The verification habit your AI policy does not mention — What six cohorts of AidGPT training have surfaced

AI disclosure: This piece predates MarketImpact's per-article disclosure standard, introduced June 2026. During this period AI tools assisted with research synthesis, multilingual data collection, and supported drafting under our verification discipline; all analysis, conclusions, and editorial judgement are the author's own. How we use and verify AI:marketimpact.org/how-we-use-ai.

What the April cohort taught us, and why the June one matters

Register for the June cohort ataidgpt.org/training. Starts Tuesday 2 June, six sessions, twenty seats, EUR 350 (EUR 280 for national NGO staff, local organisations, and aid workers between roles).Apply here.

The moment that keeps recurring

Across every cohort of this course we have run, one moment recurs. A participant takes a piece of work they have already used, or are about to use, and runs the verification technique on it.

I cannot tell you which participant or which document because we do not have permission to publish the specifics, and would not seek it for this kind of detail anyway. But I can tell you the shape of it, because the shape is identical every time.

The participant is experienced. They have used AI to draft something for a real audience: a briefing note, a presentation, a proposal section, a position paper. They are confident in the draft. They are about to send it, or have just sent it. We are in Session 2 of the course, where we teach the adversarial audit. They run the technique on the draft, not as an exercise but on the actual work they care about.

The AI finds something. A statistic invented to fill a gap the AI was not told to leave empty. A phrase lifted from a different source and presented as the participant's own analysis. A quote attributed to an organisation that, when checked, said something subtly different. A claim that sounds reasonable but that the AI confesses, under the prompt, it could not actually source.

The technique is simple. You give the AI a prompt that explicitly tells it: assume you made at least one mistake in the response you just gave. Your job is to find it. Don't defend the work. Don't hedge. Go back through line by line and point to where you invented a fact, overstated a claim, smoothed over uncertainty, or made something sound more solid than it is. Quote the exact phrase and say what is wrong with it.

The mechanism is why it works. The AI's default is to please you. When you stop talking, it stops existing, which means it has every incentive to keep producing things you nod at. By making the new task "find an error," you give it permission to keep pleasing you by surfacing errors instead of by defending its work. The same model that resists admitting a mistake when you push back becomes ruthlessly self-critical when you reframe the job.

The moment in the room is always the same. Quiet. A bit of disbelief. Then: "I had not noticed. Nobody had noticed."

That moment is what this training is actually for.

What the April cohort surfaced

In April we ran the first open cohort of the course. Participants from FAO, Concern Worldwide, Caritas Internationalis, DG ECHO and independent consultants working across EU grants, social protection, anticipatory action, and humanitarian programming. Roles from Head of Communications to MEAL Adviser. Some had been using AI tools daily for over a year. Some were near the start.

A few patterns came out of the six sessions that are worth naming, because they have shown up consistently enough across cohorts that I now treat them as findings rather than anecdotes.

Verification has to be a habit, not an extra step.

Several participants told us in the pre-course survey that they "always check every output against source material." I told them, gently, that nobody always checks. Running an AI training on the ethical use of AI, I sometimes forget to check. The realistic target is that verification becomes mechanical enough that you do not think about whether to do it, the same way you do not think about whether to spellcheck. The adversarial audit prompt is the closest thing we have to making that mechanical. Several participants reported by the end of the cohort that they had it pinned somewhere on their desktop and were running it on outputs they would previously have just sent.

The real risk is not the one the policy documents focus on.

Most organisational AI policies are written around the risk of beneficiary data leaking into a training set. That is a real risk and it should be addressed. But it is theoretical. The risk that is happening right now, today, in every humanitarian operation that has access to ChatGPT, Claude, Copilot or Gemini, is different. It is a country office team taking a 20-page food security briefing and asking the AI for a one-page summary. The summary drops a village. Or adds one. The team operates on the summary. The decision flows downstream from a fact that was not in the source document. Nobody verifies, because the source was already verified. That is the failure mode the course is built around, and it is the one most policies still do not name.

