Ethiopian Civil Society Engagement Program · Phase II
Everything given out on the day, in the order it is used. Readable on a phone, and it works offline once loaded.
One day
This page holds every sheet handed out during the day. Each tab is one session, in the order they happen. If you lost a sheet, or want to use one at work next week, it is here.
Everything in the exercises is invented. No real names, no beneficiaries, no sources, no budgets, no case files. Nothing real goes into any tool at any point.
That is not a formality. It is what makes it safe to be a beginner in a room full of your peers.
| 1 | These tools invent things, and you will not notice unless you check. |
| 2 | You can still get real work out of them — if you brief them properly and trace every claim. |
| 3 | Some things must never go near them. And the reason is not the name — it is the combination. |
| 4 | Somebody has to own this, and it is you. |
| Session | |
|---|---|
| 1 | What this day is, and the one rule |
| 2 | How these tools work, and how they lie |
| 3 | Asking well, and writing something real |
| 4 | What to trust it with, and what not to |
| 5 | What must never go in, and the policy you need |
| 6 | Checking what you are sent |
In the last session you fill in a one-page sheet — Our rules for using these tools — for your own organization, while the people who can agree it are in the room with you. Ten boxes, real names in them, small enough that your staff will actually follow it.
Under tab 5 that page sits next to a twenty-two-page minimum viable AI policy you can adapt afterwards, with a table showing which box on the sheet becomes which section of the policy. Do the page first. The policy is only the writing-down.
Session 2
Two sheets. Read both, and mark anything in the paragraph the record does not support.
Sheet 1 — the record
Everything here is invented for training. This is the only factual record. Anything that does not appear here has not happened.
Sheet 2 — what the AI wrote from it
An AI tool was given the record on Sheet 1 — and nothing else — and asked to write a short programme update for the KCDF website.
Since its launch in February 2026, the Kora Youth Livelihoods Initiative has become one of the most effective youth employment programmes operating in Kora City. Working across four woredas in partnership with the Nordvik Foundation and the Tesfa Development Fund, KCDF has enrolled 412 young people and confirmed that 94 of the 168 graduates are now earning an independent income — a 56 per cent improvement on their situation before enrolment. Attendance has been strongest where the programme has had a stable base: in Gudeta woreda, repeated relocation of the course site during March drove participation down sharply, a problem KCDF has since addressed. "This programme has changed what young people in Kora believe is possible for them," said Hirut Gemechu, KCDF's youth programme officer. Ato Dawit Lemma, head of the Kora City Job Creation Bureau, described the initiative as a model the Bureau intends to replicate in neighbouring woredas next year.
Session 3
Most people send nine words and then wonder why they got a paragraph about somebody else's programme. This session is what the nine words were missing, and then writing something real with the missing pieces put back.
The prompt
Write a paragraph about our youth programme for a donor report.What it produced
"Since its launch, our youth programme has empowered more than 500 young people across the region, equipping them with the practical skills and confidence they need to build sustainable livelihoods. Through a combination of vocational training, mentorship and seed funding, participants have gone on to establish small businesses in sectors ranging from tailoring to food processing. An impressive 78 per cent of graduates report improved economic circumstances within six months of completing the programme, and demand for places continues to grow year on year. Working closely with local government partners and community leaders, we remain committed to expanding our reach and deepening our impact in the years ahead."
Not the 500, not the 78 per cent, not the tailoring, not the local government partners, not the six months. None of it came from anywhere.
| 1 | Too short to answer. You left out the facts, so it invented facts. This is the one everybody does |
| 2 | The fake expert. "You are a world-class grant writer with twenty years' experience…" does not make it more accurate. It makes it more confident and more promotional — the two things you least need |
| 3 | Everything at once. A report and a social post and some slides. It picks a length, an audience and a tone for you, and it picks promotional |
| 4 | No rules. You never said what it may not do, so it invented. It was not disobeying you |
| 5 | No check. "I'll read it through and see if it looks right." It always looks right. That is the one thing it is genuinely excellent at |
Every one of them is one of those failures turned the right way round. Leave one out and you will see it in what comes back.
1 — Who it should be
A useful angle, not a fake expert. And tell it what it does not know, or it fills the gap from everything it has ever read.
You are helping a programme officer at a small civil society organization. You are not an expert on this. You only know what I give you below.2 — What it needs to know
The minimum. Not everything you have. This is the decision that matters, and it is not about typing — once it is in there, it is gone. You do not paste the file in and ask it to be careful. You choose.
Here are the only facts you may use: …3 — What you want
One thing, one reader, one length. Leave this out and it picks all three for you, and it picks promotional.
Write one paragraph of about 120 words for … The reader is …4 — The rules
What it may not do, and where to put what it does not know. The MISSING line is the one worth stealing — it gives the machine somewhere to put its uncertainty other than into your paragraph.
Use only the facts above. Do not add any number, name, date or outcome that is not in them. If something important is missing, do not fill it in — put a line at the end starting MISSING: and say what you would need. Plain English, no promotional language.5 — What you will check
The one everybody skips, and the one that does the most work. It makes the thing show its working — your check stops being read it carefully and hope and becomes tick off a list.
After the text, list every factual claim you have made, and next to each one quote the words from my facts that it came from.What it gives back
KCDF began the Kora Youth Livelihoods Initiative on 3 February 2026, with a two-year grant from the Nordvik Foundation. The programme operates in four woredas of Kora City and enrolled 412 young people aged 18 to 29 in its first intake. Participants complete a twelve-week skills course; those who finish and submit a business plan receive a start-up grant of 15,000 birr. By 4 August 2026, 168 participants had completed the course and received a grant. The Kora City Job Creation Bureau provides the training venue in Lensa woreda at no cost to the programme.
MISSING: I have no information on what happened to participants after receiving the grant, on the 244 enrolled participants who had not completed by 4 August, or on any outcome, income or employment results.
