Responsible AI
AI that defends the writer, not the other way around.
We exist because an AI system treated non-native writers unfairly. That makes the way we build our own AI a matter of principle, not a press release. Here is how we use it — and the lines we won’t cross.
OUR PRINCIPLES
Six commitments we build against
These aren’t aspirations for a future version. They shape what ships.
No training on your writing
Your drafts are yours. We don’t use customer writing to train models, and we don’t sell it. The text you write to get work done isn’t raw material for our next release.
Consent-first by default
Features that record or process your work — like the Authorship Certificate edit chain — are opt-in, with plain-language explanations of what’s captured and why. Off is a real default, not a buried setting.
On-device where we can
When a task can run locally, we prefer to keep it on your machine. Less data in transit is less data at risk. Where a model has to run server-side, we minimize what we send and how long we keep it.
Human-in-control
Our tools suggest; you decide. Every AI edit is shown as a reviewable change you accept or reject — never a silent rewrite. The work stays yours, and provably so.
Transparency about limits
We say what our tools can and can’t do. The Authorship Certificate documents your process; it does not prove innocence or guarantee any detector result. We won’t market a capability we can’t stand behind.
Data minimization & retention
We collect what a feature needs and keep it only as long as it’s useful, then delete it. Append-only authorship records are the deliberate exception, retained so they stay verifiable — and only when you opt in.
A DIFFERENT DEFAULT
The common pattern, and what we do instead
Plenty of AI tools train on whatever users type. We made the opposite choices — not as a swipe at anyone, but because our users have already been on the wrong end of careless defaults.
| The common default | How Writence does it |
|---|---|
| Train models on whatever users type, by default. | No training on customer writing. Opt-in is required for anything beyond running the feature you asked for. |
| Bury consent in a terms-of-service update. | Plain-language, in-product consent for recording or processing — and an honest default of “off.” |
| Send everything to the cloud because it’s easier. | On-device where the task allows; minimal payloads and short retention where it doesn’t. |
| Rewrite the user’s text automatically. | Suggest as a reviewable diff. The human accepts or rejects every change. |
We’re describing patterns, not naming companies. The point is the standard we hold ourselves to — not a comparison scoreboard.
WHAT WE WON’T SAY
Honest about what AI can’t do
Responsible AI includes refusing to overclaim. The Authorship Certificate documents how a piece of writing was made; it does not prove innocence and it cannot guarantee how any third-party detector will score your text. We won’t tell you our tools make you “undetectable,” because that would be both untrue and beside the point.
Read why detectors are the wrong test →- We document process — we don’t promise verdicts.
- We surface suggestions — we don’t rewrite silently.
- We minimize data — we don’t hoard it for later.
- We say where models run and what we keep.
- We make consent legible, not buried.
Principles you
can actually check.
See how these commitments show up in our security posture and the products we ship.
