For notices and newsletters, yes, once a bilingual member has checked every time, date, name and number before it goes out. For liturgy, scripture and pastoral messages, no: use published texts or a human translator. Live captions fall in between, and the only way to know is to test your own speakers' voices, vocabulary and sound system.
AI translation is strongest on plain announcements and weakest where one word carries doctrine, history or a person's name. A sentence such as "the soup lunch follows the morning service" comes through almost any tool intact; a line from a creed may not, and even a plain notice needs its times and names checked. A test on your own material takes about two evenings and costs little or nothing, and its results are what you put in front of vendors instead of their demo clips.
Why notices, liturgy and sermons get different pass marks
The mistake that causes most trouble is treating "church translation" as one job. A weekly notice sheet, a prayer, a sermon heard live and a hall-hire contract each have a different cost of being wrong, so each needs its own rule about who drafts, who checks and what counts as a pass. The marks below are my suggestions for your own test, not an industry standard.
| Material | AI draft acceptable? | Who must check | Suggested pass mark |
|---|---|---|---|
| Weekly notices and event details | Yes | Bilingual member, every week | No error in any time, date, place, name or number |
| Newsletter and news items | Yes | Bilingual member | Zero critical errors; no more than two meaning slips per 400 words |
| Bereavement, pastoral and safeguarding messages | No: write or translate by hand | Pastor plus a bilingual person | Human-authored; AI not used |
| Hall-hire terms, policies, accounts | Draft only | Professional translator | Signed off by someone accountable for the wording |
| Liturgy, creeds, prayers, scripture | No: use published texts | Your tradition's approved versions | Word for word from the approved text |
| Hymn and song lyrics | No new translations without permission | Rights holder or licence terms | Projected as published |
| Live sermon captions | Test first | Bilingual member watching the output | Names and scripture references right; delay readable |
Only the notice, newsletter and live-caption rows are worth testing with AI. The rest are listed so nobody is tempted to widen the test to them.
Build the test pack from your own bulletin
Use real material, because tools score well on the invented sentences in vendor demos. Take the last six weeks of bulletins and pick 20 items. Eight should be event notices with times, dates and rooms. Four should be newsletter paragraphs. Three should be names-heavy: a rota, a volunteer thank-you, a list of speakers. Two should contain church vocabulary that has more than one rendering, such as service names, ministries and course titles. Three should be "sensitive" stand-ins: a bereavement notice, a prayer request and a safeguarding statement. You don't plan to use AI for these, and they are in the pack to confirm that decision.
The sensitive stand-ins are invented. Write them with the same structure and made-up names, because real pastoral details should not go through any online tool, and the point is to see how badly a tool handles the tone. If you're unsure what may be pasted into AI at all, what member data never goes into ChatGPT sets out the line, and admin versus pastoral care explains why the sensitive items are stand-ins.
Count the pack before you start. Twenty items of this kind come to roughly 1,500 words, or about 9,000 characters at six characters a word. That matters because free tiers cap characters: DeepL's plan comparison lists 50,000 characters a month on its free version, so one pack fits several times over, though the weekly routine later will use part of the same allowance.
Run the pack through two or three tools with one prompt
Pick tools your volunteers could realistically use: a chat assistant such as ChatGPT, Claude or Gemini on a free or low-cost plan, a dedicated translator such as DeepL, and optionally Google Translate as a baseline. Use the same instructions for the chat assistants so the comparison is fair, and keep a short glossary of the words your church uses in a particular way.
Translate the church notice sheet below into [language] for a congregation of mixed ages.
Rules:
- Keep all times in 24-hour format and keep every digit unchanged.
- Keep names of people, rooms and programmes in English, in brackets after a translation only if I list one in the glossary.
- Use the glossary terms exactly. If a term isn't in the glossary and could be translated more than one way, do not guess: list it under QUESTIONS at the end.
- Do not add or remove any information.
Glossary: Upper Room = Upper Room (unchanged); Harvest Supper = [approved translation]; fellowship = [approved term]
Notice sheet:
[paste items]Run each tool once, paste the results into a table with a column per tool, and label the columns Tool A, Tool B and Tool C so the checker doesn't know which is which. Ten minutes of preparation here prevents the checker from marking a tool they have heard of more kindly.
A single marked item shows what the checker's sheet looks like. The notice is invented, and the translation is the kind of slip a tool might make, not a record of any particular product.
ENGLISH SOURCE
Harvest Supper: Saturday [date], 18:30 in the Upper Room hall. Bring a dish to share. Free, but please tell [name] by Wednesday. Children welcome.