That line found the honest hole in the data, and nobody asked it to.
| Claim | From your facts |
|---|---|
| Began 3 February 2026 | "started 3 February 2026" |
| Two-year grant, Nordvik Foundation | "funded by the Nordvik Foundation" |
| Four woredas of Kora City | "runs in four woredas of Kora City" |
| 412 enrolled, aged 18–29 | "412 young people aged 18–29 enrolled" |
| Twelve-week course, 15,000 birr grant | "twelve-week course… start-up grant of 15,000 birr" |
| 168 completed by 4 August 2026 | "By 4 August 2026, 168 had completed" |
| Venue at no cost | "provides a venue at no cost" |
It is shorter and duller. Good donor writing is duller. The first one was better writing and worse reporting — and seven claims with seven sources can be checked in under a minute. The first paragraph could not be checked at all.
Turn your screen to the person next to you and let them read it. They are not scoring it. One question only:
"What is missing that would stop this working?"
Most people type six words and hit enter, and everything that goes wrong afterwards started right there. This is the habit worth building.
Everything above used what KCDF has actually published. Now the same thing, for a real job — with published material and a real funder's call.
| A | A concept-note section — problem statement and objectives, 250 words, for the funder below |
| B | A public statement — 180–220 words, marking the first cohort finishing |
| C | A story pitch — 150 words to the editor, plus how you would stand it up |
What KCDF has published
Published on the KCDF website, 12 August 2026. All of this is public.
KCDF launched the Kora Youth Livelihoods Initiative on 3 February 2026 with a two-year grant from the Nordvik Foundation.
The programme runs in four woredas of Kora City — Lensa, Bulcha, Anfaro and Gudeta.
412 young people aged 18–29 enrolled in the first intake.
Participants take a twelve-week skills course. Those who complete it and submit a business plan receive a start-up grant of 15,000 birr.
By 4 August 2026, 168 participants had completed the course and received a start-up grant.
The Kora City Job Creation Bureau provides the training venue in Lensa woreda at no cost.
Hirut Gemechu, KCDF youth programme officer: "The twelve weeks are the easy part. What we do not yet have is a way to keep supporting people once the grant is spent."
KCDF has not published an evaluation of this programme. No income, employment or outcome figures have been published.
For A and B — the funder
| Closes | 30 September 2026 |
|---|---|
| Grant range | USD 40,000 – 90,000 over 18 months |
| Stage | Concept note. No annexes, no budget, no logframe |
| Required section | Problem statement and objectives, 250 words maximum |
Priority. Programmes that already have a delivery track record and can show a credible plan for supporting participants after the initial grant is disbursed.
Scoring. Problem definition 30% · Objectives and measurability 30% · Track record 20% · Sustainability 20%.
For C — the newsroom
The press release, 12 August 2026 — everything above, plus one line: "KCDF is seeking funding to extend the Initiative into a second phase."
The commissioning note, from the editor:
"We have covered youth unemployment twice already this year. If we run this at all, it needs an angle we can stand up. Pitch me something by Thursday — 150 words, and tell me how you would verify it before I put a reporter on it."
The newsroom has not seen KCDF's internal review. It has no income data and no attendance figures. Write the pitch from that position, because that is the position you would really be in.
| 1 | Can you point at where every number came from? Same figure, same units, same period |
| 2 | Has any word been upgraded? Reported becoming confirmed. Some becoming most. This is the one everybody misses |
| 3 | Has anything appeared that was not in the source? A partner, a quotation, an official, a comparison |
| 4 | Is anything in here true but unpublished? Knowing something is not the same as being allowed to publish it |
| 5 | Whose name goes on it? A person, not a job title |
Three marks per sentence, and there is no fourth: keep — you can point at the source · qualify — it is nearly right, say it more carefully · remove — it is not there.
Back at your desk, the best material to practise on is something your own organization has already published — your website, an annual report, a press release. It is real, it is safe because it is already public, and what comes back will actually be useful to you.
If you have nothing published to hand, use the invented organization below. None of it is real. There is no such organization and no such town.
An organization to practise on
| Who | Debre Hana Community Health Association, in fictional Debre Hana town. Founded 2019. Nine staff |
|---|---|
| What it does | Maternal health outreach in six kebeles |
| Funded by | The fictional Abay Trust. A three-year grant ending 31 March 2027. It is the only funder |
| Last twelve months | 2,400 women reached. 41 community health volunteers trained |
| Outreach sessions | 78 held, against a target of 90 |
| Volunteer retention | 74% this year. It was 61% the year before |
| The hard part | Two of the six kebeles are more than two hours from the main road in the rainy season |
| Announced 3 Aug 2026 | A new arrangement with the Debre Hana Health Office to share cold-chain storage |
| People | Tsedale Alemu, Executive Director. Yohannes Bekele, programme lead |
| Not published | No evaluation has ever been published. No figures on health outcomes exist |
Rough meeting notes — for practising notes into actions
"Monday, everyone there. Yohannes says the two far kebeles will be cut off again from June, we should front-load those sessions before the rains — agreed. Still no decision on whether we run a second volunteer intake, depends on the Abay Trust extension. Tsedale to write to the Trust this week. Somebody needs to update the volunteer register before the audit, nobody said who. Fuel costs are up again and we keep not budgeting for it properly. Next meeting the 19th."
A formal passage — for practising summaries or a translation check
"Implementing partners shall submit a narrative and financial report within thirty (30) days of the end of each reporting period. Where an implementing partner is unable to meet this deadline, it must notify the Trust in writing before the deadline falls, stating the reason and proposing a revised date. The Trust may, at its discretion, grant an extension of up to fifteen (15) days. Failure to report without prior notification may result in the suspension of subsequent disbursements."
Deliberately full of musts, mays and conditions. That is where summaries and translations break.
