Turn your monitoring data into a donor report draft without letting a single name leave your office

What happens to your information

What it costs

Before you start

• Your monitoring data in whatever form you already keep it, spreadsheet, Word file, incident log, or a notebook you can type from.
• The donor's report template or call document: exact section headings, word limits, and the logframe or indicator list if there is one.
• A laptop or phone with internet and an email address for a free account. No credit card, no installation, no terminal.
• One quiet hour where nobody interrupts you, because the de-identification step is the part that must not be rushed.
• One colleague who knows the cases well enough to recognise a person from a description. You need ten minutes of their time at the end.
• If your monitoring data sits under a donor confidentiality clause or a protection-team data-sharing rule, read that clause before you start. Some clauses forbid transfer to third-party processors outright, and no amount of de-identification cures that.

The steps

1. Step 1 / 16

   Before you open any AI tool, put two documents side by side on your computer. Name them exactly as shown. The first holds the truth. The second is the only one you are ever allowed to copy from.

   Copy this exactly

   REPORT-MASTER-DO-NOT-SHARE.docx
   REPORT-SAFE-TO-PASTE.docx

   Note

   This one habit is the entire recipe; everything else is detail. Keep the master file where you already keep sensitive material, the same encrypted drive or locked folder your protection policy already names. Never have a browser tab and the master file open at the same moment when you are tired. That is how names get pasted.

2. Step 2 / 16

   Copy your monitoring data into the SAFE file. Then build a small code key at the top of the MASTER file only, so you can still trace every claim back after the AI has finished. The key never goes in the safe file.

   Copy this exactly

   P1 = [real name, real case number]
   P2 = [real name, real case number]
   L1 = [real village / neighbourhood]
   L2 = [real village / neighbourhood]
   D1 = [real date, e.g. 14 March 2026]

   Note

   In the safe file the people become P1 and P2 and the places become L1 and L2. You keep full traceability for your own audit trail and for a donor verification annex, but the key never crosses the internet.

3. Step 3 / 16

   Now run the de-identification checklist over the SAFE file, top to bottom. Remove or replace every item in this list. Do not skip a category because it looks harmless, re-identification works by combining harmless things.

   Copy this exactly

   1. NAMES: every one. Survivors, witnesses, families, your own staff, drivers, interpreters, fixers, named officials, named soldiers, named lawyers, named doctors. Replace with P1, P2, "a witness", "a lawyer".
   2. EXACT DATES: replace with the month, the quarter, or "early 2026". A date plus a district identifies an incident everyone locally remembers.
   3. EXACT PLACES: no village, neighbourhood, street, checkpoint, camp block, school or clinic name, no GPS coordinate. Go up to governorate, or "a district in Nineveh governorate".
   4. DISTINGUISHING DETAILS: occupation, disability, injury, scar, tattoo, clothing, vehicle, tribe, family size and composition, "the only teacher in the village", "the pharmacy owner", "the family that returned from Turkey".
   5. SMALL NUMBERS: see the next step.
   6. DIRECT QUOTES: a quote is a fingerprint. Paraphrase in your own words, and never paste a testimony transcript.
   7. IDENTIFIERS: case and file numbers, ID and ration card numbers, phone numbers, emails, social media handles, vehicle plates.
   8. YOUR SIDE: names of partner organisations, field monitors, safe houses, shelters, referral clinics, and any office address.
   9. FILES: do not upload the original document, photo, scan or PDF. Photos carry GPS and camera metadata; PDFs carry author names and edit history. Type or paste cleaned TEXT only.

   Note

   Categories 4 and 5 are the ones people miss. Category 9 is the one that silently undoes the other eight, a single uploaded photo can carry the coordinates you spent forty minutes removing.

4. Step 4 / 16

   Deal with small numbers separately, because this is where documenters most often expose someone by arithmetic. Replace every small count with a band, then check that no combination of categories narrows to one person.

   Copy this exactly

   RULE: any count of people below 10 becomes a band, never an exact number.
   1-9  becomes  "fewer than 10"
   10-24 becomes "between 10 and 25"
   
   Then re-read every sentence and ask: does any COMBINATION point to one person?
   BAD:  "3 women in Sinjar district reported X; 1 was a widow with a disabled child."
   BAD:  "Of the 4 cases in the camp, 2 involved Christian families."
   GOOD: "Fewer than 10 women in Nineveh governorate reported X, including cases involving heads of household caring for dependants with disabilities."

   Note

   The arithmetic attack is simple. If you report 4 cases, then elsewhere say 2 were from one minority group and 1 involved a wheelchair user, anyone working in that camp knows exactly who all four are. Cross-tabulating small cells is the most common re-identification failure in human rights reporting. If a category holds fewer than 10 people, do not break it down by anything else.

