What we do with your text

We do not train on it

Your thesis, your sources and your prompts are never used to train a model — not ours, not anyone’s. We send text to our AI provider only to answer the request you made, under an agreement that forbids them training on it.

We also ask the provider not to keep it. Every request we make is marked zero-retention, so the provider does not store your text or its own reply once the answer has been returned — not for their dashboard, not for review, not for thirty days. There is nothing on their side for us, or anyone else, to go back and read.

What is sent, and when

Only when you ask for something. A suggestion sends the paragraph you are in, your outline and glossary, and passages from the sources you pinned. A draft sends the section’s scope note and its passages. Nothing is sent while you type, unless you switch on suggest-without-asking, and then only after you pause.

The Chrome add-on

It reads a page only when you click its button (or choose it from the right-click menu), and then only the page’s address and the tags publishers put on it to describe the article — its title, authors, journal, year and DOI. On a PubMed, arXiv or Google Scholar results page it reads the titles and identifiers of the results shown, nothing more. Never the page’s text, never another tab, never your browsing history. When you choose Save, it sends those references to your thesis, using the sign-in you already have on this site. If the tab is showing a PDF and you leave Attach this PDF ticked, it downloads that file from the site you are reading — as your own request, with your access to that site — and uploads it to your library, where it is stored like any PDF you upload. It sends nothing anywhere else and has no analytics. What it keeps in your browser is which thesis and collection you saved to last, and for at most two minutes a link you right-clicked.

From version 0.3.0, on Google Scholar, PubMed, arXiv and MDPI pages only, it also puts an Add to Thesis Copilot button into the page. To do that it reads the same tags and results as the page loads, inside your browser; it sends nothing until you press a button. Then it sends the paper’s identifier (its DOI, arXiv id or PubMed id) to look it up and, when you choose Save, to add it — or, for a search result that has no identifier, its title, authors and year as shown. It never contacts those sites itself. From version 0.3.1 it also keeps, in your browser only, the identifiers of the papers you saved with those buttons, so a page you come back to shows them as saved; that list is never sent anywhere.

From version 0.4.0 the buttons are on every site from install, so Chrome asks, when you add it, for access to all websites. The add-on uses that access only to read, inside your browser, the tags that describe an article and the DOI links in a page’s reference list (on Wikipedia, for example), to put a Save button beside them — never the rest of the page, and never on this site. It still sends nothing until you press a button, and then only that paper’s DOI or details (or, when you tick several, theirs). You can turn the buttons off for every site, or for one, in the add-on’s settings; that choice, and your theme, stay in your browser. Its window shows the name and email of the account you are signed in with, read from this site; they are not stored or sent anywhere.

From version 0.4.1, when you save an article from its button and the article’s page names its own PDF, the add-on downloads that PDF from the same site — as your own request, with your access to that site, at your press — and attaches it to the paper in your library, as Attach this PDF does. It does this only for a PDF on the very site of the page you are on, and only for a paper your library did not already have.

What we log

Every AI call is recorded with its model, its token counts, its cost and how long it took — this is how the monthly allowance is counted and how we keep the price honest. The request and response bodies are not logged. Our application logs never contain your thesis text.

Who can see it

You. Someone you share a document with sees only that document, in the role you chose: reading, reading and commenting, or — if you make them a co-author — editing with you. If you turn on a read-only link for a document, anyone who has that link can read its text (not your sources, comments or email) until you turn the link off. Our administrators can see your email, your plan, your usage and your document titles — the things needed to run the service and answer support. When needed for support or to prevent misuse, an administrator can also open your thesis to read it, never to change it. Every time that happens it is recorded, and we email you straight away to say so.

When someone comments or replies on a thesis, we email the others who can see its comments (never a reader) the commenter’s name, the thesis title and the first few lines of the comment — never the thesis text — at most once an hour per comment. Turn this off under Account or with the link in any of those emails.

Where it lives

On our own server, in a Postgres database and an object store we run. Files are encrypted in transit. Backups are taken nightly and kept for 30 days.

Leaving

Export any chapter as .docx or PDF at any time, with or without a subscription — do that first, because the rest of this is final.

Delete your account from the account page. You are signed out everywhere immediately, and seven days later your theses, chapters, sources, uploaded files, exports and comments are erased outright — not hidden, not archived. The seven days are yours to change your mind in; nothing has happened until they are up. Backups age out within thirty days of that.

Two things survive: your payment records, which we are required to keep, and a note that the deletion was asked for and carried out. Neither contains anything you wrote.

Integrity

We do not build “humanising” or detector-evasion features and we will not. What the AI wrote stays marked as such in your document, and the AI-usage log is yours to export and disclose. See the pricing page for the full position.