Yes, a church office can use ChatGPT under the GDPR if no identifiable member data goes in. Never paste names alongside health, pastoral, bereavement, safeguarding, family or giving details, or whole rotas and member lists. Use invented or redacted examples, check the plan's data settings, and put borderline cases to a data-protection adviser.
A church needs a stricter rule than a shop because of Article 9. Data that reveals religious beliefs is special-category data, and a member list, and often a rota or prayer list, reveals belief simply by showing who belongs. Article 9(2)(d) lets a religious not-for-profit body process such data, with appropriate safeguards, where it relates solely to its members, former members and people in regular contact, but only if the data is not disclosed outside the body without the consent of those people. Whether pasting into an AI tool counts as disclosure is a question for your adviser. Until you have an answer, act as though it does. None of this is legal advice.
Why a church's ordinary lists count as sensitive
Article 9(1) names the special categories: racial or ethnic origin, political opinions, religious or philosophical beliefs, trade union membership, genetic and biometric data used to identify a person, health, and sex life or sexual orientation. Processing them is prohibited unless one of the listed exceptions applies, such as explicit consent or the religious-body provision above. Regulators' guidance adds that data which "reveals" a category counts, not only data that states it outright.
A church office holds a great deal of this by accident. The directory shows who belongs. The rota shows who serves and sometimes why they cannot ("mornings only since her treatment began"). The visiting list shows who is ill, bereaved or struggling. The giving record shows who gives and how much. Add youth-group forms with allergies and parents' numbers, and a safeguarding file, and the office is a small archive of the most private facts in people's lives.
Two other Articles shape the habit. Article 5(1)(c) asks that data be "adequate, relevant and limited to what is necessary", which in practice means paste less than you think the task needs. Article 10 gives criminal convictions and offences, including the results of criminal-record checks on volunteers, their own tight rules. So the question for every paste is not "is this a secret?" but "does this task need a real person in it at all?" Usually it does not.
The never-paste list
These stay out of ChatGPT in every form, redacted or not, on every plan. If a task seems to need one of them, do the task by hand, or ask AI for a blank template and fill it in yourself.
| What it is | Why it stays out | Ask AI for this instead |
|---|---|---|
| Pastoral-care lists and visit notes | Names plus circumstances: health, family trouble, money, grief | A blank visiting-schedule layout, filled in by hand |
| Prayer requests naming a person and their situation | Health and personal detail, shared on trust for one purpose | Wording for a prayer-request form, with no real requests |
| Health, bereavement and welfare details | Health data is special category; deaths and illness identify families fast | A condolence-note template with [name] and [relationship] left blank |
| Safeguarding concerns, disclosures and referrals | The most sensitive records you hold; they follow your safeguarding procedure, not an AI chat | A checklist of what your procedure says, written from your own policy |
| Criminal-record check results | Article 10 gives this data its own tight rules | A one-page explainer of why checks are done, with no individual's result |
| Children's and youth-group records | Names with ages, parents' contacts, allergies and photo consent | A registration form layout with empty fields |
| Giving records and tax-relief declarations | Who gives, how much, and bank or address details | A thank-you letter with [amount] left out, merged by hand |
| Staff and volunteer HR matters | Grievances, references, sickness, pay | A policy or letter template with placeholders |
| Whole member directories and rotas with contact details | Names, numbers, addresses and availability notes in one place | A rota grid with roles, never people |
| Minutes of closed sessions | Individual staffing and welfare cases | An agenda layout and a minutes template |
| Anything said in confidence | It was shared for one purpose with one person | Nothing: do not paste it |
| Logins, bank details, door codes | Security, not only privacy | Nothing |
The pattern in the right-hand column is the useful one. AI is good at shape: layouts, templates, wording, structure. The people go in afterwards, by hand, outside the tool. For where the line falls between admin and pastoral tasks, the tutorial on using ChatGPT for church work sorts the common jobs.
The redact-first list
These tasks are fine once the personal detail is gone. The redaction habit below is what makes them safe.
- A volunteer rota. Roles and dates only: "Welcome team, first Sunday". Names, numbers and notes about availability stay out.
- A newsletter draft. Keep the structure and the general news. Cut stories about individuals unless the people concerned have agreed to have them in print.