Specific agents outperform generalist ones, but cost more to maintain.

A participant who builds a donor report writer agent calibrated to one specific donor's house style will get better drafts than a participant using a generic writing assistant. But that specific agent has to be updated when the donor's template changes, when the organisation's positioning shifts, when the country context evolves. The trade-off is real. Cohorts that grasp this stop trying to build one all-purpose agent and start building a small team of narrow ones, each with a defined job.

Two AIs cross-checking each other catches things one does not.

The verification technique works in a single chat. But it works better when you start a second chat, give the second AI the source document and the first AI's output, and ask the second AI to find claims in the response that are not supported by the source. The mechanism is statistical. A single AI is unlikely to hallucinate on most well-scoped tasks, and it is very unlikely that two different AIs will hallucinate in the same way on the same content. Participants used this for high-stakes outputs by the end of the course. Several reported it had become standard practice.

The discipline shifts the work, not the workload.

A consistent observation from participants at the end of the course: AI did not make them work less. It moved their work from drafting to framing and approval. One participant put it as everyone becoming a senior exec. This is closer to the truth of what AI does to a humanitarian professional's job than either the doom framing or the productivity framing. You spend less time generating words. You spend more time deciding what should be said, what sources to anchor to, what the AI got wrong, and what is actually safe to send.

Knowledge management is the unsolved governance problem.

Most organisations treat AI governance as a policy compliance problem. Who can use what, what data goes where. The April cohort kept circling back to a different question. A MEAL manager in one country office spends two months building a set of carefully tuned agents that cut a six-hour weekly task to thirty minutes. Then the MEAL manager leaves the organisation. What happens to those agents? Are they portable to the MEAL manager in another country? Has anyone documented what they do? In nearly every cohort we have run, the answer is no. The organisation has policies on AI use. It has nothing on AI continuity. That is a gap that is going to bite, hard, in the next eighteen months.

Workflow before agents.

One of the strongest pieces of practitioner-discovered advice from the cohort was simple. Do not start by designing agents. Start by manually mapping the workflow you want to mechanise, with all the decision points and who should make each one. Only then work out which bits to automate. Cohorts that skip this step build impressive-looking agent stacks that automate steps no one should have been doing in the first place.

If you are on the fence

The full course is EUR 350 (EUR 280 for national NGO staff, local organisations, and aid workers between roles). That is a real commitment, particularly if you are paying it yourself.

If you want to test the method before deciding, we are running a 90-minute Live Intro Lab on Monday 1 June for EUR 75. There is a second lab on 23 June if the June cohort fills before you decide and you want to test the method ahead of the July one. You will work through one CHEF prompt and one verification cycle on your own material. If you continue to the full course, the EUR 75 is credited against the EUR 350. The intro lab exists for exactly this reason: it lets you make a small decision now and the big decision later, when you actually know what you would be getting.

Apply for the June cohort or register for the Intro Lab.

Why this is not your compliance training

There is a category of AI training that exists to check a regulatory box. Feed your AI policy into a vendor platform, the platform generates a module, your staff click through it, you have a record of completion, the auditor is satisfied. We have nothing against this kind of training. It is the right answer to the question it is designed to answer, which is "have we discharged our compliance obligation."

It is not the right answer to a different question. That question is whether your staff, when they use AI tomorrow morning on a real piece of work, will catch the things it gets wrong. The compliance module tells them what they should not put in. It does not teach them how to spot what the AI invented, lifted from somewhere else, or confidently misattributed. That is a different skill, and it cannot be acquired by clicking through a module.

This training is built around the second question, not the first. The compliance content is covered, because it has to be. But the work of the course is the work of producing a person who can use these tools and catch their failures. That is a different curriculum, a different format, and a different cost.