TOOL OUTPUT (Spanish)
Cena de la Cosecha: sábado [fecha], a las 6:30 de la mañana, en el salón Aposento Alto. Traiga un plato para compartir. La entrada es gratuita, pero avise a [nombre] antes del miércoles. Los niños son bienvenidos.
CHECKER'S MARKS
S1: "a las 6:30 de la mañana" means 6:30 in the morning. Source says 18:30. Should read "a las 18:30".
S1: "Aposento Alto" is the Bible's phrase for an upper room; the sign on our hall says "Upper Room", so a visitor won't find it. Keep the English name.
S2: "antes del miércoles" means "before Wednesday", a day earlier than "by Wednesday". Use "a más tardar el miércoles".
S4: "Traiga" and "avise" are formal; our congregation uses "tú" ("Trae", "avisa"). Preference only.
Verdict: send after edits (3 minutes).Notice what the checker caught. The tool turned a 24-hour time into a morning hour, the classic S1. It swapped the hall's name for a biblical phrase that matches no sign in the building, which no spell-checker would flag. And it quietly moved the reply deadline a day earlier. Every sentence was fluent, which is exactly why unchecked machine output is risky for notices.
The bilingual checker: who, how long and what they mark
Your checker needs to read both languages comfortably and to know how your congregation speaks. A longstanding member who works in both languages can be a better checker than a professional with no feel for your vocabulary, though a paid proofreader is the right answer when nobody qualifies. Set aside 35 to 45 minutes per language for a 20-item pack. Give them a plain scoring sheet and tell them to mark errors, not to rewrite.
| Level | What it means | Example in a notice |
|---|---|---|
| S1 critical | A fact is wrong, missing or reversed | A weekday and date that don't match; a time shifted by 12 hours; a phone number short of a digit |
| S2 meaning drift | Understandable but says something your church wouldn't | "Fellowship" rendered as a word that means the sacrament in your tradition |
| S3 awkward | Correct but stiff or unnatural | A literal rendering of a casual invitation |
| S4 style | Preference only | Formal versus informal address |
Ask the checker to note two more things per item: whether they would send it as is, after light edits, or not at all, and how many minutes it took to fix. The second number is your real running cost.
Here is a made-up result so you can see how to read a sheet. The pack is 20 items, the language is Spanish, there is one checker, and both tools are anonymous. These figures are invented to show the method; your tools will score differently.
| Measure | Tool A | Tool B |
|---|---|---|
| S1 critical errors | 2 | 1 |
| S2 meaning drift | 5 | 4 |
| S3 awkward wording | 8 | 11 |
| Items sendable with no edit | 7 of 20 | 6 of 20 |
| Checker minutes to fix the pack | 38 | 41 |
Look at what the S1 errors were, not just the count. Suppose Tool A's two were a Thursday that became a Tuesday and a room name, "the Upper Room", translated into a phrase the congregation has never used, and Tool B's one was a phone number missing its last digit. Neither tool passes unattended, and that is the finding: the process passes only because the checker catches every S1. What you've measured is how many corrections the checker faces each week and whether they can keep up.
Repeat the exercise for each language you plan to offer. A tool that handles one language well can stumble on another, and the checker for each language is a different person with different standards. For a longer treatment of the general method, how to check AI translations before customers see them applies equally to a congregation, and testing translation tools on your own documents shows the scoring in another setting.
The sums: AI plus a checker against paying per word
Take the notice sheet. Say it runs to 400 words a week in one language. The AI draft takes about four minutes to produce and paste, and the checker spends about 15 minutes on it, so the weekly cost is under 20 minutes of a volunteer's time plus whatever the tool costs, which can be nothing.
Paying a professional per word looks like this. Gengo's pricing page listed a standard level from $0.09 a word and an advanced level from $0.18 a word, with no minimum spend. For 400 words a week that is $36 or $72, which over 52 weeks is $1,872 or $3,744 for a single language. Those figures explain why churches look at AI, and also why a volunteer's checking time is worth protecting: the saving is real only if the check actually happens. Add a second language and the human route doubles while the AI route adds another checker's 15 minutes. For the broader cost picture, what documents really cost with AI against a human translator breaks it down by document type.
The weekly routine once the pack passes
A pass on the test earns the right to run a routine, not to stop checking. The routine is short. On Thursday someone pastes the English notice sheet into the approved tool with the saved prompt and glossary. On Friday the bilingual checker spends about 15 minutes on it, correcting rather than rewriting, and initials the sheet. On Saturday it's published alongside the English. Keep a one-line log each week: date, checker, number of S1 and S2 errors found. The log is how you notice drift.