- A hall-hire enquiry. Remove the sender's name, address and number. The booking question itself is fine.
- Feedback from a survey or a service. Strip names and any detail that could identify the speaker, then ask for themes.
- A funeral or wedding order of service. Paste the structure and the standard wording, not the family's details.
- A job description or advert. Fine, provided it does not describe a named person's situation.
The always-fine list
Where no person appears, there is no member data to protect. These need nothing but care with accuracy: notice-sheet layouts, a template for a welcome letter, drafts of policies and procedures, fundraising letters with no donor names, event posters, agendas, a thank-you for volunteers in general terms, and generic wording for forms. The ChatGPT prompts for church admin are written to stay on this list.
One office week, sorted by the three lists
An illustrative office has one part-time administrator and six volunteers serving a congregation of about 180. Here is a week of tasks, sorted, with the time AI might save on each (assumed minutes, so replace them with your own):
| Task | Verdict | How AI is used | Minutes saved |
|---|---|---|---|
| Sunday notice sheet | Fine | Layout and wording from approved items | 25 |
| Summary of the volunteer safeguarding policy | Fine | Plain-English version of your own policy, no individuals | 30 |
| Blank youth-group registration form | Fine | Layout with empty fields | 20 |
| Reminder to the welcome team | Redact first | Template with [first name]; names merged by hand | 10 |
| Reply to a hall-hire enquiry | Redact first | Sender's details removed before pasting | 10 |
| Order of service structure | Redact first | Standard wording only; family details added by hand | 20 |
| Giving statement emails | Template only | Wording drafted; amounts and names merged outside the tool | 15 |
| This week's visiting list | Never | Done by hand | 0 |
| Closed-session minutes | Never | Done by hand | 0 |
Seven of the nine tasks can use AI, and the saving adds up to 130 minutes, a little over two hours. The two "never" tasks are the ones where a slip hurts most, and they cost nothing in the sum because they were always done by hand. The week shows the point of the lists: the saving comes from the fine and redact-first tasks, so the strict rule on the rest costs the office very little.
A redaction habit in three passes
For anything on the redact-first list, work through three passes before you paste. Keep any key that maps letters back to people in a separate document, never in the chat.
- Remove the direct identifiers. Names, addresses, phone numbers, emails, dates of birth. Replace names with a role or a letter ("Volunteer A").
- Remove the indirect ones. In a small congregation, a few facts identify someone without a name: "the only family with triplets", "our oldest churchwarden", "the lady who plays the organ". Cut them or make them vague.
- Delete the sensitive context. Health, family, money and belief details are deleted, not masked. "Volunteer A (hospital appointments)" is still health data.
Then apply the congregation test: if someone in your pews could read the text and work out who it describes, it is not redacted enough. Four examples show the habit at work.
The rota. The raw row holds a full name, a mobile number and a note that the person can only do mornings since a medical treatment began. After the three passes it reads "Volunteer A, first Sunday, refreshments". The medical note is gone, not shortened, because the rota never needed it. The full version, with the names and the note, stays in your own spreadsheet.
The pastoral letter. Do not paste the situation. Ask for the shape and write the person in yourself:
Draft a short, warm letter from a church leader to someone who has been unwell, offering a visit. Use [name] and [date] as placeholders. Do not mention any detail about the illness. Keep it under 120 words and plain in tone.An illustrative reply:
Dear [name], I was sorry to hear that you have been unwell, and you have been in our thoughts and prayers. If it would help, I would be glad to visit on [date], or on another day that suits you better. There is no need to write back; a quick call to the office is enough. I hope your recovery is swift and that we see you back with us on Sundays soon. With warm wishes, [your name]
Two edits before it goes out. "I hope your recovery is swift" assumes an outcome nobody in the office knows, and to someone with a long-term condition it can read as careless, so it comes out. "Back with us on Sundays soon" puts gentle pressure on a person who may not be able to come, so replace it with an open offer of help. The placeholders came back intact, so the name, the date and any personal line go in by hand, in your own words. The letter is better for it, and nothing sensitive left the office.
The council minutes. A closed session about an individual never goes in. For an open item, paste the agenda heading and the decision, with names cut: "Item 4: hall roof repair, quote accepted, work to start in spring." That is enough for AI to tidy the wording of the minutes.
