Use AI's Mind, Not Its Memory: 3 Principles for Safer AI in Humanitarian Work

Over the past month, I've written about the structural shifts reshaping humanitarian funding - the $2 billion OCHA deal with its political conditions, the withdrawal from 66 international bodies, the "adapt, shrink, or die" reality facing organisations across the sector.

The response has been significant. But one question keeps coming up in DMs and conversations:

"We're already stretched. We're losing staff. Reporting requirements aren't shrinking. How do we actually do more with less without cutting corners that shouldn't be cut?"

Part of the answer is operational efficiency - which is why I published the Treasury Rails research on payment infrastructure last week. Cash assistance is around 25% more efficient than in-kind aid, yet its share of humanitarian spending has actually fallen from 23.3% to 17.7% over the past two years. We're doubling down on expensive modalities while cutting what works.

But there's another efficiency lever that's harder to talk about: AI.

Three months ago I wrote about Shadow AI - the informal, undocumented AI use happening across humanitarian operations at 11pm, off the books, without guidance. (If you missed it: Shadow AI in Humanitarian Work documented what I was seeing across Sudan, Syria, Ukraine and other contexts - staff using AI tools to survive, mostly in secret.)

That piece hit a nerve. Over 130 reactions, dozens of DMs, and one consistent follow-up:

"OK, I get it. But what do I actually tell my team to do differently?""

The funding crisis makes Shadow AI worse, not better. In that original piece, I noted the system was operating at roughly 26% of required funding. The $2 billion OCHA deal represents an 85% drop from peak - and that number has gotten worse since.

When teams shrink and workloads don't, people reach for whatever tools help them survive. They're using ChatGPT at 11pm because they're drowning. They're using it in secret because they're afraid to ask permission.

That's not going to stop because budgets are tight. If anything, it accelerates.

So this piece is about the other efficiency question: how do you use AI safely when you're already stretched too thin to do things properly?

After training more than 50 staff across NRC, Caritas Switzerland, Estonian Refugee Council, IRC and several smaller national organisations, we've landed on three core principles that actually change behaviour.

We call them the AidGPT Axioms.

Axiom 1: The 21-Year-Old Intern
Stop thinking of AI as magic or threat. Think of it as an unlimited supply of enthusiastic new graduates on their first humanitarian deployment.

Fast. Confident. Helpful. Eager to please.

Also: lacking context, needing supervision, fully capable of making a mess if left alone.

I use this framing in every training because it immediately clarifies what AI is good for and what it isn't.

You'd trust a new graduate with:

First drafts of reports
Summarising long documents
Reformatting content for different audiences
Translating field notes
Organising messy meeting minutes

You wouldn't trust them with:

Interpreting protection data unsupervised
Making targeting or allocation decisions
Understanding political nuance in a context they just arrived in
Final sign-off on anything with life-or-death consequences

Same rules apply to AI.

The 21-year-old intern speaks with enormous confidence. They'll hand you a beautifully formatted output and sound completely certain about it. That confidence has no relationship to accuracy. They might be right. They might be completely wrong. You won't know unless you check.

The implication: Every AI output is a first draft that needs supervision. Treat it that way. If you wouldn't let a new graduate submit something without review, don't let AI do it either.

Axiom 2: Use AI's Mind, Not Its Memory
This is the single most important thing I teach. It's counterintuitive, and it changes everything.

Here's the mistake I see constantly:

People ask AI to tell them things. "What's the food security situation in Sudan?" "What are the key findings from the latest OCHA report?"

This activates the AI's training data - its "memory." And that memory is a mess: millions of conflicting sources, academic papers weighted the same as Twitter threads, outdated information presented as current. The result is confident-sounding output with no source trail and hallucinations you can't detect.

Some tools will now search the web to answer your question. That sounds better - at least it's current information. But you don't control where it searches, which sources it trusts, or how it weighs a government press release against a Twitter thread against a three-year-old blog post. You just get an answer, confidently presented, with sources you didn't choose and often can't verify.

The alternative: Don't ask AI what it knows or what it can find. Give it documents to analyse.

Instead of: "What are the key findings from the latest IPC report?"

Do this: Upload the actual IPC report and prompt: "Summarise the key findings from this document. Only use information contained in the document. Flag anything you're uncertain about."

Same tool. Completely different risk profile.

When you use AI's mind - its analytical capability - on documents you provide, outputs are traceable, verifiable, and auditable. When you use AI's memory or let it search, you get confident-sounding claims with no way to check where they came from or why those sources were chosen.

The implication: Always upload the source material you want AI to work with. Never ask AI to "tell you about" something from its own knowledge or from the open web.

