Building Àdàbà AI: Ready for Testing – October 2026 Progress Update | DigiCivic

What does a survivor of online abuse need from an AI tool before she ever meets a lawyer?

 

What should happen when she tells her story the way memory delivers it, out of order, rather than the way a court expects it?

 

And who should decide which AI model is good enough to work with survivors: the engineers who build the system, or the lawyers who will rely on it?

 

These questions ran through October’s work on Àdàbà AI.

 

If September was the month the project moved from consultation to construction, October was the month the separate parts began working as one system. The survivor’s report, the organisation’s case management, the lawyer’s review and the AI layer underneath were connected for the first time, and the platform is now in testing. Early in November, DigiCivic’s Executive Director, Mojirayo Ogunlana, also explained the thinking behind the tool to an audience of justice-sector leaders and law students. We start there, because it explains why this tool is being built at all.

 

Why build a tool here, for here

 

At the Law for Humanity Summit 2026 in Abuja on 3 November, convened by the Public and Private Development Centre (PPDC) under the theme “Justice in the Age of Data and AI”, Mojirayo spoke on a panel about closing the justice gap with data and AI.

 

Her argument was simple. Many of the digital platforms people in Nigeria use every day were, in her words, “developed for Nigeria”, but “not developed for the people of Nigeria”. They do not always understand local context. She gave an example from her own experience: a light-hearted post of hers using the word “donkey” was removed by a global platform’s automated moderation, because the system read it as abusive. When she looked into it, she found the word can carry an abusive meaning in another country’s usage. The tool understood the word but not the context.

 

Her point was not that technology is the problem. It was that communities with their own data and their own context need tools built around that context. Àdàbà AI is DigiCivic’s attempt to do exactly that for one specific harm: technology-facilitated gender-based violence (TFGBV), meaning abuse that uses digital tools such as social media, messaging apps, fake accounts or AI-generated images to threaten, harass or expose someone.

 

Àdàbà is the Yoruba word for dove, a symbol of peace. It is the image behind the platform’s design.

 

 

What Àdàbà AI is designed to do

 

For readers new to this series (start with our first update and the August update), here is the short version. When a survivor reports online abuse, one of the biggest obstacles to justice is evidence: what to collect, how to preserve it, and whether a court will accept it. Àdàbà AI is a free prototype designed to help with that groundwork:

 

  • A survivor describes what happened in her own words, by typing or by voice, and uploads supporting material such as screenshots, documents, images or audio.
  • The system helps organise that account and the evidence, and prepares a first draft of supporting documents, such as an affidavit (a written statement sworn as true), for a lawyer to review.
  • The case is then assigned, through a partner organisation, to a verified lawyer, who stays responsible for every legal judgment.

 

Àdàbà AI is not an AI lawyer or a legal advice chatbot, and it does not deliver justice on its own. It is a survivor-support and evidence-organisation tool: it helps survivors tell what happened and organise their evidence, so that a qualified lawyer can review the case.

October’s milestone: the full journey works

 

Until October, Àdàbà AI existed as separate pieces: designs, a backend (the part of the system users never see), a knowledge base and an AI layer. This month they were connected.

 

The core application is now deployed. Survivors, organisations and lawyers each sign in separately, with access matched to their role. Organisations can invite staff and lawyers by email. Lawyers complete their profiles before they are verified and assigned cases. Survivors can submit a report, give consent, and attach PDFs, images and audio. Organisations can assign cases, and lawyers can see the cases assigned to them.

 

The AI layer is connected too. Documents can be added to the knowledge base, and the system can retrieve them when it answers. In early checks it pulled the right material from the documents it had been given, the first sign that the core technical approach works.

 

Survivors don’t tell their stories in a straight line

 

The most instructive moment of the month came during a demonstration of the platform late in October.

 

The intake was designed to let a survivor narrate her experience freely, in her own words. During the demo, one of the project’s advisers raised a concern: people recounting traumatic events often do not tell them in order. Memory after trauma can be fragmented. A form that insists on a tidy sequence from the first question can make an already painful account harder to give.

 

But a lawyer does need a sequence. Dates and the order of events matter in court.

 

The team’s answer is to separate the two. The survivor tells her story first, in whatever order it comes. Then a new layer of guided follow-up questions, which the team calls a question wizard, helps her clarify details and place events in order, including dates. This builds on the AI-generated timeline introduced in September, which survivors and lawyers can already correct by hand.

 

The principle is worth naming: let people speak first, and structure afterwards. The system takes on the work of organising the account, rather than asking the survivor to organise her trauma for the system.

 

Lawyers, not only engineers, are testing the AI

 

Choosing an AI model is often treated as a purely technical decision. Àdàbà AI is treating it differently.

 

The platform now includes a comparison screen that shows two AI models’ answers to the same question side by side. The legal team is using it to put survivor-centred and lawyer-centred questions to each model and judge which answers are accurate, relevant and appropriate. A model can perform well on technical benchmarks and still respond poorly to a frightened survivor or to a precise legal question. The people best placed to tell the difference are the lawyers who will use it.