The prayer notice. The raw item says that a named family's baby is in intensive care and the mother cannot attend. Even the redacted version, "a family in the congregation has a baby in hospital", can identify them to a small community. So the request to AI carries no detail at all: "Write a short, gentle notice asking the congregation to hold a family in their prayers. Give no details. Use [family] as a placeholder." The leader adds the name only if the family has agreed to it being read out.
For a longer routine covering documents of all kinds, see the tutorial on anonymising data before you paste it. Its pseudonymised-versus-anonymised distinction matters here, because under Recital 26 data that could be linked back to a person is still personal data.
Which ChatGPT account the office uses
The data categories do not change with the plan. The settings do, and they decide how forgiving a mistake is. As of OpenAI's own pages, checked on 28 October 2026:
| Account | What OpenAI says |
|---|---|
| Personal (Free, Go, Plus, Pro) | Data sharing is on by default. Switching off "Improve the model for everyone" stops new conversations being used for training, but existing chats stay in history. Temporary Chats are not used to improve models and may be kept up to 30 days for safety. Thumbs-up or thumbs-down feedback can send the whole conversation for training even after you opt out. Turning off training does not delete anything: deleting is a separate step. |
| Business | OpenAI says it does not train on your data by default and can execute a data processing addendum with customers. Deleted conversations are removed within 30 days unless the law or the protection of its services requires longer. At least two seats are needed. |
OpenAI's enterprise privacy page and data controls FAQ carry the wording. A data processing addendum is the contract that Article 28 expects when another organisation handles personal data on your behalf: the controller may use only processors giving sufficient guarantees, under a contract setting out the subject matter, duration, nature and purpose of the processing. Having one does not settle the Article 9 question above, which stays with your adviser.
Three account habits follow. Use one named office account for church work, and ask volunteers not to use personal accounts for anything beyond the always-fine list. On any personal plan, switch off the training setting before the first paste. And look at the memory settings: ChatGPT's memory features can keep details between chats, and OpenAI reworked them in June 2026, so check what is switched on and clear what has been kept. For how all the major tools compare on these points, read the GDPR comparison of ChatGPT, Claude, Gemini and Copilot.
Consent wording you can adapt
Where you do hold member details, tell people plainly what you do with them. Under Article 4(11), consent must be freely given, specific, informed and unambiguous. Article 7 adds that you must be able to show it, that the request must be clearly distinguishable from other matters, and that withdrawing must be as easy as giving it. For the special categories, Article 9(2)(a) asks for explicit consent. These samples are starting points for your adviser to check, not finished legal text:
Rota contact details: "I'm happy for [church name] to use my name and mobile number to organise the volunteer rota and send reminders. These details stay within the church and are never entered into AI tools. I can change my mind at any time by telling [contact]."
Prayer requests: "I'm happy for this request to be shared with [the prayer group by email / the congregation in the service]. Please [use / do not use] my name. I understand I can ask for it to be removed at any time."
Photos: "Photos from services and events may appear on the church website and social pages. Tick here if you are happy for photos showing you to be used. For children, a parent or guardian ticks. You can withdraw this at any time."
One line for your privacy notice also helps: "We use AI writing tools to draft notices, templates and letters. We do not enter members' names or personal details into them." Do not word consent as a condition of belonging or of receiving pastoral care, because Article 7(4) asks whether a service is made to depend on consent it does not need.
If something has already been pasted
It will happen once. Treat it as a routine to follow, not a disaster to hide. An illustrative case: a volunteer pastes the prayer-chain list into her own free account "to tidy the formatting". It holds first names, email addresses and a line on each person's need. That is contact details, health-related need and implied belief in one paste, on a personal plan where training is on by default. She mentions it at the next meeting, so the office can act within days, not weeks. The six steps below are what it should do.
- Stop and write it down. What was pasted, by whom, when, into which account and plan, and whether the chat was a Temporary Chat.
- Fix the account. Switch off training if it was on, and delete the chat. Remember that deleting and switching off training are separate steps, and check the plan's own page on how long deleted chats are kept.
- Tell the person responsible the same day. Name that person in advance. The clock in Article 33(1) runs from when the organisation becomes aware: unless the breach is unlikely to put people at risk, the regulator must be told without undue delay and, where feasible, within 72 hours of that moment.
