Can you upload unpublished research to an AI tool? What happens to your files
Sometimes, but check three things first: whether the work is yours alone to share, whether a journal, funder or ethics rule forbids it, and what your specific account does with uploads. Personal ChatGPT content and Perplexity searches may be used for training unless you opt out. Do not upload a manuscript you are reviewing.
Three questions before you upload
Do you have permission to share it? Check with co-authors, your supervisor and the people responsible for the data. Do not assume that access to a draft gives you permission to send it to an external service.
Does a journal policy, funder condition, ethics approval or data agreement limit external processing? Then check the exact account and feature: training, retention and access are separate questions. A vendor's no-training promise is not permission to upload.
Upload privacy by account, with evidence gaps visible
This matrix uses only the supplied vendor-policy checks from 11 and 13 October 2026. An unconfirmed cell is a gap in that evidence set, not a claim that the vendor has no policy. Settings and contracts can change; check them again before sending unpublished material.
| Tool / account | Training and control | Retention / review limits |
|---|---|---|
| ChatGPT personal: Free / Plus / ProDossierSource 1Source 2 | OpenAI may use individual-service content to improve models; opt out at Settings > Data controls > Improve the model for everyone. New conversations after opt-out are not used for training. Feedback is an exception. | Temporary Chat: not used to improve models; may be kept up to 30 days for safety. General upload retention and human-review limits are not confirmed in this evidence set. |
| ChatGPT Business / Enterprise / Edu; OpenAI APIDossierSource 1Source 2 | Inputs and outputs are not used to improve models by default. | API: inputs and outputs retained up to 30 days; zero data retention (ZDR) on request for eligible endpoints with a qualifying use case. Do not transfer API limits to Edu or every file-upload feature. |
| Perplexity personal: Free / Pro / MaxDossierSource 1Source 2Source 3 | Searches may train by default. Opt out in Account Settings; the help page places AI Data Retention under Preferences. Opt-out applies only to subsequently collected data. | File-upload help: files stay private, content only customises responses; files retained for 30 days. General human-review limits are not confirmed here. |
| Perplexity Enterprise / APIDossierSource 1Source 2 | The API docs say no training by default, but certain APIs may be subject to product-specific terms. | Per the API docs, zero data retention is not the default; it is available only to eligible customers through an account-level configuration under a written agreement. File-upload help separately says 7 days for Enterprise Pro files; confirm feature scope. |
| SciSpace uploaded materialsSource 1Source 2 | Terms: no training of a general-purpose model unless separately disclosed and permitted by applicable rights, Source Terms and law. No training opt-out was found in the checked privacy policy. This is not an unconditional no-training guarantee. | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. |
| Gemini Notebook (formerly NotebookLM): personal Google accountDossierSource 1 | Google: content “will not be used to directly train our foundational AI models, unless you choose to provide feedback.” Feedback and included content (queries, uploads, responses) can be used to improve Google products and machine-learning technologies. | Feedback is “Reviewed by specially trained teams”; reviewed feedback and included content are “retained for up to 3 years, disconnected from your Google Account.” No retention period is stated for ordinary notebook content. |
| Gemini Notebook: Google Workspace or Workspace for Education accountDossierSource 1 | Google: uploads, queries and responses “will not be used to train AI models.” Check that your account really is a Workspace licence; a university email alone does not prove it. | Google: content “will not be reviewed by human reviewers even when you provide thumbs up or down feedback”. Work accounts fall under the Google Cloud Terms of Service, school accounts under the Workspace for Education Terms. |
| Claude: confirm the exact accountDossier | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. |
| Gemini: confirm the exact accountDossier | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. |
| Microsoft Copilot: confirm the exact accountDossier | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. |
| Elicit: confirm the exact accountDossier | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. |
| Consensus: confirm the exact accountDossier | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. |
| Scite: confirm the exact accountDossier | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. | Not confirmed in the supplied policy evidence. Check the vendor policy and your account before uploading. |
No training does not mean no storage or review
OpenAI says feedback can make the entire associated conversation available for training. Temporary Chat is not used to improve OpenAI models, but may be retained for up to 30 days for safety purposes. Do not apply that Temporary Chat limit to every uploaded file or ordinary conversation.
Perplexity's file-upload help says files stay private and their content is used only to customise responses. That page specifies 30 days of retention, or 7 days for Enterprise Pro. Per the API docs, zero data retention is not the default; it is available only to eligible customers through an account-level configuration under a written agreement. Those statements have different scopes; confirm the applicable file feature and contract rather than assuming every Enterprise upload has zero retention.
