How AI is used

What is generated

• A transcript of the service from its sound, made by a speech model that marks who is speaking. When it cannot transcribe the sound, a second speech model, run on our own machines, does. Nothing is taken from YouTube or any other video platform: the recording is the church’s own upload.
• An understanding of the sermon: its big idea, its themes, the scripture it quoted, the lines worth keeping, the events announced from the front.
• Drafts: one post a day, shaped for each connected account; the blog post; the weekly email; the text; and the moments cut into reels, with their captions.
• Pictures for posts and events, in the style the church chose.
• The frame of each reel, which follows the person speaking: a small model finds where faces are in the church’s own recording so the picture can follow the speaker, and a vision model checks the finished reel shows them whole. Nobody is identified, and no face is kept.
• The cover of each reel, the picture a profile grid shows before anyone plays it: a real frame of the speaker, taken from the church’s own recording at a pause with the speaker looking toward the camera, sharpened to high definition and cleaned of background clutter by a picture model, with the person kept as they were. An edit is used only after a face check says it is the same person and both picture reviewers agree; otherwise a plainly sharpened frame is used. Only when no frame of the moment looks toward the camera may one more edit turn the eyes toward it, and a post that carries such a cover is marked as AI-altered to Instagram and YouTube. TikTok’s cover is a frame of the video itself.
• Music under each reel, composed by a music model from the church’s own sound and the worship in that service, set under the speaker’s voice.
• Answers in the Help panel, from the product’s own guide and the church’s settings.

What is checked

• Invention. Every draft is written from the transcript and told to say only what was said. The person who approves a piece is the check that it does: nothing is posted before they read it.
• Quotations. A quotation of six words or more in a draft is looked for in the transcript, by plain matching and not by a model. One that is not found carries a warning, “This quotation was not found in the sermon”, unless it gives its Scripture reference. The church decides what to do with it.
• The safeguards. Hate or derogatory speech and profanity or explicit content are screened for, with the policy the church chose for each: block, warn or off. A blocking finding holds only what it is in until a person resolves it. Every time the words change, the screen reads them again, and nothing is approved until those exact words pass; if the screen cannot be reached, nothing is approved. Each refusal is recorded, and we look into an account that is refused repeatedly. Nothing is rewritten or removed by a model; what a person reads is what was said.
• Four hard limits. In every church, whatever its settings, two models, each reading on its own, look for a threat or call to harm a real person or group, encouragement of suicide or self-harm, sexual content involving a minor, and the exploitation of a minor (grooming, enticing or trafficking a minor); a finding by either is enough. A piece holding one is never approved or sent until its words change, and only a person at Sunday Everywhere can clear one found in error. Both models are told the same rule: scripture, doctrine, preaching and testimony, and teaching on protecting children, are never screened as these. These are the lines the law draws; what the church may still choose to screen is its own.
• Private details. In public pieces only (social posts, reel captions and the blog post), a safety model looks for private details about a named or identifiable person: health, a prayer request, money, family matters or a child’s name. It only warns; the church decides. The email, the text and the notes are not checked for this, because prayer requests name people by design.
• Pictures. Every generated picture is reviewed against the church’s style and the rules below before it is offered. A picture safety model and both picture reviewers also check it for sexual content or nudity, a minor shown in any sexual or suggestive way, gore, self-harm or a hateful symbol: a picture that fails is never kept and another is painted, and one that cannot be checked is not kept.
• The public tier. A blog post or an event is published on a church’s page only if the exact words approved still pass: no hold or pause on the week, nothing blocked or open as a hard limit, every picture passed, and no private details about a person unless an Owner or Admin confirms them. A check that cannot be made means nothing is published. The public never causes a model to run: a report or a like starts no check of its own except, for the two reasons that name a child or hate, one re-check of that item a day.
• The nightly re-check. When the safety models or the policies change, the published pieces are read again, a bounded number a night, and one that now holds a hard limit is taken off the page and the church is told; a check that fails does not take a piece down, a finding does.
• What the church uploads. A frame from every thirty seconds of each recording, and every picture the church uploads (its photos, its leader’s picture and its logo), is checked by the picture safety model. A recording that appears to show sexual content holds its week until a person at Sunday Everywhere has reviewed it, usually within a day; such a picture is refused. Until a recording’s frames are checked, its reels cannot be approved.

When a check is wrong

The safety models can be wrong both ways: they can miss something, and they can find what is not there. So a hard limit found in error is cleared only by a person at Sunday Everywhere, a held week is reviewed by a person, and an account paused by a check is lifted only by a person after review. To ask for a review, ask in Help and a person answers by email. The checks are a safeguard, not a guarantee, and the church remains responsible for what it publishes, as the Terms (section 5) say.

Who approves

A person at the church. Every piece waits as a draft until someone with the right role reads it, edits what they like, and approves it under their own name; posting and sending happen only after that, and the record says who did it. An approved reel or post then goes out at its scheduled time, or a little later when the posting policy spaces it, and no model ever approves or posts on its own.

What is never done

• No impersonation. We never clone a pastor’s voice, or anyone’s, and never write as if a named person said something they did not.
• No real people in generated pictures. An AI picture never shows a recognisable real person, a member of the church or anyone else; the people in it are nobody in particular, even when the church’s own photos guide its style. A church’s own photos, added with its confirmation that the people shown agreed, may appear as themselves.
• No invented sound. A reel carries the room’s own sound or church music, never manufactured applause or a crowd that was not there. The speaker’s voice is cleaned by fixed signal processing, not a model: hum and harshness taken down, the level evened, and a room’s echo eased where it was measured high; nothing in a voice is generated.
• No training on your data. Your sermons, drafts and settings are not used to train models, by us or at our direction.
• No decisions about people. The models draft and screen; they never decide who may join, who is contacted, or what a person is owed. The one thing a check does on its own is stop: a piece, a week, or, only when an upload appears to show sexual material involving a minor, an account, until a person reviews it.
• Nothing is published by a model. A church’s page shows only the blog posts and events a person at the church approved and published, as approved; its reels and posts go to the accounts it connects, and its email and text to the list it keeps.

Who makes the models

Every text, vision, picture and short-audio model is reached through OpenRouter: today Anthropic’s Claude models read and write, Google’s Gemini models screen, review pictures and paint them, and an OpenAI model answers in Help. They run only on providers that keep nothing and train on nothing we send. ElevenLabs transcribes the service and composes the music; the fallback transcription runs on our own machines. The Privacy Policy lists the same companies and what each receives.

Four kinds of model do the checking: the screen’s model, which reads every transcript and draft against the church’s safeguards and the hard limits; a text safety model, which reads for the hard limits a second time and for private details in public pieces; a picture safety model, which checks every painted picture, every uploaded picture and the frames of each recording; and two picture reviewers, which judge every painted picture. Today Meta’s Llama Guard is the picture safety model, and an open safety model from OpenAI reads the words a second time.

Which model does each job is chosen by the operators who run the service. Every week the platform compares the models on offer, on the service’s own work, for quality and cost, and recommends a change only when one does better; nothing changes until an operator approves it, and every model, old or new, runs only on providers that keep nothing and train on nothing we send.

What it costs the church

Work the models do is counted as points against the church’s plan and shown on the Account page, Help answers and Selah’s minutes included. A church sees points, never a price for a piece.