Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJun 21, 2026Johannesburg

The African-Language Safety Gap Is Model-Dependent: A Comprehension-Controlled Audit of Vision–Language Model Refusal in isiZulu and Hausa

Rayhaan Hanslod , Altaaf Ally · Team Nexus

Submitted to Global South AI Safety Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: The African-Language Safety Gap Is Model-Dependent: A Comprehension-Controlled Audit of Vision–Language Model Refusal in isiZulu and Hausa

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

We investigated whether AI models apply the same safety rules in African languages as they do in English. To separate language understanding from safety failures, whenever a model answered a harmful request instead of refusing it, we also asked it to explain what the request meant. We tested two leading AI models in English, isiZulu, and Hausa. One model, Claude, consistently refused all harmful requests across all languages. The other, Kimi, refused harmful requests in English but sometimes complied in isiZulu, including generating election disinformation that it had refused to produce in English. Since the model clearly understood the request, the issue was not language comprehension but a failure of its safety mechanisms. These results show that safety protections can vary across languages and highlight the importance of testing AI systems in African languages before deployment.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. This one stood out to me. The comprehension check is the smart move: by separately confirming the model actually understood the request, you can tell a real safety failure apart from the model just not getting the language. That is the thing that trips up most multilingual red-teaming, and you handled it cleanly. I also liked that one model held up and the other did not, since it shows the gap is not inevitable. And you were honest in a way I do not see often, calling it a small-sample existence proof, leading with effect sizes, and even catching a false positive in your own setup and killing it. The clear limit is sample size, the whole thing rests on a handful of non-refusals, so scaling up and finishing the human-judge agreement you have pending is the next step. One of the cleanest writeups in the batch.

  2. Document extends existing research. Refusal testing for isiZulu already in 2605.18239. New approach using VLMs is interesting. Methodology is strong and supports the results. Document was very clear and fun to read.

  3. This project tackles a valid concern regarding the capabilities of VLMs to identify and refuse malicious instructions in different languages, particularly West African languages, in comparison with English. The authors highlight the importance of identifying the difference between failure to comply due to misunderstanding and refusal given safety concerns and demonstrate that for the Kimi-K2.6 model this is effectively a concern. However, the paper is not clear on how the content of the images are supposed to impact the result of the experiment. The images appear to be used to add some implicit context, with the prompt not using the image and serving more as noise for the VLM. Future work could benefit from a clearer link between the prompt and the image or by stating the intended role of the image and the possible complexities or malicious intentions that it might provoke to the model.

Cite this project

@misc{hanslod2026africanlanguage,
  title = {{The African-Language Safety Gap Is Model-Dependent: A Comprehension-Controlled Audit of Vision–Language Model Refusal in isiZulu and Hausa}},
  author = {Rayhaan Hanslod and Altaaf Ally},
  year = {2026},
  month = jun,
  note = {Submitted to Global South AI Safety Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/the-africanlanguage-safety-gap-is-modeldependent-a-comprehensioncontrolled-audit-of-visionlanguage-model-refusal-in-isizulu-and-hausa-ib1j}},
  url = {https://apartresearch.com/sprints/projects/the-africanlanguage-safety-gap-is-modeldependent-a-comprehensioncontrolled-audit-of-visionlanguage-model-refusal-in-isizulu-and-hausa-ib1j}
}

More from Global South AI Safety Hackathon

All 216 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026