Àdàbà AI: Building a Survivor-Centred Legal AI Tool for Gender Justice in Nigeria

When a woman in Nigeria is targeted by a deepfake, or has intimate images shared without her consent, the slowest part of her ordeal often comes after the abuse. She has to save screenshots before they disappear, piece together who posted what and when, and turn a painful personal account into evidence a court will accept. Then she waits, sometimes for weeks, for a lawyer to prepare her case.

Digital harm spreads at the speed of a repost. Legal response moves far more slowly.

Àdàbà AI was created to close that gap. It is being developed by DigiCivic Initiative, a Nigerian digital rights organisation, as one of the projects in the A+ Alliance Gender & AI Innovation Collective (GAIC), a cohort of feminist AI projects supported by Women At The Table and Code for Africa. 

What is Àdàbà AI?

Àdàbà AI is a digital tool that helps survivors of technology-facilitated gender-based violence (TFGBV) document incidents, organise evidence and move more quickly towards legal support. TFGBV covers harm carried out through digital tools such as social media, messaging apps, fake accounts and AI-generated images, used to threaten, harass or expose someone, most often a woman or girl.

The tool takes on the repetitive groundwork that sits between a survivor’s report and a lawyer’s desk. It organises evidence, suggests what kind of legal case the evidence might support, and prepares a first draft of the paperwork. A qualified lawyer reviews every output and makes every final decision.

Àdàbà AI is designed to work alongside existing reporting platforms in Nigeria’s digital rights ecosystem, including TechHer’s Kuram and Paradigm Initiative’s Ripoti. Those platforms give survivors safe places to report abuse. Àdàbà AI picks up at the next stage, turning a reported case into a case file a lawyer can use in court.

Discovery, design and a key scoring decision

The first two months focused on understanding the problem with the people who will use the tool: survivors, lawyers and women’s rights organisations. The team set a five-month plan (understand, design, build, test, pilot) and mapped its first user journey with legal, technical and programme staff. Consultations followed in stages, with women’s rights organisation partners and lawyers. The lawyers’ session produced the most important design decision of the early phase.

Àdàbà AI keeps two scores separate:

• An Evidence Completeness Score, measuring how much of what a lawyer needs is present: the survivor’s account, the people involved, the digital evidence, whether it has been preserved intact, its legal relevance, and supporting material such as witness statements or police reports.
  
  
• A Survivor Risk Assessment, measuring how urgently the survivor needs help, based on physical danger, severity, escalation, vulnerability, psychological impact and how widely the abuse has spread.

A case with very little documentation can still be an emergency and a well-documented case can be low risk. Keeping the two scores apart means a lawyer opening a case sees both results and a recommended next step and urgency stays visible.

Two architectural principles were also set in this period. A human always makes the final call, and every AI output is labelled as a draft needing lawyer review the moment it is produced. The system is also built with swappable components, so an organisation can run the same core case logic on a paid AI service or on a free, self-hosted alternative. That flexibility makes the tool realistic for smaller women’s rights organisations with tight budgets.

Legal foundations, data protection and RAG

In august the team moved with the project from high-level design into the foundations of a working prototype.

Survivor-centred data protection: The team drafted a data protection framework and plain-language terms and conditions grounded in feminist and community-centred approaches to privacy, consent and safety. These documents are treated as part of the product, reviewed with the same care as a wireframe, so survivors can clearly understand what they are agreeing to.

A structured legal knowledge base: the team began organising Nigerian and ECOWAS laws, court decisions and legal guidance into a corpus tagged by jurisdiction, document type, legal category, source, date and applicable law. Court records that appear in the corpus are redacted where needed to protect survivors whose details appear in them.

Retrieval-Augmented Generation: DigiCivic chose a Retrieval-Augmented Generation (RAG) approach, combined with rule-based controls, to ground the AI’s outputs in retrieved legal sources. RAG lets a small team maintain and update a curated knowledge base without the cost of retraining a large model, which keeps the system sustainable for a civil society organisation.