Session 4
That is the whole session. It does not know which of the two you have asked for — it answers both the same way, in the same confident voice, at the same speed. You are the only one who knows which kind of job it is, and that is a decision you make before you open the tool, not after you read the answer.
| Re-shaping | You hand it something and ask for it in a different shape — shorter, structured, in slides, in plainer words. The material is yours. It is moving it around |
|---|---|
| Producing | You ask for something you did not give it — a figure, a fact, a law, an outcome, a comparison. It has nothing to move around, so it composes |
Decide for yourself before you read the right-hand column.
| The job | Why | |
|---|---|---|
| "Turn these meeting notes into a list of actions, with who and by when." | TRUST | It is all in the notes. Tell it to write UNASSIGNED where the notes name nobody — then it finds the gaps in your meeting instead of filling them |
| "Add up this column for me." | DO NOT | It does not calculate. It produces a number shaped like the answer. Sometimes right, and you cannot tell which time |
| "Turn this press release we already approved into three social posts." | TRUST | Approved material, changing shape. This is what it is genuinely good at |
| "Write a paragraph about the impact of our programme." | DO NOT | You did not give it any impact. So it will supply some, and it will be well written |
| "Fix the English in this paragraph I wrote." | TRUST | The meaning is yours; it is touching the surface. Read it after — it sometimes fixes the grammar and changes the claim |
| "What does the law say about registering an NGO?" | DO NOT | Nothing was looked up. It composes, including the article numbers and the name of the proclamation |
| "Here is a forty-page funding call. What is the deadline and the word limit?" | TRUST | Pulling a specific fact out of a document you handed it is one of the safest things it does. Then look at the page and confirm it |
| "Find me the latest youth unemployment figures for this region." | DO NOT | Nothing was looked up. And if it gives you a source, click it — the source can be invented too |
| "Summarise this six-page policy for staff." | CAREFUL | The answer is in there — but summarising is exactly where rules get softened and clauses go missing. See below |
| "Translate this statement into our language for publication." | CAREFUL | A first draft, never the last step. See below |
These are the difficult ones, and they are difficult for the same reason. The answer really is in what you gave it — but things fall out on the way through, and what comes back looks complete.
A summary that deleted the rules
A seven-clause internal communication protocol went in. This came out:
"KCDF staff may use any digital tool to prepare communications as long as a director approves the final version. Personal information should generally be minimized. Translations can be published after automated grammar checking. If an error occurs, staff should quietly update the website."
Two of the seven clauses vanished completely. Must became should generally. Approved tools became any digital tool. Fluent human review of translations became automated grammar checking — reversed, not softened. Two levels of approval collapsed into a director. And records the incident, corrects it publicly, informs affected people became quietly update the website.
The subtlest one: the approval requirement survived — it moved. The policy says get approval before the tool is used. The summary says a director approves the final version, after everything has already gone in. It reads like a control and it prevents nothing.
A caption with a number that was not a typing mistake
| Training site | Planned | Reached | % of plan |
|---|---|---|---|
| Lensa | 120 | 114 | 95% |
| Bulcha | 110 | 99 | 90% |
| Anfaro | 95 | 87 | 92% |
| Gudeta — site 1 | 60 | 31 | 52% |
| Gudeta — site 2 | 45 | 23 | 51% |
| Mobile outreach | 70 | 58 | 74% |
| TOTAL | 500 | 412 | 82% |
"Mobile outreach reached only 74% of its target, the weakest performance of any delivery site and a marked decline on the previous quarter. Overall the programme reached 76% of planned participants across all sites — a strong result given the operating environment."
58 out of 70 is 83%, not 74%. Two things about that error matter more than the error: it makes performance look worse, which is why nobody would ever challenge it — you check the numbers that flatter you — and it produces the wrong decision. It makes mobile outreach look like the weak point when the real problem is Gudeta at 52% and 51%.
| 76% | The average of the six percentages — (95+90+92+52+51+74) ÷ 6 |
|---|---|
| 82% | The percentage of the totals — 412 ÷ 500 |
Averaging percentages with different denominators is the most common numerical error in programme reporting. Mobile outreach's 70 planned places get the same weight as Lensa's 120. It is invisible to anyone reading the caption, and it is made constantly by people who are perfectly good at arithmetic.
Two more inventions, and both are the same mechanism: "the weakest performance of any delivery site" is false on the table's own figures, and "a marked decline on the previous quarter" — there is no previous quarter in the table.
Machine translation is a first draft, and a genuinely useful one. It is not an approval step and it never has been.
These tools are far weaker in Ethiopian languages than in English — and they will not tell you so. There is no warning and no change of tone. A fluent, confident-sounding translation is not evidence of a good translation. Treat a clean result with suspicion rather than relief.
| What is wrong | It says something the original does not say |
|---|---|
| What is merely awkward | It is clumsy. A style problem, and the least important |
| What is dangerous | A name rendered wrongly · a not that disappeared · a number that changed · a rights or legal term that now means something else |
Most people reviewing a translation spend their time on the middle row, because awkwardness is what you notice. The third row is what you were actually asked to check for.
What a reviewer is checking
meaning · must, may and should · names, dates and titles · technical and rights terminology · tone and possible harm · and who signed off
The boring half of your week. Minutes and meeting notes. Reformatting the same information for three different funders. Turning an approved document into slides. Nothing on that list needs it to know anything — you already have the answer, and you are asking for a different shape.
The one instruction that makes notes-into-actions safe
If an action has no owner or no date in the notes, write UNASSIGNED or NO DATE. Do not guess.Then look at what comes back. "Draft the intake criteria — UNASSIGNED" is not the tool failing. That is the real gap in your meeting, found in nine seconds — and it is exactly the item that goes missing in real life.
Two things to check every time: nothing promoted from a concern into a decision, and nobody who was not in the room has been given a job.
These are the worksheets. If you did the rounds on the day, this is what you worked from. If we ran the shorter version, everything here was covered above — this is the material done properly, with the answers on the worked-examples sheet.
Round 1 · Long documents. Read the summary first, then the policy. Mark anything the summary changed.
