AI and your content
Exactly which actions send your content to a model, what is sent, which models run, and what is kept afterwards.
ShareMyPage has AI features, and AI features mean your content leaves our infrastructure and reaches a model provider. This page says precisely when that happens and what is sent, because "we use AI responsibly" is not an answer anyone can assess.
Hosting a page never calls a model
Uploading, storing, sharing, viewing, commenting on, exporting and password-protecting a page involve no model call of any kind. If you never use an AI feature, no content you put in ShareMyPage is ever sent to a model provider.
This matters for the common case: if you are publishing an artifact an AI already wrote somewhere else, ShareMyPage is only the host.
The one exception is "Ask this page", below, which an author switches on for a single page and which is off everywhere until they do. Viewing such a page still calls nothing; only asking it a question does.
The actions that do call a model
Each one is something somebody explicitly turns on or triggers. All but the last are slide-deck features used by the author of the deck.
| Action | What is sent |
|---|---|
| Ask this page | The page's text and the reader's question |
| Create a deck with AI | Your written brief |
| Edit a slide with AI | Your instruction plus that slide's current HTML |
| Change a slide's layout | The slide's current content and the target layout |
| Suggest a theme | Your description, or the brand colours you supplied |
| Import a PowerPoint (re-author mode) | Each slide's parsed text and the URLs of its hosted images |
PowerPoint import re-authors by default
Re-author is the default import mode, and it hands every slide's parsed content to a model in order to lay it out well. If you want a PowerPoint imported without any model seeing it, choose faithful mode, which is fully deterministic.
Only the slide, deck or page being acted on is sent. The model does not receive your other pages, your other decks, your workspace, your members, or your comments.
Ask this page is the one a READER can trigger
Every other action on this list is started by the author, on their own deck. "Ask this page" is started by whoever opens the link, so it is off for every page until the page's author turns it on, one page at a time, in the page's settings. A workspace with AI switched off cannot use it at all, whatever any page says.
It is per page rather than per workspace on purpose. A country-by-country regulatory reference is public knowledge that is merely hard to read, and answering questions about it is the entire point. A strategy document, or a file of customer feedback whose authors agreed to nothing, is the opposite. Neither we nor a workspace admin can tell those apart from the outside. The person who uploaded it can.
The answer is drawn only from the page, and every quote supporting it is checked against the page's own text before you see it: a quote the model produced that is not in the document is discarded, and an answer left with no real quotes is shown as unsupported rather than as an answer. Questions and answers are not stored in our database and not sent to our analytics; what we record is whether the page could answer, how many quotes survived checking, and how long it took.
Which models, and where
Model calls are routed through the Vercel AI Gateway to Anthropic:
anthropic/claude-sonnet-5for text edits, deck outlines, layout planning, themes and PowerPoint re-authoringanthropic/claude-opus-4-8for edits that require the model to look at a rendered slide
Inference happens in the United States, covered by the EU Standard Contractual Clauses. This is the one part of the product where content leaves the EU. Everything else, including where the content is stored at rest, stays in Frankfurt. Both providers are listed on the subprocessors page.
Training
We do not use your content to train or fine-tune any model. We do not build models, and there is no code path that exports customer content for that purpose.
Retention at the model provider is governed by their API terms rather than by us. We can share the current terms and our agreement with a reviewer on request.
What we keep
Prompts are not stored in our database. The instruction you type is sent to the model, the result is applied to your slide, and the instruction itself is not written to a table.
Prompt text does reach our product analytics
The instruction you type into an AI field is captured as a deck_ai_prompt
event in PostHog (EU, Frankfurt) along with the outcome, so we can see which
instructions fail and fix them.
That means the text of your AI instructions is retained in our analytics, even though the slide content it operated on is not. If your instructions would themselves contain something sensitive, treat that as a disclosure. We are reviewing whether this event should be truncated or dropped.
Slide content sent to the model is not separately logged by us. What we do record per call is metadata: which feature ran, which model, latency, step count, and an outcome or error code.
If you want none of this
A workspace admin can turn AI off for the whole workspace, in Settings, General, AI features. With it off, none of the actions in the table above will run for anybody in the workspace, through the app or over MCP, and no content from it reaches a model provider at all.
It is on by default, because it governs features that already existed when the switch was added and turning them off underneath people would be a change we made to their product rather than one they chose. Any AI capability added after this switch is opt-in separately, so a new surface never arrives already enabled.
The switch is enforced at every entry point that can reach a model, and a test fails the build if a new one is added without it.
See also
- Subprocessors
- For security reviewers
- Creating decks with AI: the product-side guide.