What does it mean for a legal AI to say “I don’t know,” honestly, when the law itself barely has an answer?

 

What happens when the people you built a tool for finally see it, not as a wireframe, but as something close to real?

 

And how do you keep a system that recommends legal categories from quietly replacing the lawyer’s judgment it’s meant to support?

 

These were the questions running through September’s work on Àdàbà AI.

 

If August was the month of foundations, data protection frameworks, a RAG architecture decision, wireframes taking shape, September was the month those foundations met their first real tests. Survivors sat in the room. A genuinely difficult data problem surfaced. And the team made a small but telling decision about what to call a button.

 

Closing the loop: survivors at the table

 

Àdàbà AI’s August progress post was explicit about something the team hadn’t done yet: survivor consultation hadn’t happened. That changed on September 11, when the team held its first direct session with survivors to walk through the design.

 

The response was, by the team’s account, strongly positive. Survivors responded well to the platform’s privacy controls and its lawyer-client confidentiality model, a “need-to-know” approach that restricts who can see what. Voice-to-text testimony input was well received, easing the burden of typing out a difficult account. There was real interest in the platform’s reach: lawyers accessible across nearly all Nigerian states, not just one city or region. Several survivors expressed genuine eagerness to test the platform once it’s built, and the team is already planning how to bring survivors who couldn’t attend into future rounds of consultation.

 

This matters beyond the “we did the thing we said we’d do” logic. Consultation isn’t a single event to check off. This was one round, with mid-fidelity wireframes, and the team is treating it as the start of survivors acting as core testers through the platform’s development, not a one-time validation stamp.

 

A small relabeling, a real design principle

 

One of the more instructive moments in September wasn’t a big technical milestone. It was a naming decision.

 

The lawyer dashboard originally had the AI’s case categorization labeled simply as “triage.” During review, the team’s design advisor pushed back: the label made it sound like the AI was making the call. The team relabeled it “AI triage review” — a small change, but a deliberate one, meant to keep the human lawyer’s role visible in the interface itself, not just in a policy document somewhere else.

 

The same instinct shaped a second change to the dashboard that month. An ambiguous “evidence review” checkmark, which had no clear end state since evidence review is often ongoing, was replaced with a “case reviewed” checkmark, marking specifically that a lawyer has confirmed they understand the case, not that every possible evidentiary question has been resolved.

 

Neither change altered what the system actually does. Both changed what the interface tells the lawyer they’re responsible for. That distinction, between a system being accurate and a system being honest about the limits of its own authority, is a design principle worth naming on its own, and one other teams building legal-adjacent AI tools would do well to borrow.

 

The hardest problem so far: a legal system with almost no accessible case law

 

The most significant, and most honest, finding of the month came out of the effort to build Àdàbà AI’s legal knowledge base.

 

To ground its outputs in real law rather than invented text, the system needs access to actual Nigerian court judgments, particularly on cases involving online and technology-facilitated gender-based violence. The team found that, right now, only one High Court judgment is readily, publicly accessible for this purpose. Most High Court decisions in Nigeria are simply unreported, and many relevant cases never reach a formal judgment at all. Court of Appeal and Supreme Court decisions are reported and legally binding, but cases involving online gender-based violence rarely make it that far up the court system. Regional court judgments, from bodies like the ECOWAS Court, carry persuasive weight but aren’t binding on Nigerian courts.

 

The only real way to close this gap, the team concluded, is paid research assistance: people physically retrieving judgments from court archives, the way past projects have done. That requires funding the current project doesn’t yet have.

 

This is worth sitting with rather than glossing over. It means Àdàbà AI’s legal knowledge base, at least for now, will lean more heavily on published legal articles, statutes, and regional or international instruments than on a deep well of Nigerian case precedent, not because the team chose that, but because that precedent largely doesn’t exist in accessible form. The system is being designed accordingly. Rather than promising comprehensive case law citation, the AI’s role is to help lawyers find relevant legal instruments and reasoning, whether Nigerian, regional, or international, and flag where Nigerian law is genuinely silent, so a lawyer can build a cross-jurisdictional argument rather than assume one doesn’t exist.

