Methodology
Every fact carries
its confidence with it.
An assessment of what it takes to adopt an AI tool in humanitarian and development work, not just a list of what exists. Every rating here is contestable. This page is how you contest it.
Scope
The register covers named AI tools and projects in humanitarian and development work. Staff use of general-purpose chatbots such as ChatGPT or Claude for everyday work (verification, handling sensitive data, and handing over judgement) is out of scope for now and may become a future section.
Evidence stage and conditions for use
Entries are not all the same kind of evidence, and the register says which kind each one is. Evidence stage separates tools in use, pilots and prototypes, proposals, models or code with no deployment shown, and entries where deployment could not be established. These support different conclusions, so filter by stage before drawing one.
Each entry also shows four conditions for use: local knowledge, reliable data, ownership and sustained capacity. They are read from the entry's own fields, and "not stated" means the source is silent, not that the condition is met. Where an entry states how data is handled (for example, that it stays on the local machine, or that it stores health data), that appears on the entry too. Verifying a tool's output and deciding who controls it are not yet assessed; most sources say nothing about them.
The core rule
Every assessed field is paired with a basis. Nothing is published without one.
Documented
The source states it directly. The evidence quotes where.
Inferred
Reasoned from what the source states, with the reasoning written out.
Verified
Confirmed by the implementing organisation, with a name and a date.
Unknown
Not stated, and never guessed.
Inferred does not mean weaker. It means a judgement was made, and the judgement is visible enough that you can disagree with it.
What every tool is assessed on
Four practical questions, asked the same way every time.
Connectivity it needs
What the tool needs, not what the country has — last-mile differences go in the note, not this field.
Compute it needs
Electricity is deliberately not a tool attribute — power is a property of the context; compute profile is what you match against it.
Skill floor — split on purpose
A chatbot a displaced person uses with no training may still need an engineer to keep it running. One number conflates the two and flatters the tool.
Cost — three fields, because one number hides the question
What the end user pays
What the person using it pays.
What it costs to adopt
What an organisation pays to join or integrate a service someone else runs.
What it costs to replicate
What it would take to stand this up independently, under your own control. This is the localisation question in one field: where it diverges sharply from the adopt cost, the tool creates dependency. The register records that; it doesn't judge it.
These come apart sharply. Google Flood Hub is free to the end user, cheap to adopt, and effectively impossible to replicate. A source that says "free of charge" is almost always describing what the end user pays; reading it as the cost to adopt or replicate is a common error.
How entries are produced
Finding
Tools are found in published reports, directories, research and public announcements, and through suggestions from practitioners.
Assessing
Each tool is checked against a public source, described in our own words and rated on the questions above.
Keeping current
Entries are re-checked regularly. A tool is only marked discontinued on clear evidence.
Verifying
When the organisation behind a tool confirms an entry, it is marked verified with their name and the date.
Field notes
Names
Where a source never names a tool, the entry is titled from its description and marked described, not named.
Tool link and source link
The tool link is where you get or use the tool; the source link is what describes it. Where there is no tool link, we say so rather than substitute one.
Type of AI
Categories follow the Humanitarian Leadership Academy and Data Friendly Space practitioner survey, so what is built can be compared with what people use.
Models and code only
Published models with no known humanitarian deployment are listed but kept out of the counts, so they never inflate what the sector has done.
Language
The languages a tool states it supports are kept apart from the languages spoken where it operates.
Status
Going quiet is not the same as closing. Without clear evidence of closure, a tool is marked quiet, not discontinued.
Ranking humanitarian needs
Countries are ranked by people in need relative to population, with refugee-hosting countries included. How the ranking works →
What this doesn't measure, and where it's weakest
Not adoption
This is a supply-side record: what's been built, what it would take to run it — not what practitioners actually use. The same HLA/DFS survey found 75% of humanitarians using AI regularly, while fewer than one in ten work somewhere it's widely integrated. Read this register alongside that finding, not instead of it.
Mostly single-sourced
Most entries derive from one source; cost bands are overwhelmingly inferred rather than confirmed by implementers. Coverage reflects what's published — tools used by national organisations that never generate a report are systematically invisible here, a finding about the literature, not the sector.
One researcher, no second coder
Assessment is currently by one person. There is no inter-rater check. Every entry carries a correction link for exactly this reason.
Sources, acknowledgements and rights
This register is a not-for-profit initiative. It publishes assessments of other organisations' work and does not own the underlying material. Every entry names its own source, and the fields below record who owns it.
source_document | The document the assessment was made from |
source_publisher | Who published it |
source_rights_holder | Who holds copyright |
source_licence_note | Terms, and what may be redistributed |
How sources are credited.
Most entries are credited to the project's own public page: this register
found each project, checked its facts against that page and wrote its own
description from it. Each entry links to that page, and the organisation that runs
the project keeps the rights to its own material. Entries drawn from a published
report name that report, for example Artificial Intelligence for Syria:
applications within the Syrian humanitarian response, March 2026, published by
Elrha with ACAPS Middle East Analysis Hub (rights: Elrha and ACAPS).
Every entry was checked against a public source in October 2026. Entries for which no public source could be found were removed; entries only partly confirmed say so in their source line. Where an entry is a commercial product or rests on the provider's own claims, the entry says so.
Acknowledgement. Some projects were first identified in NetHope's Directory of AI-enabled humanitarian projects (WS4, March 2026, with the UK Humanitarian Innovation Hub). This register does not republish that directory's text.
One entry, the Financial Report Reviewer, is built by this register's maintainer. It is listed like any other tool, assessed on the same basis, given no preferential ranking, and marked with a disclosure wherever it appears.
Commercial tools. Tools that are sold or licensed for a fee are listed when they are built for, or used by, humanitarian or development organisations, and are marked "Commercial product". They are assessed on the same basis as every other entry; where a price is not published it is shown as unknown rather than guessed. No listing is paid for, nothing is ranked higher for being commercial, and the register carries no affiliate links.
Writing a report is not owning it. Commissioned research belongs to the commissioning organisations regardless of who drafted the text. If you own material that appears here and want an entry changed or removed, use the correction link on that entry or the submission form; the entry will be changed or taken down and the change recorded.
When this will be retired
An experiment, not an institution. If it turns out not to be useful, it will be closed rather than left to go quietly stale — the data archived openly, with a note saying when it stopped and why. That commitment is here because the failure this register exists to document is precisely the one where a tool stops being maintained and nobody says so — several entries in it are already discontinued, and none of their publishers announced it. It would be hard to make that argument while running something on the same terms.
Corrections
Every entry carries a correction link. If an entry about your organisation is wrong, say so and it will be changed. Every change is logged with the date, the evidence and who raised it. A correction from the organisation that built the tool moves the field's basis to verified, with their name and the date recorded; a correction from anyone else, or from published material, moves it to documented. Entries can be unattributed or withdrawn on request where publication creates a protection risk.
Version. This method is versioned, and every change to it is logged.