Security designed around the survivor: Planned safeguards include role-based and case-level access controls, a secure evidence vault, file hashing, timestamps, encryption in transit and at rest, and protections against prompt injection. Evidence items carry statuses such as Accepted, Under Review and Needs Context, creating a feedback loop between survivors and lawyers. The team also noted an important distinction: a file hash shows a file has not been altered, and proving when an event actually happened requires further evidence.

Two interfaces for two users: The survivor interface is calm and gentle, built around the Àdàbà dove symbol. Survivors can save progress and return later, skip evidence upload until they are ready, type or speak their testimony, and reach emergency support before entering personal details. The lawyer interface is more professional and focused on case review.

In August, the DigiCivic team attented the GAIC bootcamp, which covered data audits, data cleaning, chunking, embeddings, RAG, prompting, guardrails and model selection, with hands-on work in the AI Innovation sandbox. The team took away a clear lesson: an AI system’s quality depends heavily on how carefully its data and instructions are prepared.

By the end of the month, backend development had started on PostgreSQL, frontend work was underway, and the team had set a pilot target of 31 October 2026.

Survivors at the table

On 11 September, DigiCivic held its first direct consultation with survivors, walking them through the design. Survivors responded warmly to the privacy controls and the “need-to-know” confidentiality model, welcomed voice-to-text testimony, and showed interest in the network of lawyers available across nearly all Nigerian states. Several asked to test the platform once it is built. The team now plans to involve survivors as core testers throughout development.

Honest labels in the interface. During design review, the lawyer dashboard’s “triage” label was renamed “AI triage review” to keep the lawyer’s role visible. An ambiguous “evidence review” checkmark became “case reviewed”, marking that a lawyer has confirmed they understand the case. These small changes make the interface communicate exactly where human responsibility sits.

A data protection impact assessment. The team drafted a Data Protection Impact Assessment (DPIA) covering ten risk areas, including AI hallucination and unauthorised evidence disclosure. Evidence is annotated on working copies so originals stay intact. DigiCivic plans to publish the DPIA and privacy policy openly once finalised, and found Nigeria’s Data Protection Act a strong legal foundation for this work.

What was built. Backend authentication was substantially completed. The survivor dashboard gained an AI-generated event timeline that survivors and lawyers can correct by hand. The lawyer dashboard now includes case overviews, a workload snapshot, a review queue that flags overdue tasks, and a carousel for reviewing evidence. Work began on integrating RAG with Amazon Bedrock.

The challenge ahead: Nigeria’s case-law gap

September also surfaced the project’s hardest problem. Very little Nigerian case law on online gender-based violence is publicly accessible. The team found only one readily available High Court judgment on the subject. Most High Court decisions go unreported, and many cases never reach judgment. Appellate decisions are reported and binding, yet TFGBV cases rarely reach those courts. Judgments from the ECOWAS Court carry persuasive weight in Nigeria.

Closing this gap would require funded research assistance to retrieve judgments from court archives. For now, Àdàbà AI is designed to help lawyers find relevant Nigerian, regional and international legal instruments and to flag where Nigerian law is silent, so lawyers can build well-grounded cross-jurisdictional arguments. Courts in Nigeria also remain largely paper-based, so the platform’s scope covers case preparation, evidence and testimony up to the point of filing.

What other teams can learn from Àdàbà AI

DigiCivic’s work offers several practical lessons for organisations building AI for gender justice in under-resourced settings:

1. Keep evidence readiness and survivor urgency as two separate measures.
2. Build systems with swappable components so smaller organisations can afford to run them.
3. Write consent and data-use terms as part of the user experience.
4. Use RAG to keep a legal AI tool maintainable on a modest budget.
5. Audit interface labels so they reflect where human judgment sits.
6. Design the system’s scope around the legal material that actually exists, and say so openly.

October’s priorities are finalising AI model selection and sandbox testing, turning the DPIA into a public privacy policy, building the knowledge hub of explainers, glossaries and legal resources, and running further demos with lawyers ahead of the pilot. The A+ Alliance looks forward to sharing what the pilot teaches us.

Àdàbà AI reflects what the Gender & AI Innovation Collective was set up to support: African feminist organisations building AI on their own terms, grounded in the needs and safety of the communities they serve.