 

Data protection: built as a foundation, not bolted on after

 

Building on August’s data protection framework, September saw the team formalise a Data Protection Impact Assessment (DPIA), a structured review of what personal data the system processes, what could go wrong, and what rights are at stake. The preliminary version covers ten risk areas, including AI hallucination and the risk of unauthorised evidence disclosure.

 

A few decisions are worth flagging as, again, reusable by other teams:

 

• The DPIA is explicitly not the privacy policy itself. It’s the analysis that informs the policy, kept as a separate document so the reasoning behind each privacy commitment stays visible rather than getting flattened into legal boilerplate.
• Evidence integrity work happens on copies, not originals, preserving an unaltered original while allowing annotation and review on a working copy. A straightforward but important safeguard for material that may end up in court.
• The team intends to publish the DPIA and resulting privacy policies openly once finalised, treating rigorous privacy documentation as a demonstrable strength, not a compliance afterthought to be minimised.

 

Nigeria’s Data Protection Act was specifically noted by the team as well suited to this kind of work, described as tightly scoped and contextually relevant, comparable in rigor to GDPR, which gives the project a solid legal floor to build on rather than having to invent data protection standards from scratch.

 

What actually got built

 

Underneath the design and policy work, September was a genuinely productive month on the technical side.

 

Backend authentication, the system managing who can log in and what they can access, was substantially completed, unlocking real registration and login on the development server. The survivor dashboard gained an AI-generated event timeline, mapping the sequence of an incident from a survivor’s testimony, with both survivors and lawyers able to correct or add missing events by hand. The lawyer dashboard took shape around case overviews, a workload snapshot, and a new review queue, a single list consolidating every pending task across a lawyer’s cases, with overdue items flagged automatically, so a lawyer isn’t required to click through each case individually to find what needs attention. Evidence review moved to a carousel interface, letting lawyers browse and respond to uploaded material without losing their place. Work began on integrating Retrieval-Augmented Generation with Amazon Bedrock, continuing the architecture direction set in August.

 

One operational reality surfaced clearly this month, worth stating plainly. Nigerian courts are not digitised. Even as Àdàbà AI streamlines evidence collection and case preparation, a lawyer still has to physically file printed documents with the court. The platform’s scope is deliberately bounded around what happens before that point: case management, evidence, and testimony, not the court filing process itself, which remains outside any single tool’s control.

 

What we haven’t solved yet

 

In the spirit of the honesty this project has tried to hold itself to: the legal data scarcity problem above is unsolved, and solving it requires funding the project doesn’t currently have. The DPIA, while substantially drafted, still needs to be translated into a finalised, public-facing privacy policy. And the AI model selection process, which specific approach will actually power triage and drafting, is still in early testing, not finalised.

 

None of these are hidden. They’re the actual state of the work at the end of September.

 

Why this matters beyond one organisation

 

Two things from this month are worth other organisations building similar tools taking note of. First, the “AI triage review” relabeling: a one-word interface change is a cheap, fast way to keep a system’s human-in-the-loop design honest in practice, not just on paper. Worth auditing your own interface for places where a label implies more authority than the system actually has. Second, the case-law scarcity problem is very unlikely to be unique to Nigeria or to TFGBV cases specifically. Any organisation building legal AI in a jurisdiction with limited digitised, reported case law will hit the same wall. The honest response, rather than quietly lowering the bar on what the AI claims to know, is to say so and design the system’s scope around what legal material actually exists.

 

Looking ahead

 

October’s focus is finishing AI model selection and sandbox testing, translating the DPIA into a finalised privacy policy, and continuing to build out the knowledge hub, the layer of explainers, glossaries, and legal resources that the AI’s triage tags will link out to. The team also plans further lawyer demos to stress test the dashboard before wider rollout.

 

Àdàbà AI is still a prototype. In September, it started meeting the people it’s meant to serve, and the limits of the legal record it has to work with.

 

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