AI transparency

What the system does

Sanctavo listens to the person speaking at a service, recognises what they said, translates it, and speaks the translation in the listener's language, within a few seconds. Afterwards it can write a summary, a study guide, an article or a newsletter from the transcript.

Every one of those steps is performed by an artificial intelligence model. There is no human interpreter and no human transcriber in the loop while the service is running.

What it gets wrong

Speech recognition on a live room is imperfect, and translation of theological language is genuinely hard. In our own measurements the most common failures are:

• Names. Personal and place names are frequently transcribed as they sounded rather than as they are spelled.
• Numbers and scripture references. A chapter and verse can come out wrong, which matters more here than almost anywhere else.
• Doctrinal vocabulary. A word with a specific meaning in one tradition can be translated into a word with a different one. Churches can supply their own glossary, and we encourage it.
• Missing speech. A very quiet passage, a person away from the microphone, or heavy background noise can be lost.

None of this makes the translation useless, it makes it a translation, not a transcript of record. Treat it as you would a volunteer interpreter doing their best in real time.

A person reads it before you do

Nothing written by AI is published to a congregation until someone at the church has read it and approved it. That is enforced by the software, not by a policy: a draft cannot be sent or published while it is unapproved, we record who approved it and when, and editing an approved text withdraws the approval, because the approval belonged to the words that changed.

Live subtitles and live translated audio are the exception, they are produced and delivered in the same few seconds, so nobody can read them first. That is the trade the service makes, and it is why the live screen says the translation is automatic.

How AI content is marked

Two different things, and they are worth separating.

• A notice you can read. Present everywhere: on the live screen, above any content published to listeners, and on the cover of every Word and PDF document, in the reader's own language.
• A marker a machine can detect. Partial. Our responses carry an X-AI-Generated header and our documents carry a marker in their file properties. The live translated audio carries no watermark, and we are not going to claim otherwise.

Watermarking a real-time synthesised voice is not currently feasible at this scale: none of the providers we use offers one on the streaming interface, and the marking schemes that exist do not survive the chunking and compression a live stream needs. We have written down the reasoning, we revisit it as vendors ship capabilities, and we would rather publish the gap than a claim we cannot support.

Who processes the audio

Producing a translation means sending the audio to specialist providers. Which ones, what they receive, and where they process it is published at /subprocessors.

What we do not do

• We do not use your content to train AI models.
• We do not identify listeners. There is no account, no login and no record connecting a person to what they listened to.
• We do not make automated decisions about anybody that would have legal or similarly significant effects.
• We do not synthesise a voice imitating the speaker. The translated voice is audibly an interpreter, not the preacher.