What the June cohort will be like

We refined the curriculum based on April. The method is the same. The sequencing is tighter, the agent frameworks are introduced earlier, and the Cowork session has been reworked to do the capstone work it was designed for. That is the session where chat prompts become file-aware agentic workflows running on your own documents. April showed us it deserves more time and a clearer ramp-up. June reflects that.

CHEF as the prompting backbone. Verification as the core discipline. Reusable workflows as the output. Responsible use embedded across every session, not pushed to a final compliance lecture.

Every session is instructor-led with two facilitators in the room, not pre-recorded. That matters because the moments that change how people think about AI cannot be scripted. They happen when a facilitator reads the room and adjusts.

The course is tool-agnostic. The techniques work in Claude, ChatGPT, Copilot or Gemini, because they are built on the underlying logic of how these systems behave, not on the surface features of any one of them. You will use Claude during the sessions because the cohort licence supports it, but everything you learn transfers.

You do not need to be technical. You need to understand what the tools do well enough to manage their use, in your own work and in the work of your team. The analogy I keep coming back to is the football manager. A manager does not need to be the best striker on the pitch. They need to understand what a striker does, what good and bad performance look like, when to substitute, when something has gone wrong on the field. A manager who has never played the position cannot do the job. Most senior people in this sector are about to find themselves managing teams that use AI daily, and the skill they need is the manager's skill, not the striker's.

What this means in practice is that the participants who get the most out of the course are willing to test the methods on their actual work. The techniques only land when you run them against documents you care about. The participants who got the most out of April brought real material into the sessions. The ones who got less brought hypothetical examples.

You will leave with a personal CHEF Prompt Catalogue, at least one documented Workflow Card, and verification habits that work across any AI tool. The course includes all session recordings, a completion certificate, a Claude Pro licence for the duration, and an optional follow-up call. We also run 30-day and 60-day check-ins, because training without sustained practice structures is a six-week injection of confidence followed by a six-week slide back to where you started. I wrote that in April and I stand by it.

For organisations

If you want this for your team rather than as individuals, we scope dedicated cohorts separately, adapted to your tools, workflows and policies. Previous private cohorts have been delivered to NRC Sudan, Caritas Switzerland, the IRC, ACF UK, and the Estonian Refugee Council.

The team format works particularly well when an organisation wants to embed AI use as a shared practice rather than as individual skill-building. The Workflow Cards and Prompt Catalogues that participants build become collective resources rather than personal ones. That is how the knowledge-management problem I named earlier starts to get solved.

Discuss a private team cohort or email info@marketimpact.org.

The details

Dates: 2, 4, 9, 11, 16, 18 June 2026, Tuesdays and Thursdays.

Time: 16:00 East Africa Time, 15:00 CEST, 14:00 BST, 09:00 EDT.

Duration: Six 90-minute sessions over three weeks.

Price: EUR 350 per person. EUR 280 for staff from national NGOs, local organisations, and aid workers between roles.

Cohort size: Capped at 20.

Apply at aidgpt.org/training.

Why now

I have made the case for why this matters across five previous pieces in the AidGPT series. This one is about what we learned when we actually did it.

The participants from April who got the most value were not the ones who built the most impressive agents during the sessions. They were the ones who, the following week, ran the verification technique on something real, found something they had not expected, and changed what they did next.

The course does not end with Session 6. It starts with what you try on Monday morning.

aidgpt.org/training

Tom

This is the sixth instalment in the AidGPT series on responsible AI in humanitarian and development practice. Earlier pieces covered Shadow AI, the Liar's Dividend, the three principles, the ACF UK evaluation data, and local AI. All are on the newsletter archive.

Tom's Aid and Dev Dispatches is a weekly newsletter on humanitarian and development trends, read by 9,000+ subscribers. If someone forwarded this to you and you'd like the next one in your inbox, you can subscribe on LinkedIn.

#HumanitarianAI #AidGPT #ResponsibleAI #AITraining #HumanitarianTech #DigitalTransformation