Safeguarding the person you are speaking with

How a translation is produced and checked

Safe use starts with knowing what happens to a turn before the other person reads it. Nothing here is a setting you have to switch on. It happens on every turn, every time.

1. It translates in context. The turn goes to a language model with the conversation so far, the facts recorded about the situation, and your domain vocabulary in scope. It is not translated as an isolated sentence.
2. It assesses its own work. A second, separate call scores that translation for fidelity against the source: does it carry the same meaning, the same register, the same specifics. The score is a number, and it is attached to that turn.
3. A poor score is acted on automatically. If the score falls below the configured threshold, the tool rewrites the translation before the other person ever sees it. Both versions stay in the record. Nobody has to notice the problem for it to be fixed.
4. Where it cannot be confident, it asks instead of guessing. If the source itself cannot be translated confidently, an unfinished sentence, or a term that could mean two very different things, the tool puts a short, specific question back to the speaker in their own language and waits. It does not pick an interpretation and carry on.

The worker sees a short note beside the turn saying what that check found. Where no score exists for a turn, the space is empty: an unchecked turn is never drawn as a checked one.

A tool that scores its own work and puts a question back when it cannot be sure gives you something to act on while the conversation is still live. That is what the rest of this page assumes you have in front of you. When it is not sure goes through the two mechanisms in more detail.

Where to stop

This tool is built for the straightforward conversations: the twenty minutes at a front door that would otherwise be postponed for want of a shared language. Handling those here is what keeps interpreter time available for the conversations that genuinely need it. My Words is not intended to take the place of a qualified human interpreter in formal, legal, clinical or high-risk situations, and the categories below are never in scope whatever the tool appears to be managing.

1. Statutory assessments and decisionsHomelessness duty decisions, care assessments, capacity assessments, anything where the record is the basis of a legal duty.
2. Legal proceedings and adviceCourt, tribunal, police interview, formal legal advice, anything that will be tested as evidence.
3. Clinical consultation and consentDiagnosis, treatment options, consent to a procedure, medication instructions.
4. Formal safeguarding interviewsWhere the account itself is being taken and will be relied on.
5. Any conversation where an error would be hard to reverseIf a misunderstanding could not be caught and corrected in the next few minutes, this is not the tool.

Context the system picks up, including risk factors

As a conversation runs, My Words extracts structured facts from what is said and carries them into the turns that follow. Each one is stored with the category it belongs to and a confidence figure, and can be pinned, corrected or superseded as the picture changes.

A good deal of what surfaces this way is risk-relevant, because it is what people actually talk about at a front door: mental health, rent arrears, tenancy status, a benefits application that has stalled, who is in the household, the age of a child. It arrives without anyone running an assessment, which is the point. Somebody explaining their situation in their second language should not also have to know which detail was the important one.

How that is read matters as much as what it catches, so there are rules about it.

1. It surfaces. You assess. An extracted fact is a prompt to your professional judgment. It is never a finding, a score against the person, or a safeguarding assessment, and nothing it produces belongs in a case record as though it were an assessment you carried out.
2. An empty panel means nothing. Absence of a fact is not a clearance and is not evidence that nothing is there. It will miss things, particularly where a person is being careful about what they say, which is exactly the circumstance that matters most. Never read a quiet screen as reassurance.
3. Every fact carries a confidence figure, and you can see it. Confidence describes how sure the extraction was, and nothing about whether the thing is true. Check anything you are about to act on.
4. It does not replace your safeguarding procedure. Your thresholds, your escalation routes and your statutory duties are unchanged. If your procedure says refer, refer, whether or not the tool noticed anything.
5. It is not a screening or triage instrument. It has not been validated as one, it is not offered as one, and it must not be used to decide who gets seen, in what order, or at what level of service.
6. Correct what is wrong. A fact left uncorrected carries into every following turn and shapes the translations that come after it. That applies as much to age band and gender as to anything else.

The reason this exists is that language barriers hide risk: a person explaining a difficult situation through a tool gives less away than they would in their own words to someone who speaks them. The reason to be careful with it is that a tool which appears to watch for danger invites people to stop watching themselves.

It does not replace professional responsibility or judgment

My Words is a support tool inside a conversation that a person is running. It translates, it scores its own work, and it tells you when it is unsure. It does not assess, it does not decide, and it carries no professional responsibility for anything that happens in the conversation. You do.

That matters most at the moment the tool looks most convincing. A fluent translation with a good score is still a translation, and the judgment about what the person means, whether they have understood you, what risk is present and what should happen next stays with the practitioner throughout. Nothing on the screen transfers any part of that, and no output of the tool is a defense for a decision you would not otherwise have made.

What remains yours:

1. Deciding whether this conversation is appropriate for the tool at all.
2. Noticing risk. The tool will help you see it and will never be the reason you did not.
3. Checking understanding directly, in the way you would with any conversation that matters.
4. Stopping when something is not landing, even where the screen looks confident.
5. Recording the substance of the conversation in your own systems, in your own words.

Consent and the other person

The person you are speaking with should know what is happening. In practice that means:

1. They are told that a machine translation tool is being used, and that they can ask for an interpreter instead.
2. They are told if a supervisor is present. The product says so itself, so it does not depend on the worker remembering.
3. On their own device, the interface is in their language, so the explanation reaches them as well as the translations do.