Set a stop rule in advance, tied to what the test showed. A month of notice sheets is roughly the size of the 20-item pack, so the pack's S1 count is a fair monthly baseline: if the checker finds more S1 errors in a month than the chosen tool made in the test, or if any S1 error ever reaches the published sheet, stop using the AI draft for that language until you have retested, and let a person translate in the meantime. A stop rule agreed in a calm moment is much easier to follow than a judgement made after a notice has gone out with the wrong time.
Liturgy, scripture and hymns stay outside the test
A machine can turn a creed or a prayer into fluent text, and the fluency is the danger. Many traditions have approved translations of their prayers and forms of service, and a congregation hears the familiar wording as part of what it is saying. Replacing it with a fresh machine rendering changes the words people have learned by heart, and may change meaning the tradition protects. If an approved translation exists, paste it. If none exists, the translation is a human job done by someone your leaders trust.
Scripture follows the same rule. Quote the published Bible that your congregation reads in that language, with the reference written the way that edition writes it, rather than asking a model to translate a verse. Hymn and song lyrics add a legal wrinkle: a translation of lyrics is a new version of someone's work. Your music licence usually covers projecting lyrics as published, and may not cover new translations, so read its terms or ask the rights holder before projecting one.
A five-minute audio test for live captions
Live captions need a different test because the failure modes are different: names, scripture references, numbers, speed and sound. Record five minutes of a typical sermon segment with the speaker's agreement, using the same microphone feed the captions would receive. Choose a stretch that includes a scripture reference, a few names, a number and a reading. Have a person type a reference transcript of the first 200 words.
Play the recording through each tool you are considering, including any free trial hours a vendor offers. Your checker counts three things: word errors in the first 200 words, any name or scripture reference that is wrong, and the delay, measured with a stopwatch from a spoken word to its caption. A reasonable rule of thumb for your own use is that any wrong name or reference is a fail, and that the delay should be short enough that a reader can follow along without losing the speaker. Five errors in 200 words is 2.5%, and worth knowing, but the named-term failures matter more than the percentage.
Check each tool's own warnings before the test, because they differ. Microsoft's support page says PowerPoint for Microsoft 365 can show captions in the spoken language or subtitles translated into another, recommends a headset microphone and needs a reliable internet connection. It lists more than 60 on-screen languages, but splits spoken languages into fully supported and preview, and says preview languages generally have lower accuracy (PowerPoint live captions and subtitles). Google's page for captions in Google Slides describes captioning the speaker's voice only and warns the captions might not be complete or accurate. YouTube's help page says automatic captions on live streams are available in English only and can't be switched on for a whole channel, and it lists overlapping speakers and poor sound as reasons captions fail (YouTube automatic captions). Those are the vendors' own statements, checked in October 2026, and any of them can change.
Two practical results usually come out of this test. One is that the quality of the audio feed matters as much as the tool, so a clean feed from the sound desk beats a laptop microphone at the back of the room. The other is that every tool stumbles on a pastor's own name and on unfamiliar place names unless it can be taught them, which leads to the next question to put to vendors.
What four vendors say on their own pages, and what to ask them
Four live-translation services for churches rank for this search: ChurchTranslate, Glossa, LiveSunday and OneAccord. I haven't tested any of them. The table records what each page itself said on 29 October 2026, in USD, so you can compare the pricing models. Treat it as a starting list for questions, and confirm everything with the vendor.
| Vendor | Pricing as published | Quality controls it names |
|---|---|---|
| ChurchTranslate | Pay as you go at $5 per active-speech language-hour, $25 minimum purchase; a $99 a month plan with 20 language-hours | Each language reviewed with real audio and by a native speaker before it can be selected; failed routes shown rather than hidden |
| Glossa | Pay as you go at $5 an hour per language; $99 a month plan with 25 translation hours, $4 for each extra hour | Custom vocabulary for names and ministries; says silence, songs and music don't count towards hours |
| LiveSunday | First two hours free; pay as you go at $5 an hour with one translation language, plus $5 a month for each extra language | Full session transcript history; no review step named in the home page text I could read |
| OneAccord | From $150 a month for five hours; no setup fee; free credits on request | Optional moderation mode to review the transcript before translation; says it favours human interpreters where possible |
The pricing models differ more than the headline rates suggest. ChurchTranslate and Glossa bill per language-hour, so hours multiply by the number of languages. A 75-minute service with 45 minutes of speaking, translated into three languages, is 2.25 language-hours if only speech counts and 3.75 if the whole service does. Over a month of about 4.33 services (52 a year divided by 12) that is 9.75 or 16.25 language-hours, which at $5 comes to $48.75 or $81.25. LiveSunday's pay-as-you-go page prices the hour instead, with $5 a month for each language after the first, so the same month works out at roughly $26 to $37 depending on which minutes count. OneAccord quotes on languages, hours and users. The definition of billable time matters more than the rate, so ask each vendor to state in writing how songs, silence and prayers are counted, and whether extra languages multiply the hours.