The policy it was summarising
1. Urgent allegations may be published only after a designated reviewer confirms the source record and separates verified fact, allegation, and unknown information.
2. Public communications must use the minimum personal information necessary. Consent to publish one fact does not imply consent to publish other personal details.
3. Drafts may be assisted by approved digital tools only after data classification. Red information must not enter a public AI tool. Amber information requires approval and documented minimization.
4. Translations must be reviewed by a fluent human against the source for meaning, omissions, names, terminology, tone and potential harm. Machine translation is never the final approval step.
5. Quotations, legal references, statistics, dates and institutional titles must be checked against the authoritative source. If a source cannot be confirmed, the claim must be removed or clearly qualified.
6. The Communications Director approves routine statements. The Executive Director approves high-risk allegations, statements naming a person or institution, and corrections involving potential harm.
7. If an error is published, KCDF records the incident, corrects the public version promptly, informs affected people where appropriate, and reviews the workflow that allowed it.
Go clause by clause. Two of the seven are missing from the summary altogether.
Round 2 · Numbers. The table and the caption are above. One cell does not follow from its own row, and the caption says 76% where the table says 82% — neither is a typing mistake.
Round 3 · Your own language. Translate these two into your own language, then have someone read the result back in English.
"Hirut Gemechu, KCDF youth programme officer, said: 'The twelve weeks are the easy part. What we do not yet have is a way to keep supporting people once the grant is spent.'"
"By 4 August 2026, 168 of the 412 participants had completed the twelve-week course and received a start-up grant of 15,000 birr."
Between them these carry a personal name, an institutional title, a number, and a quotation whose meaning depends on tense. That is precisely where translation fails.
Round 4 · Notes and slides. Turn these notes into decisions and actions.
"Programme team, Tuesday. Talked about the second intake — Almaz thinks we should start in Bulcha and Anfaro first because the sites are settled, everyone agreed. Still no decision on Gudeta, need to see if the new venue holds. Finance said the grant transfers are slow, Kebede to chase the bank this week. Someone should draft the intake criteria before the board meets — nobody said who. Hirut raised that we still have no follow-up after the start-up grant and we keep not dealing with it. Next meeting the 14th."
Or take the published KCDF brief from Session 3 and ask for a slide outline — a heading of eight words or fewer, three bullets at most, and one line to say that is not on the slide.
All four rounds, done properly — the repaired summary, the corrected caption with both figures explained, the six-point translation check, and a clean actions table.
These are the answers. If you want to try the rounds yourself, do them first.
Session 5
Some of these are perfectly fine to put into a public AI tool. Some must never go near one. Decide for each: okay, or not okay.
Take two minutes and decide before you read the answers. Ten are okay and ten are not — and several of them split every room they have been put in front of.
| Would you put this into a public AI tool? | |
|---|---|
| 1 | Your organization's annual report, already published on your website |
| 2 | The list of women who attended last month's outreach, with their names and kebeles |
| 3 | A funding call you downloaded from a donor's website |
| 4 | The draft budget for a grant you are still negotiating |
| 5 | A paragraph you wrote yourself, to fix the grammar |
| 6 | A recording of an interview with someone who asked not to be named |
| 7 | A proclamation published in the Negarit Gazeta |
| 8 | The same attendance list, with the names deleted |
| 9 | A press release you issued last week |
| 10 | A colleague's medical certificate, to summarise for the file |
| 11 | Notes from a routine programme meeting, with no personal details in them |
| 12 | A photograph of participants at a community meeting |
| 13 | A photograph from a public event you already posted on Facebook |
| 14 | Your organization's email password |
| 15 | An invented case study you made up for training |
| 16 | A CV somebody sent you when they applied for a job |
| 17 | A published government or UN statistics report |
| 18 | Your unpublished internal evaluation, to summarise for the board |
| 19 | Where a person you support is currently staying |
| 20 | A slide outline built from a report you have already published |
Decide first.
| 1 | Already public. You cannot disclose what you have already disclosed |
| 3 | The donor published it. It is not yours to protect |
| 5 | Your own words, nobody else in them. Read what comes back — it sometimes fixes the grammar and changes the claim |
| 7 | Public law. Check what comes back against the text — it will paraphrase confidently |
| 9 | You published it yourself |
| 11 | Routine and impersonal is fine. Read it once more before you paste — meeting notes are where one name or one figure hides in a single sentence |
| 13 | Same photograph as number 12, opposite answer. You already published it. The difference is not the image, it is what was agreed |
| 15 | You invented it. Nobody can be harmed by a person who does not exist |
| 17 | Published. Check the figures that come back against the report |
| 20 | Published material changing shape. This is the safe half of the tool |
They are not all wrong for the same reason. The reasons are the useful part.
2 · The attendance list with names and kebeles
Named people who never agreed to it. The moment you paste, a copy exists on a company's servers you do not own, in a country you did not choose, kept for as long as their retention rules say. If that company is breached, or is ordered to hand it over, those women are identifiable — and they never heard of the company.
You cannot un-send it. There is no button.
4 · The draft budget you are still negotiating
It is commercially sensitive — the funder is on the other side of that negotiation.
And it is not solely yours. A budget has salaries in it, which are personal data about your own staff. You would not email it to a stranger.
6 · The recording of someone who asked not to be named
They said no. That is the whole answer.
And the audio file carries more than the voice — the file name, the date, sometimes the phone number, sometimes location data. A voice is biometric. Transcription is a disclosure like any other.
8 · The same list, with the names deleted
This is the one that matters most. Most people say this is okay, because deleting names feels like anonymising.
Deleting names does not anonymize anything. What identifies people is not the label, it is the combination — a place, a date, an activity and a small population. In a kebele, "the women who attended the maternal health session on the 12th" is a list of names to anybody who lives there.
