Use of AI

Who this covers

INT6 Ltd, a company registered in Scotland, number SC547705, trading as INT6, which builds and holds InMyWords. It covers every use of a language model or speech service in the application. It is reviewed once a year and whenever a model or provider changes.

Where models are used

1. Translation. Each turn of a conversation is translated by a large language model, using what the conversation has carried forward so far.
2. Scoring. A second model call scores each translation for how faithfully it carries what was said, and the score is shown beside it.
3. Clarifying questions. Where the meaning is unclear, the model proposes a question to ask rather than a guess.
4. Facts and summaries. A model records facts said in a conversation and, where an organisation switches it on, writes a review or summary of it.
5. Documents. A letter or form sent for translation is translated by the same route.
6. Speech. Where an organisation switches speech on, words spoken are turned into text, and text is read aloud, by speech services.

What is sent, and what is not

1. Sent: the words of a turn or document, the languages, and the context the conversation carries forward, such as formality and how to address the person.
2. Not sent: passwords, payment details, or anything from another conversation.
3. Through our own gateway. Model calls go through a gateway on our server, which holds the provider credentials; the browser never talks to a model.
4. Speech stays on our server by default. Speech is sent to an outside speech service only where our administrators have switched that on for an organisation and the organisation has itself accepted that service's data policy.

The providers

The model provider is a setting, not something built into the software, so it can be changed without a new release. We tell an organisation which providers handle its conversations before its service starts. We do not use conversations to train or fine-tune any model.

A person stays responsible

1. The tool does not decide anything. It translates between two people. What is decided in the conversation is decided by them.
2. A low score is a warning. A translation that scores badly is marked, so the member of staff can ask again or check another way.
3. The scorer is a model too. It can be wrong in the same direction as the translation it scores. That limit, and the others, are set out on known limits.
4. Where it should not be used. Safeguards says where an interpreter, not this tool, is the right answer.