Then put the same four requests to every vendor. First, run your 20-item pack and your five-minute recording and send back the output. Second, show how a vocabulary list is entered and whether it applies to captions and audio. Third, explain what happens when a language route fails mid-service. Fourth, state what is stored: audio, transcripts, how long, and whether any of it is used to train models, because sermons carry names and prayer requests. Score the answers with your checker, and ask for a pay-as-you-go or trial period before any annual commitment.
Cases where a human interpreter or translator is the right call
AI plus a checker suits routine, low-stakes text. A person should lead in these situations: funerals, weddings and baptisms, where the words carry weight on the day; pastoral conversations and counselling; safeguarding statements and anything a regulator or insurer might read; hall-hire contracts and annual accounts; statements of belief; and any service where a visitor is relying on the interpretation to take part. For a document such as a 1,200-word safeguarding policy, a professional at Gengo's published rates would cost roughly $108 at the standard level or $216 at the advanced level per language, which is modest against the risk.
A volunteer interpreter who whispers or speaks live also remains the best live option when you have one. One vendor, OneAccord, says on its own page that it champions human interpreters whenever possible, and treats its AI as support for a church starting out or a small interpreting team. ChurchTranslate's page makes the same point from the other side: AI can mishear names, theology and Scripture, so use a qualified interpreter wherever mistakes could cause harm. Captions then become a back-up for the interpreter, which is a more comfortable role for the technology than replacing the person.
Write down the result of your test as a one-page rule: which items use an AI draft with a checker, which are always human, who the checker is for each language, and when you will retest. Retest after any change of tool, any new speaker with a strong accent and at least once a year, because the tools change faster than a congregation's habits.
Questions churches ask about AI translation
Is Google Translate good enough for church notices?
It can produce a usable first draft of plain announcements, like any other tool, but it gets no exemption from the check. Put it in the same test pack as the others and let a bilingual member score it. Judge it on how many time, date, name and number errors it makes, not on how fluent it sounds.
Can ChatGPT translate scripture or hymns?
It will produce text, but that text is a new translation nobody has approved. For scripture, use the published Bible your congregation reads in that language. For hymns, translating lyrics creates a derivative work, so check the terms of your music licence or ask the rights holder before projecting a new version.
What if nobody in the congregation speaks the language well enough to check?
Then you can't run the test honestly. Pay a professional translator or proofreader to check the first test pack and the notice template once, then reuse the approved wording. Published per-word rates, such as Gengo's from $0.09 a word, make that a small one-off cost compared with sending unchecked text.
Do live captions cope with singing and several speakers?
Speech recognition is built for one clear speaker. Sung words, overlapping voices and poor room audio are where captions fail first, and YouTube's own help page lists overlapping speech and poor sound as causes of missing automatic captions. Project lyrics from published text and give the speaker a clean microphone feed.
Further reads
- ChatGPT Prompts for Churches: 10 Admin Jobs, No Sermons — Ten admin prompts for the English side of the notice sheet.
- Sample Church AI Policy: One Page on Staff, Volunteers and Member Data — A one-page policy so volunteers know what may go into AI.
- AI Chatbot for Churches: Weddings, Baptisms and Hall-Hire Questions — Where a chatbot can answer hall-hire questions in other languages.
- How Translators Use AI to Keep Terminology Consistent — How translators keep a glossary consistent, useful for church vocabulary.
- How to Offer Multilingual Customer Support With AI Translation — A wider look at replying to people in several languages.
- AI Church Website Builder: Wix, Squarespace or GoDaddy Airo Compared — Website builders compared, for a multilingual notices page.
- AI Receptionist for Churches: Office Phone Cost and Pastoral Call Rules — A monthly price range for a church AI receptionist, a list of calls that always go to a person, and a script for the office phone.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Microsoft Support page on PowerPoint live captions and subtitles; YouTube Help page on automatic captions; Google Docs Editors Help page on captions in Google Slides; DeepL plan comparison page; Gengo pricing and languages page; the home and pricing pages of ChurchTranslate, Glossa, LiveSunday and OneAccord, all read on 29 October 2026.