Threat model
What Proselike protects against, and what it doesn't. Read this before you rely on it.
Assets
- The persona ↔ person link. If someone learns that pseudonym X is real person Y, the harm can be total: retaliation, job loss, legal exposure. We treat a breach of this link as the existential risk.
- Your writing, both drafts and published work.
- Your account and payments.
Adversaries we design for
| Adversary | Capability | Our mitigation |
|---|---|---|
| Stylometric analyst. Compares your pseudonymous posts with your real-name writing. | Off-the-shelf authorship tools, many samples of both. | The judge measures the same signals they would. The loop pushes distance from Real Voice up and keeps model-likeness low, so you don't trade one fingerprint for the "written by AI" one. Tells show you which habits leak. |
| Platform or employer with a pile of suspect texts. | Clusters writing by style and checks whether the persona is internally consistent. | Persona consistency keeps the pseudonym a single, stable, human-sounding author. A pseudonym that drifts week to week stands out. |
| Breach of our database. | Full copy of our DB. | The DB has no Real Voice data at all. Account numbers are stored as salted hashes, with no email unless you added one. An attacker learns persona names, specs and published outputs: text you had already made public. |
| Legal demand served on us. | Compels us to produce records. | We can produce only what we hold. There's no Real Voice, no originals, and no account-to-identity link unless you gave us an email or paid by card. Pay with crypto and skip the email if this matters to you. |
| Network observer. | Sees traffic to our domain. | TLS everywhere, HSTS. They can see that you use Proselike, but not what you write. If being seen to use it is itself a risk, use Tor. |
What we do not protect against
- Content. No style tool hides what you say. If only five people knew a fact, publishing it narrows the field to five, however it's written.
- Metadata. Posting times, time zone, IP address, account-creation details, payment trails and cross-account follows all identify people more often than prose does.
- A compromised device. Your Real Voice fingerprint is in your browser's localStorage. Anyone with your unlocked browser has it.
- The model provider. Paragraphs being rewritten go to the configured frontier-model API, under that provider's data terms. Don't paste anything that would be fatal in a provider breach. For Lane C users at high risk, a self-hosted model deployment is on the roadmap.
- Guarantees. Stylometry is probabilistic. A high "distance" score means our judge, and judges like it, find no strong link. It doesn't mean no one ever could. Our published error rates are in Benchmarks.
Design decisions that follow from this
- Drift-from-me personas. The stored spec never contains your real-voice markers. Signature openers are steered away from on each request, using the Real Voice fingerprint your browser sends. A drift persona is stored with the same method label as a crafted one. One residual risk remains: a drift spec is, by construction, anti-correlated with your habits ("never uses em-dashes" because you love them). Someone holding our database and already suspecting you could treat that as weak corroboration, though it isn't evidence on its own. If that matters to you, use Craft, or edit the spec after creation.
- Real Voice fingerprints are computed server-side but never persisted. The next step is moving the encoder into the browser, so samples never leave your device at all. Our encoder is linear plus character n-grams, which is small enough to ship to a browser.
- No bulk or multi-account API in v1, and per-IP rate limits, so we aren't a sockpuppet factory.
- Request bodies are never logged.
- Credit codes for anonymize are bearer tokens with no account, revealed once.
Abuse we refuse to enable
Impersonating real, named people. Modelling a third party's voice without their consent (Voice fidelity requires a consent attestation). Running many accounts for one operator (sockpuppetry). Evading platform bans. See Terms.