Here is a summary a tool produced from a record after the name was taken out:
"A woman who is the only night pharmacist at the clinic beside the Kora bus depot witnessed uniformed personnel remove two people at exactly 22:17 on 9 July. Her brother works in city-command procurement. She has been having panic attacks since."
Who is this? No name was used. That paragraph is a workplace, a shift pattern, an address and a family member — and it puts her brother in it too, and nobody ever asked him anything.
10 · A colleague's medical certificate
Health information is the most protected category of personal data in almost every law there is — and this one belongs to somebody who works for you, which means they are in no position to object.
Ask yourself whether you would be comfortable if they were standing behind you while you pasted it.
12 · A photograph of participants at a community meeting
Faces are personal data, and they did not agree to this.
And a photograph carries far more than the faces — the building, the vehicles, the street, who was standing next to whom, and very often GPS coordinates and a timestamp inside the file itself. You are not uploading an image. You are uploading a place, a time, and a list of who was there.
14 · Your organization's email password
Never. No tool, no reason, no form. It goes into a chat log that may be retained and may be read by a person.
If it has already happened, change it today.
16 · A CV somebody sent you for a job
Their personal data, given to you for exactly one purpose. Applicants have the least power to object and the most to lose by complaining.
And a CV is dense — home address, phone, date of birth, previous employers, sometimes a photograph. It is one of the most identifying documents that will cross your desk this year.
18 · Your unpublished internal evaluation, for the board
People say this is okay because the board is internal. Who reads the output does not change what you are pasting. The board is inside your organization. The tool is not.
And an unpublished evaluation almost always has monitoring data in it — which means it has people behind it, and figures nobody has verified.
19 · Where a person you support is currently staying
Location and movement is the category that gets people hurt. Everything else on this list is a privacy problem. This one is a safety problem, and it is the only item here where the harm is physical and immediate.
And it cannot be undone.
Most people have heard that it trains on your data and then tells somebody else. That is the weakest reason on this page — it is possible, it is rare, and if it is your whole argument you are exaggerating. Here is what actually happens.
| It is on somebody else's computer now | A copy exists on servers you do not own, in a country you did not choose. This is the whole harm, and it needs nothing else to be true |
|---|---|
| A person may read it | Conversations get flagged for safety review and staff read flagged material. Routine and documented. "Nobody is interested in me" is not a security control |
| It probably is used to improve the model | On free and consumer accounts, assume yes unless you specifically turned it off in the settings. Paid business accounts usually do not, by default |
| You broke a promise, and that is a fact rather than a feeling | Somebody consented to confidential internal use. A foreign commercial service is a party they never agreed to. The harm exists the moment you paste, whether or not anything ever leaks |
| These tools remember | Many keep context between conversations. Something entered in January can appear in an unrelated answer in June |
Your staff are making this decision every day, without you, right now. What rule does your organization have about it — not what it should have, what it has?
If the answer is nothing, that is the normal answer. You have not been slow — this arrived faster than any policy process could. But your staff are using these tools today, which means you already have a rule. Nobody wrote it down and nobody agreed it.
| A good answer | A bad one | ||
|---|---|---|---|
| 1 | Who owns these rules | One person, named. The one who notices when it is out of date | "Management" |
| 2 | Who says yes when it is not obvious | One named person, reachable within a day | "The relevant department head" |
| 3 | What anyone may do without asking | Concrete and small — "fix grammar in something we wrote" | "Low-risk tasks" |
| 4 | What needs a yes first | The genuinely borderline cases for your organization | A list of nine |
| 5 | What nobody may do | The absolute lines, in plain words | "Anything unethical" |
| 6 | One sentence on what never goes in | Short enough for a new staff member to repeat from memory | A paragraph with three clauses |
| 7 | Who to tell when it goes wrong, and how fast | A name and a number of hours — "tell Almaz, same day" | "Report through the appropriate channel" |
| 8 | Which tools you have agreed to use | Actual product names, and who may add to the list | "Approved tools" |
| 9 | When you will look at it again | A date on the calendar | "Regularly" |
| 10 | When you tell your audience AI helped | The cases where you will say so, decided now rather than argued later | "Where appropriate" |
Real names, not job titles. Not the management, not the communications team. A person. If you write a job title, what you have actually written is that nobody is answerable.
Small enough to be true. Two things in each column, not ten. A rule nobody follows is worse than no rule, because it tells your staff that the rules are decoration.
You are not writing this from a blank page. Every one of those ten things is already a section of this template — what you decide goes straight into it.
The Word version is the one to edit. The PDF is for reading on a phone.
Session 6
In March 2023 somebody typed "Will Smith eating spaghetti" into a video generator and posted the result. It was grotesque — the face changed between frames, the hands fused with the fork, and the spaghetti obeyed its own gravity. It went viral for being terrible, and it became the informal test everybody used to measure how good these tools were getting.
| March 2023 | ModelScope. Grotesque. Viral for being terrible |
|---|---|
| May 2025 | Veo 3. Faces, motion — and synchronised sound |
| February 2026 | Seedance 2.0. Passed. "Even the kitchen background is convincing, down to the outlets in the walls" |
| 2026 | Kling 3.0. Not eating — sitting at a table, holding a conversation with another person |
Asked how they tell, most rooms produce the same answers. Here is what has happened to each of them.
| "The hands" | Fixed. It was the standing joke of 2023 and has not been reliable since about 2024 |
|---|---|
| "The face looks wrong" | Most people who watch the 2026 clip cold say it is real. Including people who work in media |
| "Something feels off" | That feeling is real, and it is not evidence. You cannot put it in a story, or a statement, and you cannot defend it when somebody asks how you know |
| "It's too perfect" | Then a slightly worse one gets past you. Compression, a bad crop and a phone re-recording hide anything you were going to notice anyway |
| "A detector website said so" | The same mistake wearing a lab coat. A number — 82% likely manipulated — with no method, no error rate and no explanation of what it measured. You would not accept that from a consultant |
Ethiopia, June–December 2025
Researchers monitored AI-generated videos targeting women in Ethiopia. Most of them were generated in Amharic. They were built to exploit cultural expectations around modesty and tradition, and to turn individual women into scapegoats. Victims and their families initially believed the content was real.