5. Step 5 / 16

   Read the SAFE file once more, out loud, pretending you are a hostile security officer who works in that district and wants to know who talked. If you can name anyone, cut more. Then have a colleague who knows the cases do the same for ten minutes.

   Note

   You will lose some vividness. That is the correct trade. A donor report saying "a returnee family in Nineveh" and safe is worth more than one saying "a returnee family in Hardan village" and not. If the donor genuinely needs exact detail, it goes in a separate annex through your normal secure channel, never through a chat box.

6. Step 6 / 16

   Open your chosen free tool and turn off model training BEFORE you paste anything. Do this once per account; it does not carry across accounts or across colleagues' logins.

   Copy this exactly

   ChatGPT: profile icon (top right) > Settings > Data Controls > turn OFF "Improve the model for everyone"
   Claude: profile icon > Settings > Privacy > turn OFF the model-improvement / training setting
   Gemini: profile icon > Gemini Apps Activity > turn it off

   Note

   Neither OpenAI's nor Anthropic's public help page states plainly whether this is on by default, so look at the toggle yourself. Anthropic's help centre says that if you do allow training, "we may retain your data in a de-identified format for up to 5 years in our model training pipelines." Five years is longer than most grant cycles and longer than many protection risks last.

7. Step 7 / 16

   Start the conversation in the private chat mode rather than a normal chat, so the text is not kept in your history and is not used for training.

   Copy this exactly

   ChatGPT: open the chat menu and choose "Temporary Chat"
   Claude: choose the Incognito chat option

   Note

   OpenAI states Temporary Chats "are deleted from our systems after 30 days", "aren't used to train our models" and do not get saved in history. Anthropic states Incognito chats are never used for model training. The cost is that you lose the conversation when you close it, which is exactly why step 12 tells you to move the draft into your own document as you go.

8. Step 8 / 16

   First prompt: do not ask for a draft yet. Ask only for a map of your evidence onto the donor's headings, so you learn immediately where you have no evidence. Copy this text exactly and fill in the two blocks at the bottom.

   Copy this exactly

   I am writing a progress report for a donor, for a human rights organisation. Below are (a) the donor's required section headings and word limits, and (b) my de-identified monitoring data.
   
   Strict rules for everything you do in this conversation:
   - Use only the facts I give you. Do not add any number, name, date, place, quote or claim that is not in my data.
   - If a section has no evidence in my data, write [GAP: what is missing] and stop. Do not fill it in.
   - Do not turn an approximate number into an exact one. Keep the words "approximately", "at least" and "fewer than" exactly where I wrote them.
   
   For now, do NOT write any prose. Give me only an outline: which of my data points belongs under which donor heading, and a list of headings where I have no evidence.
   
   DONOR HEADINGS AND WORD LIMITS:
   <paste them here>
   
   MY DE-IDENTIFIED DATA:
   <paste from REPORT-SAFE-TO-PASTE.docx here>

   Note

   The [GAP] instruction is the most valuable line in this recipe. It turns the tool's biggest weakness, filling silence with plausible invention, into a to-do list for you. Expect real gaps. Finding them two weeks before the deadline is the actual win.

9. Step 9 / 16

   Second prompt: draft one section at a time, never the whole report at once, and make it show its work by tracing every number back to your data.

   Copy this exactly

   Now draft section 1 only, in about <N> words, in plain professional English for a donor audience. Same strict rules as before.
   
   After the draft, add a section called SOURCE CHECK: list every number, date, place and organisation you used in the draft, and for each one quote the exact line of MY DATA it came from. If you cannot point to a line, mark it [INVENTED] and remove it from the draft.
   
   Mark anything you are unsure of as [CHECK].

   Note

   The SOURCE CHECK list is what makes verification take ten minutes instead of an hour. Anything the tool cannot trace back to your data is, by definition, something it made up. Work section by section, long single-shot reports drift furthest from the evidence and burn free-tier limits fastest.

10. Step 10 / 16

    Third prompt: cut to the donor's word limit without losing your caveats. Word limits are where careful language quietly disappears.

    Copy this exactly

    Cut the text below to exactly <N> words or fewer. You must keep every number, every caveat word ("approximately", "at least", "fewer than", "reported", "alleged"), and every [GAP] and [CHECK] marker. Do not add anything. Show only the shortened version.
    
    <paste your draft>

    Note

    Watch specifically for "alleged" and "reported" being dropped. If your report states as fact something your monitoring recorded as an allegation, you have created a legal and protection problem for your organisation, not a style problem.

11. Step 11 / 16

    If you drafted in Arabic or Kurdish and the donor needs English, do the translation in a SEPARATE chat, and demand translation only, not improvement.

    Copy this exactly

    Translate the text below into English for a donor report. Translate only: do not summarise, do not improve, do not add or remove anything. Keep every number and date exactly as written. If a term is ambiguous, give your translation and put the original word in square brackets after it.
    
