbs up or downFeedback can pull a whole conversation into training or long storage
6. Clean upDelete the chat and any copy saved to LibraryDeleting a chat can leave the file behind

The Spencer Foundation states the redaction principle plainly: "Any information that is uploaded to generative AI systems should be anonymized or deidentified wherever possible to protect individual and community privacy" (Spencer Foundation AI policy, listed August 21, 2025). Its policy names the kinds of personal data an unpublished proposal can expose, such as "home addresses, phone numbers, salary details" (Spencer Foundation AI policy, listed August 21, 2025).

Practitioners on Reddit describe their own versions of this. One commenter wrote: "i omit details that are proprietary like names, organization, structure of the organization, but i do ask for general reviews of concepts of my idea" (Reddit thread 1ubd86x, read September 2026). Another, in the same thread, said "I don't upload any confidential or sensitive information" (Reddit thread 1ubd86x, read September 2026).

A hand blacking out several lines on a printed page with a thick marker

Check the Funding Announcement Before Any Upload

The funder's own documents come first, because some funders put their AI terms inside the announcement or a policy it cites. DOE's standard Financial Assistance NOFO Part 2 (Version 4.0) has a section II.B.4 titled "Artificial Intelligence (AI) Application Use" (DOE NOFO Part 2, copy attached to DE-FOA-0003627, created July 6, 2026). NEH's AI policy, dated October 23, 2024, is still cited in NEH's 2026 notices, such as Scholarly Editions (NEH Scholarly Editions notice, read September 2026). NASA's Science Mission Directorate answers AI questions in its SARA FAQs, question 32 (NASA Science, page updated October 13, 2026).

Where the documents say nothing, ask. Our guide to contacting a program officer before you apply covers how to put a short question in writing.

Announcements are long, and the terms you need are spread through them. GrantCue's NOFO PDF Summarizer is built for that reading job: "AI extracts key information from lengthy Notice of Funding Opportunity documents in seconds," covering key dates and deadlines, eligibility, funding amounts and priorities, and cost sharing and the application process (grantcue.com, September 2026). GrantCue's Free plan includes 3 AI NOFO summaries a month (grantcue.com, September 2026). A summary is a starting point: the announcement's own AI and data clauses still need a full read, and the NOFO summarizer on the features page lists what each summary covers.

FAQ

Does ChatGPT Plus Train on Uploaded Grant Proposals?

It can. Plus is one of OpenAI's services for individuals, where OpenAI says "we may use your content to train our models" (OpenAI Help Center, read September 2026). Turn off "Improve the model for everyone" under Settings, then Data Controls, or use a temporary chat. OpenAI's no-training default covers business products such as ChatGPT Business, ChatGPT Enterprise and the API (OpenAI Help Center, read September 2026), and Plus is not on that list.

Does Deleting a ChatGPT Chat Delete the Files I Uploaded?

No, on plans with Library. OpenAI says uploaded files "are saved to your account until you delete them manually" and "Deleting a chat containing files does not delete those files saved to Library" (OpenAI Help Center, read September 2026). Those plans include ChatGPT Free, Go, Plus, Pro and Business, and Enterprise, Edu and Healthcare workspaces. Delete the file from Library as a separate step.

Are Temporary Chats Safe for Confidential Grant Drafts?

They reduce training exposure but still leave a stored copy. A ChatGPT temporary chat is not used to improve OpenAI's models while it stays temporary, but OpenAI may keep a copy for up to 30 days for safety purposes (OpenAI Help Center, read September 2026). Gemini temporary chats on a personal account are kept with your account for 72 hours (Gemini Apps Privacy Hub, updated July 15, 2026). For Wellcome applicants the answer is no: confidential information stays out of generative AI tools altogether (Wellcome, April 8, 2026).

Is Microsoft Copilot at Work Safe for Grant Documents?

Check which account you are signed in with first, since Microsoft 365 Copilot is now named Microsoft Copilot (Microsoft Learn, updated August 18, 2026). With a work Entra ID account, prompts and responses are not used to train foundation models, and use is covered by Microsoft's Data Protection Addendum with Microsoft as data processor (Microsoft Learn, updated August 18, 2026). That is a training commitment, not a promise that nothing is stored.

Can I Upload My NIH Application to an AI Tool?

NIH's upload ban in NOT-OD-23-149 binds peer reviewers, not applicants (NIH, June 23, 2023). The firm data rule is NOT-OD-25-081: sharing controlled-access genomic data with public generative AI tools violates the Data Use Certification (NIH, March 28, 2025). Your own draft text falls under your account's training and retention terms, and NOT-OD-25-132 limits how much of the application AI may develop (NIH, July 17, 2025).

Write a One-Page AI Rule Before the Next Deadline

Put the decisions in writing once, so nobody makes them late at night before a submission. List the AI accounts your organization approves and whether training is off on each one.

Sort your standard grant documents into public, internal and restricted, and name the restricted ones: budgets with salaries, letters of support, unpublished results, controlled-access data and export-controlled material. Add the funders you apply to most often, with the date you last read their AI and confidentiality terms.

Recheck the provider pages when you review the rule, because their dates move. Then find the next announcement and read its terms before anything is uploaded.