Skip to content

What we send to AI (disclosure facts)

Technical version: AI data egress — the disclosure source of truth

This is the plain twin of the disclosure source-of-truth doc. It is the same honest story as What we send to AI, and where it goes, kept under the matching basename so the two-track rule does not invent a silent exception for “disclosure” vs “egress.”

Not legal terms. Legal review still has to turn these facts into ToS and privacy copy. Until that review lands, treat this as engineering honesty for humans, not published policy.

The one fact that must not blur

Edition Whose AI account What that means for you
Desktop Yours — your API key or your signed-in command-line AI tool WorkWingman packs the request on your machine. It does not route the prompt through a WorkWingman server account. Your provider’s own retention and training rules still apply.
Cloud / pilot Ours (and sometimes tenant-supplied keys for some features) We are in the middle. Responsibility for vendor choice and retention is ours, and the exact vendor list depends on how that cloud instance is configured.

“Your own key” does not mean “private forever.” Consumer plans sometimes train or retain; that is between you and the vendor.

What actually leaves (high level)

For writing tasks, the model typically needs career facts: name (preferred or legal), history, education, skills, stories, writing samples, the job description, and sometimes federal extras such as supervisor contacts. Your email, phone, and postal address are not stuffed into the drafting prompt — the app puts those into the finished document afterward.

Media features send whatever you type as an image, speech, or music prompt — and some image tools send the image file itself.

We do not claim the content is anonymized before the model sees it. A resume without who you are is not a resume. What we do is pick the fields for the task, fence untrusted third-party text where that fencing is wired, refuse invented employers/degrees/dates on the main draft path, and keep usage ledgers free of your document text.

Still open on cloud

How long cloud providers keep what we send, and whether they train on it, is not something we invent from memory. Until commercial terms are confirmed in writing, user-facing promises about retention and training stay out of published copy.

Where engineers check the list

The checkable inventory of “paths that leave the machine” lives in code as an egress surface registry. Treat it as a floor, not a perfect census — missing entries have been found before. Disclosure copy should track that inventory rather than freestyle marketing language.