This single shift eliminates most of the risk I see in humanitarian AI use.

Axiom 3: Specificity Is Safety
The first two axioms tell you how to think about AI. This one tells you how to use it.

Vague prompt → vague output. Specific prompt → useful output.

"Make this better" is useless. The AI doesn't know what "better" means to you. "Summarise this" is almost useless. Summarise for whom? In what format? How long?

The staff who get consistently good results are the ones who invest time being specific about three things:

Role - Who should the AI be?

"You are a MEAL reporting assistant for an international NGO operating in Sudan, writing for a non-technical senior management audience."

Giving AI a role constrains its responses to relevant knowledge and appropriate tone.

Rules - What constraints apply?

"Maximum 200 words. British English. Do not infer beyond the source document. Do not invent statistics. Flag any claims that cannot be verified from the source."

Rules prevent the most common AI failures: outputs that are too long, add invented details, or confidently fill gaps with hallucinated content.

Details - What specific information should it use?

Upload the actual document. Name the target audience. Include word limits and formatting requirements.

The implication: Specificity is safety. A well-structured prompt saves hours of revision. If you spend 30 seconds on your prompt and then 2 hours fixing the output, you've got it backwards.

The Tool That Makes It Stick: The AI Workflow Card
Axioms are useless if people forget them under pressure. In the middle of a real workday - three deadlines, inbox overflowing, field teams waiting - it's easy to slip back into "I'll just throw this into ChatGPT and see what happens."

That's why we built a simple tool: the AI Workflow Card.

Before anyone opens ChatGPT, Gemini, or Copilot, they answer five questions:

What is the task? (Be specific)
What data will I use? What's sensitive? (Classify before you share)
Which tool am I using, and why? (Not all tools have the same data protections)
What are my rules and safeguards? (Role, constraints, red lines)
How will I verify the output? (Who checks, how, before it goes anywhere)

Five questions. One page. Takes two minutes.

These five questions catch more unsafe AI use than any 15-page policy document I've seen. Because they force the pause. They turn "I'll just throw this into ChatGPT" into "Wait - what data am I using? What could go wrong? How will I check this before it goes to the donor / the field team / the cluster?"

When a MEAL officer, an IM specialist and a protection officer sit together and design a workflow on that card, risk becomes concrete rather than theoretical, responsibilities get named, and managers finally get visibility into how AI is actually being used.

What We've Seen Work
Let me give you a few examples from actual training cohorts - anonymised, but real.

The sitrep shortcut

A programme team was spending 3-4 hours every week compiling situation reports from multiple field locations. Different formats, different levels of detail, some in Arabic, some in English. Someone had to read everything, pull out the key points, and write it up in a consistent format. Every week.

They built a workflow using their organisation's enterprise AI tool - one covered by a data protection agreement, unlike free tools like ChatGPT.

The process: upload all the field inputs, give it a structured prompt that specifies who it should write as, what format to use, and what constraints to follow, then generate a first draft. After that, a human manually checks every factual claim against the original field reports before anything gets sent.

Time to first draft: 20 minutes instead of 3 hours. The verification step still takes an hour - that doesn't change. Net saving: 2+ hours per week, with better consistency and a clear audit trail showing what came from where.

With staff cuts looming across the sector, that 2+ hours per week isn't a nice-to-have efficiency gain. It's the difference between a sustainable workload and burnout.

The translation trap

A protection officer was using AI to translate sensitive field notes from Arabic to English. The translations were grammatically perfect. But something was off.

"Concerns about safety" became "minor worries." "Reports of violence" became "some incidents."

The AI was softening the language - making everything sound less severe than the original. This isn't malice; it's how these tools are built. They're trained to be helpful and positive, which means they often sand down sharp edges. Fine for a customer service email. Dangerous for a protection report.

The fix wasn't a different tool. It was explicit rules in the prompt: "Maintain the severity and tone of the original. Do not soften or minimise. Flag any terms where translation may lose important nuance."

Same tool, different output. The rules made the difference.

The hallucination catch

A grants team used AI to summarise a 40-page donor report for an internal briefing. The summary was clean, well-structured, and professional. It included one line about "improved food security indicators in the northern region."

The source document said the opposite: "continued deterioration despite recent interventions."

The AI invented that claim. This is called "hallucination" - when AI generates confident-sounding statements that have no basis in the source material. It doesn't happen because the tool is broken. It happens because these models are designed to produce fluent, plausible-sounding text, and sometimes plausible-sounding is not the same as true.

They caught it because they had a verification step built into their workflow: a second person checked key claims against the source before the briefing went out. Without that step, the error would have shaped internal planning decisions.