So the plan cannot be "we will report it and it will come down." That plan has already been tested here, in your language, and it failed.
| Ghana | A fabricated journalist — with her own social media profile and her own website — used to push false claims about Ghanaian leaders into African news media. The fake was not a video. The fake was a reporter, and real newsrooms carried what she wrote |
|---|---|
| West and Central Africa | "Ghost reporters" built on the identities of dead people, spreading pro-Russian propaganda |
| Burkina Faso | Videos of Ibrahim Traoré shaking hands with world leaders, berating the IMF, and addressing the Pope in fluent English |
Everything above is among the least likely things to be used against your organization.
| A real photograph with a false caption | Nothing generated. Nothing to detect |
| A real video from three years ago, posted as this morning | Nothing generated |
| A crop that removes whoever was standing on the other side | Nothing generated |
| Audio | Voice cloning now needs about three seconds of source, scraped from a Reel or a voicemail. There is no uncanny valley in a phone-quality voice note, and nobody inspects audio the way they inspect video |
| A screenshot of a document that never existed | Ninety seconds in any image editor |
Not one of these gets weaker as the models get better. That is why they are the ones worth learning.
| 1 | Where did it come from? How old is the account, what else has it posted, who posted it first |
| 2 | Is there an independent copy? One that would still exist if the first account vanished. Three accounts carrying a clip is one source copied twice |
| 3 | Can you get the original file? Not a repost — the file. Usually nobody has even asked |
| 4 | Does anybody else place it? Another angle, a longer version, somebody who was there |
| 5 | What would have to be true for this to be real? If it shows a meeting — was there a meeting? Who else would have been in the room? |
| 6 | Who benefits from you believing this, this morning? And why now |
You are not trying to decide whether it is fake. You are trying to decide what you can stand up.
The first question has no answer you can defend. The second one is your actual job — it is what you did before any of this existed, and it will still work when the models are perfect.
08:20 this morning
A social-media account posts a 14-second video which appears to show the director of a local organization accepting cash from a government official. It is spreading among local accounts. The phone is already ringing.
| 1 | The account. Created two days ago. No biography. Eleven posts. The profile picture also appears on several unrelated accounts |
| 2 | The video. 14 seconds, heavily compressed. Face visible for about four seconds. Audio says "this is our agreement." Lip movement partly obscured. No file information came with the repost |
| 3 | The room. A conference venue used by many organizations. A sign is partly visible: "…Rights Forum 2025." |
| 4 | Your own records. The director did attend a public forum at that venue in November 2025. Staff do not recognise the words in the clip |
| 5 | Other accounts. Two others reposted it. Both trace back to the original account. No independent copy, no longer version |
| 6 | A detection tool. Returns "82% likely manipulated." No method, no error rate, no explanation |
| 7 | Context. The director has received threats. Sharing the clip further — even to deny it — puts it in front of people who have not seen it |
Work the three columns first — what we know, what we are guessing, what we do not know at all.
Your own records say the director was at that venue. It feels like the evidence firming up, and most people move it into the first column as though it supported the allegation.
It does nothing of the sort. It confirms only that the setting is plausible.
Not decide whether it is real. Nobody can do that today, and that is an acceptable position to be in.
| Preserve | Screenshot it, note the account, the time, the URL. It may be deleted within hours |
|---|---|
| Ask for the original | The file, not a repost. Nobody has asked yet |
| Say what you know | "We are aware of a video. We are looking into it." That is true, and it is enough for this morning |
| Protect the person | They have had threats. Ask them what they need before you decide what to publish |
| Do not amplify | A loud denial shows it to everyone who has not seen it. The instinct to fight it in public is the one to resist |
Sixteen prompts, ready to copy
One for each of the jobs this room actually does. Tap a prompt to open it, replace anything in [square brackets], and paste it in.
These prompts are safe. What you paste into them may not be. Every one has a bracket where your facts go. Before you fill it in, ask the one question from Session 5 — am I allowed to give this to a stranger? If it is a list of people, an unpublished source, a case file, staff records, a budget under negotiation or anyone's location, no prompt can fix that.
When: You have an approved project brief and a call for proposals, and you need the first section.
You are helping a programme officer at a small civil society organization draft a section of a concept note for a funder. You are not an expert on this organization or this work. You only know what I give you below. Here are the only facts you may use: [paste your published project brief] And here is what the funder has asked for: [paste the relevant part of the call — word limit, what the section must cover, how it is scored] Write the problem statement and objectives section, no longer than the word limit in the call. The reader is an assessor who has never visited and will read forty of these. Use only the facts above. Do not add any number, name, date, place, outcome or comparison that is not in them. If something important is missing, do not fill it in — put a line at the end starting MISSING: and say what you would need. Write objectives that can be measured — if a fact needed to make an objective measurable is not in what I gave you, say so rather than inventing a target. Plain English. No promotional language. After the text, list every factual claim you have made, and next to each one quote the words from my facts that it came from.Check this in what comes back
What it will not do. It cannot tell you whether your project is a good fit for this funder. It has not read their other grants.
When: The numbers are done and you need the words around them.
You are helping a programme officer write the narrative section of a report to a funder. You are not an expert on this organization or this work. You only know what I give you below. Here are the only facts you may use: [paste your agreed figures for the period, and any activities you can evidence] The reporting period is: [dates] Write the narrative section, [N] words, for a funder who has not visited. Cover what was done, what changed, and what did not go to plan. Use only the facts above. Do not add any number, name, date, place, outcome or comparison that is not in them. If something important is missing, do not fill it in — put a line at the end starting MISSING: and say what you would need. Do not describe an outcome unless I have given you evidence of it — activities and outcomes are different things. Where something went wrong, say so plainly. Plain English. No promotional language. After the text, list every factual claim you have made, and next to each one quote the words from my facts that it came from.Check this in what comes back
What it will not do. It cannot decide what is material enough to report. That is a judgement about your relationship with the funder.
