Legal
Privacy policy
What Carrel collects, why, where it is stored, who else touches it, and how to get it back or delete it.
Last updated
This policy explains what Carrel does with your information. It is written to be read rather than to be survived, and it describes how the product actually works today.
Carrel Inc. (“Carrel”, “we”) is the controller of the personal data described here. If you are using Carrel through your university or department, that institution may also have its own agreement with us; where it does, and the two differ, that agreement governs.
Questions about anything below go to privacy@carrel-ai.com.
The short version
- Your papers, highlights, notes and drafts belong to you. We do not sell them, publish them, or use them to train AI models.
- Your library is stored in the European Union, in Frankfurt.
- Parts of your content are sent to an AI provider when you ask for an answer, an extraction or a draft — because that is how those features work. That provider does not train on it.
- Only you and the people you invite to a project can read what is in it.
- You can export everything or delete it at any time.
What we collect
Your account. Your email address, your name if you give one, and a password hash — or, if you sign in with Google, the email address and profile name Google returns. We never see your Google password.
Your workspace. The organisations and projects you create, who is in them, and what role each person has.
What you put into Carrel. The PDFs you upload, the papers you import from a DOI, an arXiv identifier or Zotero, and everything the product derives from them: extracted text passages, embeddings, claims, methods and findings, the relations between them, and gap candidates. Also everything you write yourself: highlights, notes, questions to the research assistant and the answers it gave, synthesis sections, manuscripts and drafts.
Billing details. If you subscribe, our payment processor holds your card; we hold your plan, your subscription status, your billing period, seat counts and invoice history. We never receive or store your card number.
Usage data. Which features are used, how often, and whether they succeeded. This is deliberately narrow: our analytics events carry counts, kinds, plan names and booleans, never the titles of your papers, your questions, passage text, free text, or the identifiers of your own objects. That restriction is enforced in the code, not just in policy.
How the app is used on screen. To find where the product is confusing or broken, we record interaction with the app itself: where on a page people click and how far they scroll, and a replay of a session’s layout and pointer movement. The text on the page is replaced with asterisks in your browser before anything is sent, as is everything you type, and the pages of your PDFs are not recorded at all. What reaches us shows that a button was clicked or a panel opened, not what a paper, a note or a draft says.
Technical data. IP address, browser and device type, and timestamps, as part of ordinary request logs and for security.
Correspondence. If you email us or write in our community server, we keep what you send so we can answer it.
Why we use it, and on what basis
| What for | Legal basis |
|---|---|
| Running the product you asked for — storing your papers, answering questions, drafting sections | Performance of a contract |
| Billing, invoices and tax records | Contract, and legal obligation |
| Keeping accounts secure and preventing abuse | Legitimate interests |
| Understanding which features work, and fixing what does not | Legitimate interests |
| Service emails you cannot opt out of — receipts, security notices, an invitation someone sent you | Contract |
| Product news and marketing email | Consent, withdrawable in one click |
Your papers and AI
This is the part most researchers want to be clear about, so it gets its own section.
When you ask the research assistant a question, extract claims from a paper, build the graph, or draft a section, the relevant passages from your own library are sent to our AI provider’s API along with the instruction, and the result comes back to you. That is unavoidable: it is the feature.
What we commit to:
- We do not use your content to train models. Not ours, not anyone’s.
- Our AI provider does not train on it either. We use their API under terms that exclude API data from model training.
- Nothing is shared with other users. A passage from your library is never used to answer somebody else’s question.
- Only what is needed goes. Answers are grounded in passages retrieved from your own project, not your entire library, and not other people’s.
We do not sell personal data, and we do not share it with advertisers. If we are ever compelled by law to disclose data about you, we will tell you unless we are forbidden from doing so.
Who else touches your data
We use a small number of providers to run Carrel. Each one only gets what it needs to do its job.
| Provider | What it does | Where |
|---|---|---|
| Supabase | Database, authentication and file storage — your papers and PDFs live here, and account email such as confirmations and password resets is sent from here | Frankfurt, EU |
| Vercel | Hosts the website and the app | EU region for functions |
| Render | Runs the ingestion worker that parses PDFs, and the collaboration server | Frankfurt, EU |
| OpenAI | The AI behind answers, extraction, relations and drafting | United States |
| Stripe | Payments and subscriptions | United States / EU |
| PostHog | Product analytics, heatmaps and masked session replay, within the limits described above | United States |
| Resend | Product news and announcements, if you opt in | United States |
| Optional sign-in, if you choose it | United States |
When you import metadata, we look a paper up at public scholarly services — Crossref, OpenAlex, Semantic Scholar, arXiv and doi.org. They receive the identifier or the title being looked up. They do not receive your identity or your library.
Where a provider is outside the European Economic Area, transfers are made under the European Commission’s standard contractual clauses.
Where your data is stored
The database, the file storage and the server functions that read them are in Frankfurt, in the European Union. That is a deliberate choice, and the infrastructure is pinned to it.
Analytics and email are the exceptions in the table above, and neither carries the content of your research.
Who can see what is in your projects
Access is enforced in the database itself, not only in the application. Every row and every file belongs to an organisation and a project, and a query that is not made by a member of that project returns nothing. PDFs are kept in private storage and served through links that expire.
Inside Carrel, a project is visible to the people you invite to it, with the role you give them. Supervisors see what is shared with them, not your whole account.
Access to production systems is restricted to the people who maintain them, used only when something is broken and there is no other way to fix it, and not used for anything else.
How long we keep it
We keep your content for as long as your account exists, because it is what you came for.
When you delete a paper, a project or your whole account, the rows and the files are removed from the live database and from storage. Our database provider holds encrypted backups on a rolling schedule; deleted content disappears from those as the backups age out.
Billing and tax records are kept for as long as the law requires us to keep them.
Your rights
Wherever you are, you can ask us to give you a copy of your data, correct it, or delete it — and you can do the first and the last yourself, from inside Carrel, without asking.
If you are in the European Economic Area or the United Kingdom, you also have the right to object to or restrict processing, to withdraw consent, and to complain to your data protection authority. If you are in California, you have the right to know what we collect, to delete it, and not to be discriminated against for exercising either — and we do not sell or share personal information as those terms are defined there.
To exercise any of these, write to privacy@carrel-ai.com. We answer within 30 days.
Cookies
Carrel uses very few, and none of them are advertising cookies.
- Session cookies keep you signed in. Without them the product cannot work.
- A project cookie remembers which project you had open, so the app does not drop you somewhere else when you come back.
- Analytics cookies record the usage described above.
You can clear or block them in your browser. If you block the session cookies, you will not be able to sign in.
Security
Everything is encrypted in transit and at rest. Access between services uses scoped credentials rather than shared ones, secrets are never committed to our code, and per-workspace isolation is enforced at the database level so a bug in the application cannot hand one customer another’s papers.
No system is perfect. If we discover a breach affecting your personal data, we will tell you and the relevant authority without undue delay.
Children
Carrel is built for researchers at universities and is not intended for anyone under 16. We do not knowingly collect data from children. If you believe a child has given us personal data, write to us and we will delete it.
Changes
If we change this policy in a way that matters, we will say so in the product and by email before it takes effect, and the date at the top will change. Older versions are available on request.
Contact
privacy@carrel-ai.com for anything in this document, including requests for a copy of your data, deletion, or a data processing agreement for your institution.