Hausa Preprint Reveals AI Safety Alignment Breaks Down for Low-Resource Languages
An arXiv preprint from October 27 2026 shows that standard safety methods applied to Qwen2.5-7B produce high refusal rates on Hausa prompts mainly because the base model…
- The untuned base model appears safe on Hausa harmful prompts only because its limited generation ability leads to low-quality refusals across the board.
- Supervised fine-tuning on the new HausaSafeLM dataset, especially the larger silver machine-translated portion, improves both refusal fluency and precision.
- Adding Direct Preference Optimization after fine-tuning sharpens refusal precision but causes a steep drop in linguistic quality regardless of gold or silver data.
What the HausaSafeLM preprint found
This arXiv preprint, based on its abstract only, examines how existing safety alignment techniques perform for Hausa, a language spoken by more than 200 million people in West Africa. Researchers built HausaSafeLM, one of the first Hausa safety preference datasets, by adapting 10000 prompts from the English PKU-SafeRLHF collection across six harm categories. It includes 3000 human-corrected gold examples and 7000 machine-translated silver examples.
Using Qwen2.5-7B-Instruct, they compared supervised fine-tuning alone against fine-tuning followed by Direct Preference Optimization. Evaluation occurred on 500 harmful and 200 benign Hausa prompts. The abstract notes that the untuned base model shows high refusal rates on harmful prompts, yet these refusals are almost uniformly low quality because the model has limited ability to generate fluent Hausa.
How alignment methods performed on Hausa
According to the abstract, supervised fine-tuning improves both the fluency of refusals and their precision. It benefits substantially from the larger silver dataset despite translation noise. In contrast, adding Direct Preference Optimization after fine-tuning sharpens refusal precision but incurs a steep, largely data-invariant cost to linguistic quality.
No single configuration jointly maximizes safety, helpfulness, and fluency. The researchers conclude that current alignment techniques primarily teach models when to refuse rather than how to refuse fluently in underrepresented languages, and that refusal-rate metrics alone are unreliable signals of safety in low-resource settings.
Why it matters for uncensored AI
Our analysis starts from the preprint abstract's finding that refusal-rate metrics alone are an unreliable signal of safety for Hausa. This highlights how alignment methods developed mainly for high-resource languages can produce models that seem safe on paper but fail to deliver usable, fluent responses for millions of speakers in West Africa.
The practical consequence is that users and developers working with Hausa or similar languages face a trade-off: accept models with awkward refusals or seek alternatives that better balance helpfulness, cultural fit, and linguistic quality. This limitation in standard alignment approaches underscores the value of open models that communities can adapt locally for their specific expressive needs without rigid corporate safety layers.
