What this is bad at

What the checking does not give you

The product assumes it will sometimes be wrong, and the job that follows from that is showing you when. Being clear about what the scoring gives you means being just as clear about its limits.

1. A score is not a guarantee. It describes the check that was actually run on that turn, and nothing about the world beyond it.
2. The scorer is a language model, like the translator. It offers one model's opinion of another model's work, and it can be wrong in the same direction as the translation it is scoring.
3. A high score on a turn does not make the conversation appropriate for the tool. That judgement is yours, and the list of conversations that are never in scope on the safeguards page is where it starts.
4. Where something about the person has been picked up from the conversation rather than told to the system, it is visible to the worker and can be corrected. Age band and gender change how a translation refers to someone, so an assumption left uncorrected is an error that carries into every following turn.

Known limits

1. Language coverage is uneven. Quality is not the same across every seeded language, and the languages with the least training data are often the ones a frontline service most needs.
2. Dialect and register vary. A regional variety may be handled less well than the standard form of the same language.
3. It is written, not spoken. Today both people type. That suits a desk and a phone; it is slower than talking, and it assumes both people can read and write their own language. Speech in and out is being built and will bring its own limits with it.
4. Scoring is itself a model output. A fidelity score is one model's opinion of another model's work. It can be wrong in the same direction as the translation it is scoring.
5. It is actively developed. Behaviour changes between releases. That is how the product improves, and it means a check you ran six months ago is not a check you have run today.

Data

Conversation content has to be sent to a hosted language model to be translated, and is stored so the exchange can be reviewed and audited. For a service handling sensitive disclosures that matters, so it is written out in the privacy notice.

Where an organisation has constraints that this does not satisfy, say so early. Deployment options are an open question we are working on, and we would rather hear the constraint than argue with it.

This page describes the product's current limits. If you believe something here is inaccurate, or you have hit a limit that is not listed, tell us. That is more useful to the project than a compliment.