Ufakazi

Ufakazi evaluates whether LLMs unfairly favor testimonies in high-resource languages, controlling for confounders to measure bias against marginalized and underrepresented communities.

Endorsed by

Temilola

Ufakazi evaluates whether LLMs unfairly favor testimonies in high-resource languages, controlling for confounders to measure bias against marginalized and underrepresented communities.

Endorsed by

Temilola

Project Details

Ufakazi is an evaluation harness built to determine whether models are unfairly biased towards trusting testimonies presented in high-resource languages (specifically English). By controlling for confounders, we isolate the language bias of various current generation LLMs and show that they often unfairly discriminate against marginalised and underrepresented communities.

Hackathon paper: Ufakazi
Code: github.com/KBrand26/Ufakazi

Theory of Impact

The South African constitution was written to rectify past discrimination on the basis of race and language. Hence testimony may still be given in any of the twelve official languages. With AI becoming more prevalent in the judicial system, a model that views testimony as more reliable when it is written in English quietly reintroduces bias into the courtroom and discriminates against historically marginalised communities. Ufakazi shows that most of the models we tested do this today.

Our path towards meaningful impact starts with providing explicit, published evidence of discriminatory behaviour in models. By naming the exact models, languages and the size and significance of the bias, we pre-empt the dismissal of the problem as anecdotal or speculative. We then ensure the visibility of our work in the rooms where it matters by partnering with Prof. Willem Fourie (founding director of Stellenbosch University's Policy Innovation Lab and an advisor to the African Union) and by writing an easily digestible blog post that can be shared with non-technical readers.

This project will be considered a success if the results sway decision-makers to introduce policy that requires multi-lingual testing of models before they can be introduced into judicial systems.

People

Kevin Brand

Team Member

Track Record

Global South AIS Hackathon (Apart Research), 2026

Grants Received– no grants recorded

Add grant

Funding Asks

Funding ask

Open

Minimum

$162

Ideal

$5,852

What this ask is

Ufakazi began as a hackathon project, but its findings have real-world consequences. It would be irresponsible to leave our work at a hackathon. We have since made contact with researchers and policy advisors in South Africa, who have offered to assist with disseminating the work through the appropriate channels. For that work to be taken seriously, it needs to be polished and published in a reputable journal. Any funding we receive goes towards the extensions that help us get there. We group these extensions by necessity.

The first extension is one we believe necessary for the work to be publishable. We need to increase the repetitions per experiment from five to ten and grow the scenario set from 20 to 40. The headline findings are already statistically significant and we expect them to hold, but we would like to tighten the confidence intervals to pre-empt fair questions from reviewers.

Three further extensions would be valuable, but not necessary. Adding frontier models (Claude Fable 5.1, GPT-6 Astra, Gemini 3.1 Pro and Qwen 3.8 Max) shows whether our findings hold for the most advanced models. Acquiring native translations for isiXhosa and isiZulu removes any remaining doubt that translation fidelity could be the true explanation. Finally, varying the language of the system prompt tests whether the bias shifts to align with the language of the instructions. Each of these extensions addresses a question that could reasonably be raised in review.

The final two extensions are nice-to-have. One investigates whether commonly used legal AI products (such as Harvey and Legora) are similarly susceptible to language bias. Their domain-specific harnesses could change language sensitivity, but they are not directly comparable and are harder to access. Adding very low-resource African languages such as isiNdebele and Tshivenda would also show how language prevalence shapes model bias.

How the money will be spent

All budget line items are assuming that we extend the scenario set to 40 scenarios.

Contingency buffers account for any testing, bugs or unforeseen expenses.

Necessary extensions:

Added token usage for statistical hardening (extending scenario set and increasing repetitions): $134.83
20% contingency buffer: $26.97

Valuable extensions:

Added token usage for adding frontier models: $378.96
Added token usage for native translations: $375.26
Paying for native translators: ~$180
Added token usage for varying system prompt language: $2716.92
20% contingency buffer: $721.19

Nice-to-have extensions:

Token usage for two more low-resource languages: $757.77
Paying for native translators for the two additional languages: ~$180.00
20% contingency buffer: $151.554

Discussion

Private comments in this thread are only shown to approved funders and grant reviewers. Email hi@grantmaking.ai or to request access.

Private UserPrivate15d?

Temilola OloyedeJudge · Global South AIS HackathonEndorsement15d

This was a well-executed project with a thoughtful methodology, especially given the time constraints. I appreciated the use of clear controls and validity checks which made the results easier to interpret. I also appreciated the thoroughness in ruling out other confounding factors.

The methodology felt especially strong in how it separated different possible sources of bias.
Overall, this was a rigorous and well-scoped.