Where your text goes when a model is used
The rules of a check run on our own servers without a model. A model is used only for the AI steps and only on the providers your workspace allows. It never has the final word.
The steps
- Upload. You upload a bilingual file, or a connector fetches it from your platform. The file is stored on our servers in the Netherlands and removed within 24 hours. A workspace can agree a longer keep for its uploaded files with us; we set it per workspace. An uploaded file is not in a backup, except an original kept for the review editor or under a workspace's agreed longer keep, which is in the nightly backup for 14 days like a report.
- Rules without a model. The deterministic checks run on our servers. Nothing leaves the service in this step.
- The provider allow-list. Before any text is sent, the service reads which providers your workspace allows (Settings, AI review). A provider that is not on the list receives nothing; if the list cannot be read, no provider is called. With no provider allowed, the check ends with the rules.
- The fenced prompt. The source and target are wrapped as untrusted data, with an instruction to assess them rather than obey them. A segment that contains instructions is content, not a command. This holds in the segment review and in the whole-file review alike.
- The answer filter. Before a model's comment or suggestion is stored, it is compared with the instructions that produced it. An answer that carries a marker every prompt contains, or a run of eight words copied from the prompt, is replaced with a neutral line. This is a mitigation with limits, not a guarantee: a paraphrase of our instructions shorter than eight matching words passes it.
- The finding. What the model proposes becomes a finding in your report, in your workspace. A reviewer confirms or overrides it; no report is final without a person.
- Retention. The report stays in your workspace until a workspace admin deletes it or until the workspace's retention setting removes it. A deleted report leaves our backups within 14 days.
The whole-file review, which reads a file's segments in order to find problems that span several of them, runs on paid plans when it is switched on. Every answer is checked against your file before it becomes a finding: the segment numbers must exist and the quoted text must be in the target it names.
What a provider receives
For the segment review, a provider receives one segment at a time:
- the segment's source and target;
- the neighbouring segments, as context;
- the segment's note from the file, when it has one;
- your profile's style guidance and instructions;
- the glossary entries whose source term is in the segment;
- past reviewer corrections on the same profile, as examples.
The AI check of rule findings and the second opinion receive the same kind of text for the segments they look at. A provider does not receive your file as a file. Other places that use a model read more. The whole-file review reads the segments of one file. Learn reads the documents you give it to build a profile. Eval sends the source text to the machine translation engines you enable.
All of these run on our own accounts with each provider, not on keys you supply. Your text is never used to train our models.
The providers
| Provider | What is sent | Keys | Training use | Region | When |
|---|---|---|---|---|---|
| Google Gemini | The segment text and context listed above; the segments of one file, in chunks, when the whole-file review is on | Ours | Not used to improve Google's products per the Gemini API paid-tier terms; covered by Google's data processing addendum | Global (Google) | Allowed for new workspaces; your admin can switch it off |
| OpenAI | The segment text and context listed above; the segments of one file, in chunks, when the whole-file review is on | Ours | Not used for training per the API terms | United States | Allowed for new workspaces; your admin can switch it off |
| DeepInfra | The segment text and context listed above, on open models; the segments of one file, in chunks, when the whole-file review is on | Ours | Not used for training per the provider's data page | United States | Allowed for new workspaces; your admin can switch it off |
| DeepSeek | The segment text and context listed above; the segments of one file, in chunks, when the whole-file review is on | Ours | Inputs may be used to improve the provider's models per its policy | China | Off for new workspaces; only if your admin allows it, and a workspace admin can switch it off again |
| DeepL | The source text of an Eval run | Ours | See the provider's terms | See the provider's terms | Only in Eval, only if your workspace enables it |
| Google Translate | The source text of an Eval run | Ours | See the provider's terms | See the provider's terms | Only in Eval, only if your workspace enables it |
| Microsoft Translator | The source text of an Eval run | Ours | See the provider's terms | See the provider's terms | Only in Eval, only if your workspace enables it |
The full list of services that handle data for us, AI providers included, is on Sub-processors.