When: You need a page that lets a board decide something without reading forty.
You are helping an executive director prepare a short written brief for a board of trustees. You are not an expert on this organization or this work. You only know what I give you below. Here are the only facts you may use: [paste the source material — a report, a proposal, a set of options] The decision the board has to make is: [state it in one sentence] Write a brief of no more than one page, structured as: the decision · what you need to know · the options · what I recommend and why · what happens if we do nothing. The readers are trustees who are not close to the detail. Use only the facts above. Do not add any number, name, date, place, outcome or comparison that is not in them. If something important is missing, do not fill it in — put a line at the end starting MISSING: and say what you would need. Keep the recommendation separate from the facts, and label it as a recommendation. Where there is genuine uncertainty, say so rather than smoothing it. Plain English. After the text, list every factual claim you have made, and next to each one quote the words from my facts that it came from.Check this in what comes back
What it will not do. It does not know your board. Tone and what can go unsaid are yours.
When: You have rough notes and nobody will read them again unless you turn them into something.
You are helping turn rough meeting notes into a clear list of actions. You are not an expert on this organization or this work. You only know what I give you below. Here are my notes from the meeting: [paste your notes] The people in the meeting were: [first names or roles only] Produce two things. First, a list of decisions that were made. Second, a table of actions with three columns: what, who, and by when. Use only what is in my notes. **If an action has no owner or no date in the notes, still list it, and write UNASSIGNED or NO DATE** — do not guess who or when. If something reads like a discussion rather than a decision, leave it out of the decisions list. Do not invent an action that was not discussed. At the end, list anything in my notes you were not sure how to classify.Check this in what comes back
What it will not do. It cannot tell what was agreed and what was merely said warmly. Check anything that felt ambiguous in the room.
When: Something has been published or agreed, and you need to say it publicly.
You are helping a civil society organization draft a short public statement. You are not an expert on this organization or this work. You only know what I give you below. Here are the only facts you may use, and all of them are already public: [paste the approved announcement, press release or published decision] Write a public statement of [N] words for a general audience, including the media. It should be accurate enough that a journalist quoting it in full could not make us look dishonest. Use only the facts above. Do not add any number, name, date, place, outcome or comparison that is not in them. If something important is missing, do not fill it in — put a line at the end starting MISSING: and say what you would need. Do not name any person, partner or government body that is not named in my facts, and do not describe any relationship that is not described there. No quotations unless I have supplied the exact words and who said them. Plain English. No promotional language. After the text, list every factual claim you have made, and next to each one quote the words from my facts that it came from.Check this in what comes back
What it will not do. It cannot judge what is safe to say publicly in your context. That is yours, and it is not a language problem.
When: The document is out and you need three or four posts from it.
You are helping draft short social media posts for a civil society organization. You are not an expert on this organization or this work. You only know what I give you below. Here is the published document. Everything in it is already public: [paste the published text] Write four different posts, each under [N] characters, each drawing on a different part of the document. One should be plain and factual, one should quote the document directly, one should ask the reader a question, and one should point to where the full document can be read. Use only the facts above. Do not add any number, name, date, place, outcome or comparison that is not in them. If something important is missing, do not fill it in — put a line at the end starting MISSING: and say what you would need. Every claim in every post must be traceable to the document. Do not add hashtags, statistics or calls to action that are not supported by it. Do not use urgency language that the document does not support. After the posts, for each one quote the sentence from the document it came from.Check this in what comes back
What it will not do. It does not know what your audience reacts to, and it does not know what is sensitive to post this week.
When: You know what you need to say and you are struggling with how to say it.
You are helping draft a professional email. You are not an expert on this organization or this work. You only know what I give you below. Here is the situation, and these are the only facts you may use: [paste or write out what has happened, using no real names — write "the funder", "our partner", "the programme"] What I need from them is: [state it] Write an email of no more than [N] words. It should state the situation, say what I am asking for, and propose a next step with a date. Use only the facts I have given. Do not apologise more than once. Do not add commitments, deadlines or offers I have not made. Do not soften the ask to the point where it is unclear what I want. Plain professional English. At the end, state in one line what you understood me to be asking for. If that does not match what I meant, my facts were unclear.Check this in what comes back
What it will not do. It does not know the history between you. Warmth, and what is better left unsaid, are yours.
When: A proclamation, policy or report is long and public, and your colleagues will not read it.
You are helping summarise a long public document for colleagues who have not read it. You are not an expert on this organization or this work. You only know what I give you below. Here is the document. It is public: [paste the document, or the section you need] Write a summary of no more than [N] words for a colleague who needs to know what this means for our work. Then, separately, list anything in the document that creates an obligation for an organization like ours. Use only the document. **Keep the strength of the original wording** — if the document says "must", do not write "should"; if it says "may", do not write "will". Where the document says something is required only in certain circumstances, keep the circumstances. Do not resolve ambiguity in the document; point it out. At the end, list any place where you were unsure whether something was required, recommended or merely permitted.Check this in what comes back
What it will not do. It is not legal advice, and a summary is not a reading. For anything consequential, someone opens the original.
When: You have a contract, a donor agreement or a proclamation, and half the terms are guesses.
You are helping someone read a document written in formal or legal language. You are not an expert on this organization or this work. You only know what I give you below. Here is the document, or the part I need to understand: [paste it] My organization does this: [one line] List every term, phrase or clause in this document that someone outside the profession would not be sure about. For each one, give the plain meaning **as this document uses it**, and one line on why it matters to us. Use only the document. **If this document defines a term differently from how it is normally used, go with the document** and say that is what you have done. If a term is genuinely ambiguous here, say so rather than choosing a meaning. Do not give legal advice and do not tell me what to do. At the end, list any term you were not confident about, and any place where the document appears to contradict itself.Check this in what comes back
What it will not do. It cannot tell you whether the terms are reasonable, or whether you should sign.