Perplexity says agreements prohibit third-party model providers from training on Perplexity data. That is distinct from Perplexity's own default use of consumer searches to improve its AI. Do not confuse a provider restriction with your account's opt-out setting.
Check the account, not just the university email address
OpenAI says ChatGPT Business, Enterprise, Edu and API inputs and outputs are not used to improve models by default. Perplexity's API docs say no training by default, but certain APIs may be subject to product-specific terms. These are named product commitments, not a guarantee that any account with a university email address has those protections.
Check the signed-in workspace, product name and institutional agreement. Ask your university's IT or data-protection team which service, data categories and features are approved. This guide does not name universities as license holders or label a tool GDPR- or Swiss-data-protection-compliant.
Peer review: keep the manuscript out of cloud AI
Do not upload a manuscript you are reviewing. First obtain the venue's current instructions and ask the editor what assistance is permitted. A personal training opt-out does not resolve confidentiality or the author's permission. Work from an offline review outline while you check the rules.
Interview transcripts and personal data need a separate decision
Do not upload raw interview, survey or patient data while permissions and processing terms are unresolved. Removing names alone may leave identifying details. Check the consent wording, ethics approval, data agreement and institutional instructions with the responsible team before using any external assistant.
For a thesis, an abstracted or synthetic example may be enough to ask about structure or method without sending the underlying data. This is a workflow precaution, not legal advice or a finding that a particular service satisfies local data-protection law.
A more cautious workflow
Check permission first, then use an approved account and set the training control before starting a new conversation. Send only what the task needs. Record the setting, account type, source policy and date so a later declaration is based on what you actually did.
For work that must stay on your computer, keep the corpus and analysis offline and check every sync, plug-in or export before enabling it. Zotero and ASReview dossiers are starting points for workflow planning, not proof that every configuration stays local.
Delete chats and files when you no longer need them, but do not treat deletion as undoing training. Perplexity says opt-outs apply only to data collected after opting out and previously collected training data cannot be deleted or removed. An opt-out is a forward-looking control.
Questions this guide answers
Every answer below is assembled from the dated fields on this page. Nothing is written separately for search.
Can I upload my thesis to ChatGPT?
Only after checking your permission, institutional rules and the specific account. For personal ChatGPT accounts, turn off Improve the model for everyone before starting a new conversation if you do not want it used for training. An opt-out does not resolve co-author rights, confidentiality or ethics conditions, and is not a no-storage promise.
Does ChatGPT Plus train on my data?
Personal ChatGPT Plus is covered by the consumer training setting, not the Business, Enterprise or Edu default. OpenAI says individual-service content may be used to train models and offers an opt-out in Data controls. It also says submitting feedback can make the entire associated conversation available for training.
Does Gemini Notebook (formerly NotebookLM) train on uploads?
Google renamed NotebookLM to Gemini Notebook on 16 July 2026. Its help page says personal-account content “will not be used to directly train our foundational AI models, unless you choose to provide feedback.” If you rate a response, that content can be human-reviewed and kept for up to 3 years. For Workspace and Workspace for Education accounts, Google says content is not human-reviewed and “will not be used to train AI models.” Checked 16 October 2026.
Is a university account different?
It can be, but the product matters more than the email address. OpenAI excludes ChatGPT Business, Enterprise, Edu and API inputs and outputs from training by default; Perplexity's API docs say no training by default, but certain APIs may be subject to product-specific terms. Confirm the actual workspace and agreement. Do not assume a personal account becomes institutional because you registered with a university email.
Can I upload a manuscript I am peer-reviewing?
Do not upload the manuscript to a cloud assistant. Obtain the current venue instructions and clarify permitted assistance with the editor first. A training opt-out or paid account does not itself establish permission or resolve confidentiality. Use a local review outline and the public record while those questions are unresolved.
Can I upload interview transcripts?
Not while permissions, consent and processing terms are unresolved. Check your ethics approval, data agreement and institutional instructions with the responsible team before using a cloud tool. Removing names alone may not remove identifying details. For method or writing questions, use an abstracted example instead of sending raw participant data.
Does deleting a chat undo training?
Do not assume it does. Perplexity explicitly says its opt-out applies only to data collected after the opt-out date and that previously collected training data cannot be deleted or removed. Training controls, deleting a visible conversation and vendor retention are different mechanisms. Set the control before sending the material, not after.
How can I keep research on my computer?
Use an offline workflow and check each sync, extension and export before enabling it. Keep the source corpus local and avoid sending raw files to a cloud assistant. Read the Zotero and ASReview dossiers when planning this workflow, but do not assume every configuration is offline or that a local tool removes your data-governance duties.