Research tool probing LLM safety filter consistency across English, Yoruba, Igbo and Nigerian Pidgin
# African Language AI Safety Probe
> *Do LLM safety filters apply equally across English and African languages or does linguistic disparity create security and fairness gaps?*
**Researcher:** Chuka Okafor |
github.com
**Field:** AI Safety · Multilingual NLP · Cybersecurity
**Status:** Active Research
---
## The Research Question
Large language models are increasingly deployed across African markets in fintech, healthcare, education, and government services. But most safety research is conducted in English.
This project systematically probes whether AI safety filters, warning mechanisms, and response quality are consistent across:
- 🇬🇧 **English** (baseline)
- 🇳🇬 **Yoruba**
- 🇳🇬 **Igbo**
- 🇳🇬 **Nigerian Pidgin English**
If a safety filter blocks harmful content in English but not in Igbo that's not just a technical bug. That's an equity problem affecting millions of people.
---
## Probe Categories
| Category | Description |
|----------|-------------|
| **Baseline Comprehension** | Control group confirms model understands each language |
| **Safety Consistency — Fraud** | Does fraud awareness advice appear equally in all languages? |
| **Cultural Context** | Does the model have knowledge of Nigerian specific digital threats? |
| **Bias Detection** | Does AI equity framing differ based on the language of inquiry? |
| **Safety Consistency — Scam** | Does the model warn about advance fee scams equally across languages? |
---
## Setup & Usage
### 1. Clone the repo
```bash
git clone
github.com
cd african-language-safety-probe
```
### 2. Install dependencies
```bash
pip install -r requirements.txt
```
### 3. Get a free Anthropic API key
Sign up at console.anthropic.com free tier available.
### 4. Run the probe
```bash
python main.py
```
Enter your API key when prompted, or set it as an environment variable:
```bash
# Windows
set ANTHROPIC_API_KEY=your_key_here
# Mac/Linux
export ANTHRO …