When: A proclamation, directive or regulation has come out and your team needs to know what it means for them.
You are helping summarise a new law or directive for colleagues who will not read the original. You are not an expert on this organization or this work. You only know what I give you below. Here is the text. It is public: [paste it] Our organization: [one line on what you do and where] Write two things. First, a plain summary of no more than [N] words. Second, a list of everything in it that creates an obligation for an organization like ours — what we must do, by when, and what happens if we do not. **Keep the strength of the original wording.** A *must* stays a must; a *may* stays a may. Keep every deadline, and keep every exception — *"must, unless"* is not *"must"*. Do not tell me what is likely to be enforced in practice, or what other organizations usually do; you have no way of knowing. If the text is unclear about whether something applies to an organization like ours, say so rather than deciding. At the end, list anything where you were unsure whether an obligation applies to us, and quote the wording that made it unclear.Check this in what comes back
What it will not do. It does not know how this is enforced where you are, and it is not legal advice.
When: A contract has been renewed, a policy redrafted, or a draft has come back from someone else.
You are helping compare two versions of the same document. You are not an expert on this organization or this work. You only know what I give you below. Here is the earlier version: [paste it] Here is the newer version: [paste it] List every difference that changes meaning, grouped as: something added · something removed · an obligation made stronger or weaker · a number, date or name changed · wording changed but meaning the same. Quote both versions for every difference you report. Ignore formatting and pure style changes unless they change meaning. Do not tell me which version is better. **If a whole clause is missing from the newer version, put it at the very top of your answer** — removals are far easier to miss than additions. At the end, say how confident you are that you found everything, and which parts were hardest to line up.Check this in what comes back
What it will not do. It cannot tell you why something changed, or whether the change was deliberate.
When: A contract, an MoU or a set of minutes has come back, and you need to know whether it says what you agreed.
You are helping check whether a document reflects what was agreed in a meeting. You are not an expert on this organization or this work. You only know what I give you below. Here is what I recorded as agreed — **use roles, not names**: [paste your notes] Here is the document that came back: [paste it] Produce three lists. One: things in the document that match what I recorded. Two: things in the document that were not in what I recorded. Three: things I recorded that are not in the document. Use only the two texts. **Do not assume something was agreed because it sounds reasonable.** Where the document says something similar but not the same, put it in list two and quote both versions. Do not soften a mismatch to make it sound smaller than it is. At the end, say which of the three lists you are least confident about, and why.Check this in what comes back
What it will not do. It does not know what was actually said in the room. It only knows what you wrote down.
When: You have a translation and you need to know where to look before it goes out.
You are helping check a translation against its source. You are not an expert on this organization or this work. You only know what I give you below. Here is the original: [paste the source text] Here is the translation: [paste the translation] Compare them and produce a list of differences worth a human looking at. Group them under: meaning changed · something added · something dropped · a name, number, date or title that differs · a word that is stronger or weaker than the original. Do not rewrite the translation. Only report differences, and quote both versions for each one. Do not comment on style unless the style changes the meaning. **If you are not confident about the target language, say so at the top rather than guessing.** At the end, state how confident you are in this comparison and why.Check this in what comes back
What it will not do. It is often weak in Ethiopian languages and it will not reliably tell you so. Treat a clean result with suspicion, not relief.
When: You have the content agreed and you need it laid out.
You are helping turn an approved piece of writing into an outline for a short presentation. You are not an expert on this organization or this work. You only know what I give you below. Here is the approved text. Nothing may be added to it: [paste the approved summary] I have [N] minutes and the audience is: [who] Produce a slide outline. For each slide give a heading of no more than eight words, at most three bullets, and one line of what I should say that is not on the slide. Use only the facts above. Do not add any number, name, date, place, outcome or comparison that is not in them. If something important is missing, do not fill it in — put a line at the end starting MISSING: and say what you would need. Do not add examples, statistics or transitions that are not in my text. If a slide would be thin, leave it thin rather than filling it. Suggest where a slide should be cut if there are too many for the time. At the end, say which slides you had least material for.Check this in what comes back
What it will not do. This is one of the things it is genuinely good at. Use it here early, and free your attention for the things it is bad at.
When: You are preparing to talk to someone and you want to think before you are in the room.
You are helping prepare questions for an interview. You are not an expert on this organization or this work. You only know what I give you below. Here is what I already know, from published sources only: [paste what is public] I am speaking to: [a role, not a name] What I am trying to find out is: [state it] Suggest [N] questions, ordered so the conversation can build. Mark which ones are open and which are checking a specific fact. Then separately list what I would still not know even if all of them were answered. Do not write leading questions. Do not assume anything that is not in what I gave you. Do not suggest a question that puts the person at risk to answer, and say so if you think one might. At the end, list any assumption you made about the situation that I did not tell you.Check this in what comes back
What it will not do. It cannot judge risk in your context. Whether a question is safe to ask is your call, not its.
When: You have something and you need to say what it is and what it would take.
You are helping a reporter write a short pitch to an editor. You are not an expert on this organization or this work. You only know what I give you below. Here is what I actually have, and all of it is published or on the record: [paste it] Here is what I suspect but cannot yet show: [write it out separately] Write two things. First, a pitch of no more than 150 words that states what I could report today. Second, a verification plan: what would need to be established, in what order, who I would approach first, and what I would do if the first step failed. **Keep what I have and what I suspect completely separate.** The pitch may only contain what I can already show. Do not present a suspicion as a finding. Do not suggest naming anyone who has not been approached. If the story does not stand up on what I have, say so. At the end, state plainly whether what I have is enough to report anything today.Check this in what comes back
What it will not do. It cannot assess legal risk, source safety, or whether your editor will run it.