That verification step takes time. When teams are stretched and deadlines are tight, it's the first thing to get skipped. Which is exactly when it matters most.

The Honest Limitations
I want to be clear about what AI can't do, because the hype often obscures this:

AI cannot read your mind. If you don't specify what you want, you won't get it. Unstated requirements stay unmet.

AI cannot verify its own claims. It will confidently present hallucinations as facts. Verification must be human, or use a separate AI instance as a cross-check.

AI cannot access information that isn't in its training or your uploads. It doesn't know what happened yesterday unless you tell it. It doesn't know your organisation's internal context unless you provide it.

AI cannot replace human judgment on ethical matters. Allocation decisions, protection assessments, do-no-harm considerations - these require human accountability. AI can support the analysis, but a human must own the decision.

AI cannot be held accountable. When something goes wrong, the accountability sits with the person who used the tool and the organisation that did or didn't provide guidance. "The AI told me to" is not a defence.

These principles work. But knowing them and applying them consistently under pressure are different things. That's why we built training around them.

What We're Launching: Open Cohorts
For the past year, AidGPT training has been private - commissioned by specific organisations for their teams. That's worked well, but it's left out the many people who've messaged me saying "my organisation won't pay for this, but I need it."

So we're opening public cohorts.

Responsible AI in Practice is a 3-week, 6-session course for humanitarian and development professionals who are already using AI and want to do it better.

You get the full AidGPT framework with worked examples, practice on real humanitarian workflows, your own AI Workflow Cards designed for your actual job, and a cohort of peers facing the same challenges. You'll learn as much from each other as from us. Cohorts are capped at 20 participants to keep the learning interactive.

You don't get AI hype, generic corporate training, or a PDF and a wish of good luck.

It's designed for programme staff, MEAL teams, grants and partnerships staff, and anyone handling documentation who's already using AI informally and wants to do it safely. It's not for people looking for magic solutions or organisations wanting to tick a compliance box.

Individual enrolment: €350 per person. First open cohort launches Q1 2026.

For organisational bookings or in-house training, we tailor content to your specific context and workflows - schedule a call to discuss.

Not sure where your organisation stands? We also offer a Shadow AI Health Check - a 3-4 week diagnostic that maps how AI is actually being used across your teams, identifies risks, and builds a practical roadmap from Shadow AI to Safe AI. Useful if you suspect AI use is happening but don't know the scale.

Your Teams Are Already Experimenting
Here's what I know for certain:

In six months, organisations will be operating with smaller teams and tighter budgets. The funding crisis I've been documenting isn't going away - it's accelerating.

Your staff are already using AI tools. They're using them because they're stretched, because the tools help, and because no one told them not to - or because the policy said "no" but offered no alternative. When teams shrink and workloads don't, people reach for whatever helps them survive.

The question isn't whether AI is being used in your organisation. It's whether it's being used with guidance or without.

The organisations that invest now will have documented workflows, confident staff, and fewer incidents. The ones that wait will eventually get there too - but probably after managing consequences that didn't need to happen.

AI adoption isn't separate from the adapt-or-die question facing the sector. It's part of the answer - if done safely.

Ready to join the next cohort?

📝 Join the individual waitlist 🗓️ Schedule a call for organisational training 🌐 Explore the full programme at aidgpt.org

Already using AI and want to connect with others doing the same? Join the AidGPT Community Forum - it's free, and a good way to share prompts, ask questions, and learn from real implementation stories.

📝 Join the individual waitlist 🗓️ Schedule a call for organisational training 🌐 Explore the full programme at aidgpt.org

Already using AI and want to connect with others doing the same? Join the AidGPT Community Forum - it's free, and a good way to share prompts, ask questions, and learn from real implementation stories.

Or message me directly - I read everything, even if I'm slow to reply.

Thomas Byrnes is CEO of MarketImpact, a humanitarian consultancy bridging digital innovation and field reality. His work spans systems architecture, social protection, humanitarian finance, and responsible AI - informed by 15+ years in contexts including Syria, South Sudan, Yemen, Pakistan, Ukraine, and the Philippines.

This piece is part of an ongoing series on navigating the funding crisis:

Adapt, Shrink, or Die - What the $2B OCHA deal actually means
The Fine Print - The withdrawal from 66 international bodies
Treasury Rails - Stablecoin infrastructure for humanitarian payments
Shadow AI in Humanitarian Work - The original piece on undocumented AI use

#HumanitarianAI #AidGPT #ResponsibleAI #ShadowAI #FundingCrisis #AdaptOrDie #HumanitarianInnovation #NGOTech #CapacityBuilding #AITraining