    <paste your text>
    
    --- the same instruction in Arabic ---
    ترجم النص التالي إلى الإنجليزية لتقرير مانح. ترجمة فقط: لا تلخّص، ولا تُحسّن، ولا تُضِف ولا تحذف أي شيء. اترك الأرقام والتواريخ كما هي تماماً. إذا كان أي مصطلح غامضاً، ضع الكلمة الأصلية بين قوسين بعد ترجمتك.

    Note

    For Kurdish Sorani treat this output as a rough first pass only, and have a Sorani speaker check every sentence, see the language section for why. If the text going in is Iraqi dialect rather than MSA, expect kinship, tribal, checkpoint and land terms to come out wrong. Those are exactly the words that carry the meaning.

12. Step 12 / 16

    Move every approved paragraph out of the chat window into your own document as you go. Do not keep drafting inside the chat.

    Note

    Two reasons. Free tiers cut you off without warning and no vendor publishes the exact limit. And private or temporary chat modes are designed to disappear. Losing ninety minutes of work to a rate limit is the most common way this recipe fails in practice.

13. Step 13 / 16

    Mandatory human verification, part one: the number and name audit. Ask the tool to list its own claims, then check each one against your MASTER file yourself. This step is not optional and cannot be delegated to the tool.

    Copy this exactly

    List, as a plain numbered list, every number, date, place name, organisation name and factual claim that appears in the text below. Do not comment. Do not correct anything. Do not add anything.
    
    <paste your near-final draft>

    Note

    Take that numbered list to REPORT-MASTER-DO-NOT-SHARE.docx and tick each item against your real records. Anything you cannot tick gets deleted from the report, not softened, deleted. Expect to delete something. If you find nothing wrong on your first ever report, you checked too fast.

14. Step 14 / 16

    Mandatory human verification, part two: run the hostile-reader test again on the FINAL text, because the AI's rewriting can re-identify people you had successfully protected.

    Note

    The tool sometimes 'improves' your careful vagueness back into specificity, naming the governorate's main city, guessing a month, merging two banded figures into one exact number. Read the final draft as the security officer again, and have the colleague from step 5 read it once more. This is the step that catches damage the AI itself caused.

15. Step 15 / 16

    Decide the disclosure question before you submit, using the funder's own words rather than a rumour. Search the call document and the funder's website for "generative AI", "artificial intelligence" or "AI"; if there is nothing, email your programme officer and keep the reply on file.

    Copy this exactly

    Suggested one-line email to your programme officer:
    
    "Does your organisation require applicants to declare the use of generative AI tools in preparing an application or report? We use them for language editing and for English translation of text we have written ourselves, and we would like to follow your policy exactly."

    Note

    What we could verify in primary funder text, August 2026. EUROPEAN COMMISSION, the ERA Forum "Living guidelines on the responsible use of generative AI in research", third version, May 2026, tells funders to "Request transparency from applicants on their use of generative AI" and states that "Applicants declare if they substantially used generative AI tool(s) to prepare their application." The guidelines define "substantial" in their own footnotes: "using generative AI as a basic support tool for authors is not a substantial use", whereas "interpreting data analysis, carrying out a literature review, identifying research gaps, formulating research aims, developing hypotheses" can be substantial. The same document names "supporting non-native speakers in producing texts in multiple languages" as a legitimate, positive use. So: help writing YOUR OWN application in English is not what the disclosure rule is aimed at; having the tool generate your analysis or your aims is. NIH (United States) is the strictest we found: notice NOT-OD-25-132, effective for the 25 September 2025 receipt date onward, states "NIH will not consider applications that are either substantially developed by AI, or contain sections substantially developed by AI, to be original ideas of applicants", and caps a Principal Investigator at six applications per calendar year. OPEN SOCIETY FOUNDATIONS publishes nothing about AI on its grants pages, which is the normal situation for the funders most Iraqi CSOs actually work with, and exactly why you ask and keep the answer in writing.

16. Step 16 / 16

    Write your organisation's rule down in one page today, while the reasoning is fresh, and put it where new staff will find it. Seventy-one percent of organisations like yours have no written AI rule, which means the rule is whoever is most tired at 11pm on deadline night.

    Copy this exactly

    Our rule, in five lines:
    1. Raw monitoring data, testimony, names, exact dates and locations never go into any AI tool. Ever.
    2. Only the de-identified SAFE file goes in, and only after a second person has read it.
    3. Model training is switched off and private/temporary chat mode is on, in every account we use.
    4. Every number and name in an AI-assisted draft is checked against the master file by a human before it is sent.
    5. If a funder asks whether we used AI, we answer honestly, and we say what we used it for.

    Note

    Print it. Tape it above the desk of whoever writes the reports. This page, not the software, is what protects the people in your files.