 

No model has been chosen yet. That decision will follow testing.

 

For now, AI is used lightly and deliberately: to summarise case reports, help categorise cases, and support drafting. Anything the AI generates is labelled clearly as AI, so no user mistakes a machine’s draft for a lawyer’s conclusion.

 

A knowledge base built in chapters

 

The legal knowledge base the AI draws on is now organised into chapters by jurisdiction and topic. Nigerian law and data protection law, for example, sit in separate chapters.

 

This matters for two reasons. It helps the system retrieve the right material for a particular situation, rather than searching everything at once. And it makes the tool easier to adapt elsewhere: an organisation in Kenya, for instance, could leave out the Nigerian chapters and add its own. The knowledge base is being treated as a living document, growing as testing shows where the gaps are.

 

There are now two layers: a shared knowledge base, built from material collected across the programme and available to all participating teams, and a private knowledge base for each organisation’s own documents.

 

A change of direction, stated plainly

 

In September’s update we wrote that the team had begun building the system’s document search, known as retrieval-augmented generation or RAG (a method where the AI looks up relevant documents before it answers), on Amazon’s cloud services.

 

In October, the team changed course. The search now runs in a shared testing environment (a sandbox) provided through the programme. This avoids separate cloud costs and gives the team access to the shared knowledge base, instead of maintaining a duplicate system. The application itself continues to run in the cloud.

 

Plans change as a project learns. We would rather say so than let an earlier post stand uncorrected.

 

Data protection: from assessment to policy

 

September’s update said the Data Protection Impact Assessment (DPIA), the structured review of what personal data the system handles and what could go wrong, still needed to become a finished privacy policy. In October, the DPIA was completed and Àdàbà AI’s privacy policy was updated.

 

Testing, in small circles first

 

Testing has begun, in stages.

 

DigiCivic’s own staff are testing first, from different locations, including Jalingo, Lagos and Abuja, so the team can see how the platform behaves on different connections and in different settings. This is still under way. Invitations have also gone out to survivors, lawyers and partner women’s rights organisations for the next phase.

 

Three choices shape how that testing is designed:

 

  • Small groups before large networks. Each group is asked to find gaps and missing functions before any wider release, so problems are fixed while they are still small.
  • Separate sessions for survivors. Survivors and lawyers are tested in different sessions, so survivors can speak openly and lawyers can focus on legal accuracy.
  • A fixed testing window. Testers can work at their own pace within a set period, which respects their time and keeps feedback arriving while it can still shape the build.

 

The reasoning behind the timing is simple: bugs found early are cheaper and easier to fix than bugs found after launch.

What we haven’t solved yet

 

As in previous updates, here is what remains open:

  • The AI model. No model has been selected. Lawyer-led testing will inform the choice.
  • Timeline accuracy. The AI-generated event timeline is connected, but its accuracy has not been fully tested.
  • Messaging. The channel for lawyers and survivors to message each other has been built but not fully tested.
  • Drafting without a template. When there is no sample or guide for a particular document, the drafting feature currently falls back on the AI model’s general knowledge. Given this project’s commitment not to invent law, that fallback needs careful testing, and lawyers will review every draft.
  • Case law. The shortage of accessible Nigerian judgments described in September has not gone away.

Why this matters beyond one organisation

 

Two lessons from October may be useful to other teams building tools for survivors.

 

First, design intake around how trauma is actually told. Asking for a chronological account up front puts the burden on the person least able to carry it. Collecting the story freely and structuring it afterwards, with the person’s help, is fairer and likely to produce a fuller account.

 

Second, let the end users judge the AI. Side-by-side comparison by lawyers, using real survivor-centred and legal questions, tests what matters in practice, not only what scores well on paper.

 

Looking ahead

 

November’s focus is completing internal testing, fixing what it finds, running testing with survivors, lawyers and partner organisations, building the guided follow-up questions into the survivor intake, and moving towards a choice of AI model. The team is also preparing a product demonstration and technical documentation for review by Code for Africa.

 

Àdàbà AI is still a prototype. In October, it started to work as one system and went into the hands of its first testers.

 

Organisations interested in partnering on Àdàbà AI or supporting its next phase can contact us at contact@digicivic.org.

 

Àdàbà AI is developed by DigiCivic Initiative as part of the A+ Alliance Gender & AI Innovation Collective, supported by Code for Africa and Women at the Table. This post is published under a Creative Commons Attribution 4.0 International licence (CC BY 4.0). Reuse and adaptation are encouraged with attribution.

Leave a Reply

Your email address will not be published. Required fields are marked *

Stay Informed. Stay Empowered.

Get the latest digital rights news, research, advocacy updates, and opportunities from DigiCivic Initiative delivered straight to your inbox.

Your privacy matters. Read our Privacy Policy