Security & Data Handling

• unwrite has no model-training pipeline of its own. It does not train or fine-tune any LLM on your writing. Your material is used only to produce your drafts and run the features you invoke.
• The LLM being used is Anthropic’s Claude, accessed through Anthropic’s commercial API, under a contract that forbids training on your content. Anthropic’s Commercial Terms of Service state, in the Customer Content section: “Anthropic may not train models on Customer Content from Services.” Under those same terms, you retain all rights to your Inputs and own the Outputs, and your content is treated as your Confidential Information. (Source: Anthropic Commercial Terms of Service.)
• Your content stays yours end to end. unwrite claims no ownership of your sources, briefs, voice samples, or drafts, and never sells your data.

• Your data serves and is accessible only to your account. unwrite uses what you bring to help you. It is never pooled into a shared model, a shared dataset, or a common pattern library across customers.
• Never shared with other unwrite users. One customer’s sources, voice, drafts, or usage never inform another customer’s output. unwrite Teams plans share material only inside the one organization that owns the workspace, and only with the roles that organization grants.
• Never fed to Anthropic’s training. Your content passes through Anthropic’s API only to generate your draft, under the no-training term quoted above.
• The direction of travel. The value of unwrite grows as it fits your way of working. Any personalization we build stays scoped to your own account. Your writing and the material you bring are never pooled across customers and are never used to train any model, now or in the future.

• Encrypted in transit. Every connection uses HTTPS, and the site is served with HSTS, so browsers refuse to load it over an unencrypted connection.
• Stored in the United States. Your account and saved work live in a managed Postgres database (Neon) in a US region, encrypted at rest.
• Hardened sign-in. Sign-in is a Google account, an email and password, or a one-time email link. Passwords are stored only as a salted argon2id hash, never in the clear. Sessions live in a secure, HTTP-only cookie, so the session token is never exposed to page scripts.
• Scoped, isolated access. On the Teams plan, access is role-based (administrator, member, and review-only reviewer roles), and each organization’s workspace and material are isolated to that organization.
• Limited internal access. Access to stored content in production is restricted to authorized personnel for support and operations only, is never used for any other purpose, and is never shared.
• Analytics never see your writing. Product analytics capture how features are used; the text of what you type is stripped before any analytics event is sent, and screen recording is off.
• Audited infrastructure underneath. unwrite runs on cloud providers (hosting, database, payments) that maintain their own independent certifications such as SOC 2 and ISO 27001. Those certifications belong to the infrastructure vendors; unwrite builds on top of them. unwrite does not yet hold its own SOC 2 or ISO 27001 certification.

• Purpose limitation. The material you bring (sources, briefs, voice samples, uploaded documents, and drafts) is used only to generate your drafts and run the features you trigger. It is not repurposed.
• Drafts from your sources. Drafts are written only from the sources tied to your piece: the material you provide, plus sources unwrite gathers and cites when you ask it to research a topic. Researched sources show an authority rating, and unwrite’s built-in Source Check verifies each checkable claim against that source set and flags the result in the Studio. unwrite never invents facts, and never draws on any other customer’s material.
• You control retention and deletion. Delete a piece or a voice, and it is held for 30 days so you can restore it from Settings, then erased. Delete your account, and everything goes immediately, including anything in Recently deleted. You can export any piece at any time, or ask us to access, correct, export, or delete your data by writing to hello@unwrite.ai.

unwrite is a multi-tenant cloud service, not an isolated, single-tenant, or on-premises deployment.

• Enforced isolation. Every user request is authenticated, and every read and write is scoped to the owning account or organization. You never see another customer’s material, and they never see yours.
• Shared, managed infrastructure means it is patched and audited for you. You are not responsible for hardening a server, applying security updates, or maintaining the database. That work is done continuously on independently certified infrastructure, which for most teams is more secure than a machine they would have to run themselves.
• Encryption and access control travel with the data, in transit and at rest, regardless of tenancy.
• Isolated and embedded options for enterprises. If you need a single-tenant or isolated deployment, or an embedded, headless integration that runs unwrite inside your own tools (for example, Word, Google Docs, or Salesforce), these will be available under an enterprise agreement and scoped per engagement. They are not part of the standard multi-tenant service today.

AI-assisted output may carry a provenance mark, an industry-wide content-authenticity signal, consistent with unwrite’s stance as an authenticity tool rather than a concealment tool. These marks are created at the model level and have nothing to do with unwrite’s ability to create quality content. Moreover, because unwrite is not an AI-bypass tool (meaning unwrite is not in the business of creating content designed to fool AI content checkers), we support these marks and encourage users to use unwrite as intended: as a drafting tool that helps writers rather than replacing them.

unwrite relies on a small, named set of subprocessors, each receiving only what it needs: Anthropic (large language models; API content not used to train their models), Neon (database, US region, encrypted at rest), Stripe (payments; unwrite never stores full card numbers), Resend (transactional email), MailerLite (opted-in product-update email; address only), Google (optional single sign-on), Upstash (rate limiting / abuse protection), PostHog (product analytics, usage only), Sentry (error monitoring), and Slack (internal support routing; form contents only, never your writing). The complete, current list is in the Privacy Policy.

• unwrite operates as a commercial customer under Anthropic’s Commercial Terms of Service and Usage Policy (commercial, not consumer, terms).
• No training on API content is a contractual term (quoted above).
• Limited retention, with a zero-retention option. By default, content sent to the Anthropic API is automatically deleted within 30 days, and retained data is not used for training without express permission. A Zero Data Retention arrangement is available to eligible accounts on request. (Sources: Anthropic API and data retention; Anthropic Trust Center.)
• Anthropic publishes its own posture. Anthropic’s security certifications, subprocessor list, and data-handling commitments are at its Trust Center, which you can review directly to verify Anthropic’s side of the chain.
• A Data Processing Addendum governs the processing. Anthropic’s Commercial Terms incorporate a DPA covering how customer data is handled, and unwrite has entered into it.