How it works — ADAFP AI Laboratory

This page exists because the AI Ethics Charter requires it. Section 4.2 commits the Laboratory to stating publicly which systems it uses, what data they are given, what the tools can and cannot do, and where the known failure modes are.

It is written to be checked rather than believed. Where something is not yet done, it says so, and the compliance status page keeps a running record.

1. What this is

A working environment for ADAFP member organisations — research, drafting, translation, analysis and training. It is not a public verification service and not an independent assessor of anyone.

ADAFP is a political alliance whose members compete against the governments its tools would assess. An accountability instrument run by a partisan actor produces findings that favour that actor's interests, not because anyone intends it, but because choosing what to examine is itself a political act. So the Laboratory serves members, and where its output is published it appears under the name of the ADAFP body that takes responsibility for it.

Access requires signing in through the ADAFP member portal. Anonymous use is refused.

2. Which systems are used, and where they run

Engine	Who operates it	Licence of the weights	Where requests go
Mistral Small 4 — the default	Mistral AI	Apache 2.0	France
Mistral Large 3	Mistral AI	Apache 2.0	France
Claude	Anthropic	Closed	United States
Gemini	Google	Closed	United States
Image generation (mood boards)	Cloudflare Workers AI, FLUX.1 Schnell	Apache 2.0	Cloudflare network

The two Mistral engines are open-weights models: the weights are published under a permissive licence, so the models could in principle be run on infrastructure ADAFP controls. Today they are used through Mistral's own French-hosted service, which means those requests do not leave the European Union. Requests to Claude and Gemini do.

Whichever engine is selected, the text you type is sent to that provider to generate a response. Do not enter anything confidential, and do not enter personal information about other people that you are not entitled to share.

3. Where answers come from

Generated from the model alone

The assistant, Document Studio, the Brand and Identity Starter, the Anthem Creator and the Concept Builder produce text from the model's own training. Nothing is looked up. These are drafting tools: the output is a starting point that a person edits, not a source of facts.

Retrieved first, then written

African Intelligence, the Claim Checker and Encyclopaedia Africana search the live web first and instruct the model to write only from what was retrieved, citing the sources by number and saying where the sources are silent. Encyclopaedia Africana lists every source beneath the briefing with a working link and a retrieval timestamp, plus a standing reference to that country's ISS African Futures profile.

This distinction matters more than any other on this page. A retrieved answer can be checked against its sources. A generated answer cannot.

Source hierarchy

Every mode is instructed to prefer sources in this order, and to say which tier an answer rests on:

National electoral commissions, parliaments, courts, ministries, official gazettes, national statistics offices
The African Union and regional organisations
Election observation missions and research institutes
Established media

Social media and unattributed material are not treated as sources.

4. Rules every mode operates under

Fact, analysis and opinion are labelled separately. A fact has a nameable source; analysis interprets it; opinion is a recommendation or political judgement.
Certainty is expressed in words — high, moderate, low — never as a percentage. A percentage implies a statistical measure the system does not have.
Where no primary source exists, the output says so rather than filling the gap.
Dates, figures, quotations, case names and citations are not to be invented. Where the model does not know, it should say so.
The same evidentiary standard applies to every party and government, including ADAFP and its own member organisations. ADAFP positions are presented as positions, never as independent factual assessment.

These instructions are embedded in the page delivered to your browser and can be read by anyone who inspects it. That is deliberate: a rule nobody can verify is not a commitment.

5. What is stored

What	Where	Who can see it
Documents and concepts you save	ADAFP's server	Only your own organisation
Unsaved drafts in Document Studio	Your own browser	Only you, until you save
Audit record of actions	ADAFP's server	ADAFP administrators
Web server access logs	ADAFP's server	ADAFP administrators

The audit log records actions, not content: which organisation saved a document and when, never what it said; that a search was run, never what was searched for. A transcript of every draft written by an opposition party, sitting on a server, would be a liability rather than a safeguard.

No account is required beyond portal membership, no advertising or analytics trackers are used, and no profile of you is built. See the privacy notice inside the platform for the full statement.

6. What this cannot do

It can be wrong, confidently. Every engine here sometimes produces fluent, plausible, incorrect text. The rules above reduce it; nothing eliminates it.
It can invent a source. A fabricated citation to a real ministry is the most dangerous failure this kind of system produces, because it defeats the checking a careful reader would otherwise do. Click the links.
It is not current. Except where a tool retrieves live sources, answers reflect training data of unknown vintage.
It is not legal, electoral or financial advice, and no output is certified or authoritative.
Its knowledge of Africa is untested. These are commercial systems trained mostly on non-African material. How accurate they are about African political institutions, and whether their errors fall evenly on governments and oppositions, has not been measured. ADAFP has published a methodology to measure it and intends to.

7. Human responsibility

Nothing is published to a public record automatically. Assessments of election results, allegations of illegality, statements about identified individuals, and interpretations of law are not published under ADAFP's authority without review by a named person.

The Promise Tracker is closed to submissions until its review workflow exists. Its editor is Chofor Che Christian-Aimé, Senior Researcher at the ADAFP Think Tank.

8. Corrections

If anything here is wrong — including if you are its subject — write to office@adafp.org. Challenges receive a substantive reply within fourteen days. Corrections are made in place and logged publicly, with the date and the nature of the change. We do not delete quietly.

Corrections go to Adil Abdel Aati, Deputy Secretary-General of ADAFP.

9. Version

This page describes the platform as at 25 August 2026. It is revised whenever the systems, the data flow or the rules change. Where the Laboratory does not yet meet a commitment in its Charter, the compliance status page says so.