- Let the adviser judge the risk. Whether the breach is "unlikely to result in a risk" is a judgement for them, not the volunteer who made the slip.
- Record it anyway. Article 33(5) requires a record of every breach, including those you decide not to report, so the regulator can check how you decided.
- Change the habit. Ask what made the paste easy, such as a rushed rota or a template that asked for names, and fix that.
Five questions to take to your adviser
Your denomination's central office, a data-protection adviser or a solicitor can answer these. Your data-protection regulator may also run a helpline for small organisations. Bring the questions in writing:
- Does Article 9(2)(d) cover the way we use member data, and does passing it to an AI provider count as disclosure outside the body?
- Is a business-plan account with a data processing addendum acceptable for any member data, or should the list above stay absolute?
- Do we need a data protection impact assessment before any member data is used with AI? (Article 35 requires one for high-risk processing, and names large-scale processing of special categories.)
- What should our privacy notice say about AI tools?
- What are our deletion, retention and breach steps for a mistaken paste?
If you want a one-page version of these rules for volunteers, the sample church AI policy shows one in policy form, and the tutorial on volunteers using their own accounts covers the personal-account problem in more depth.
Questions church offices ask about ChatGPT and data
Can volunteers use their own ChatGPT accounts for church tasks?
Only for tasks with no member data, such as notice templates or wording ideas. A personal account sits outside any agreement your church holds, and on personal plans the model-improvement setting is on by default until the user switches it off. Name the one account the office uses, and make 'no member data in any other account' part of the volunteer briefing.
Is redacted data still personal data?
It can be. The GDPR treats pseudonymised data, where a key could link it back to a person, as personal data, and in a small congregation a few details can identify someone without a name. Redaction lowers the risk but does not guarantee anonymity. Apply the congregation test below, and ask your adviser about anything that still points at a person.
Do we have to tell members that we use ChatGPT?
If personal data ever goes into a tool, your privacy notice should reflect who receives it. The simplest route is to keep member data out entirely and say so in one line of the notice. Your adviser can confirm what your notice must contain, because the answer depends on what data you hold and why.
Do the same rules apply to Claude, Gemini or Copilot?
Yes. The data categories and the redaction habit do not change with the tool. What changes is the account type and its settings, which the tutorial comparing ChatGPT, Claude, Gemini and Copilot against the GDPR sets out. Apply the never-paste list first, then check the tool.
Further reads
- AI Receptionist for Churches: Office Phone Cost and Pastoral Call Rules — Which calls must always reach a person, not an AI.
- AI for Church Administration: Rotas, Notice Sheets and Room Bookings — A week of church admin with AI, rotas included.
- Charity AI Risks: Data Protection, Deepfakes, and Donor Trust — Data, deepfake and donor-trust risks for small charities.
- How to Redact Personal Data From Documents With AI Before Sharing — Detect, review and apply redaction before sharing files.
- AI Chatbot for Churches: Weddings, Baptisms and Hall-Hire Questions — A church chatbot built from a one-page fact sheet.
- OpenAI for Nonprofits: Who Qualifies and What to Check First — Who qualifies for OpenAI's nonprofit offer.
- AI Church Website Builder: Wix, Squarespace or GoDaddy Airo Compared — Wix, Squarespace or GoDaddy Airo for a small church: list prices, who keeps the site current, events pages, giving links and the cost of changing later.
- ChatGPT for Nonprofits: What Donor Data Must Never Be Pasted In — A field-by-field never-paste list for donor and beneficiary data, a six-step redaction habit, a 1,200-donor worked example and a one-page staff rule.
- Church Translation With AI: A Test to Run Before You Pay a Vendor — A two-evening test with a bilingual checker that shows whether AI can translate your notices, newsletters and live services, and what to ask vendors.
- How to Manage a Community Centre: Where AI Helps and Where It Cannot — Four tests that sort community-centre jobs into AI drafts and person-only decisions, with worked examples for bookings, rotas, newsletters and grants.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: GDPR text, Articles 4, 5, 7, 9, 10, 28, 33 and 35 and Recital 26; OpenAI enterprise privacy page and data controls FAQ; data-protection regulator guidance on special category data. Checked 28 October 2026. This is general information, not legal advice.