DisElect-Africa

The paper makes a valuable contribution by extending election disinformation benchmarking to African electoral contexts through a controlled experimental design, addressing an important gap in AI safety evaluation. This methodology is a particular strength, and the Constitutional AI intervention adds practical value by demonstrating a mitigation strategy. The main opportunity for improvement is scope: the evaluation currently covers a limited set of countries, models, and prompts, which means the findings are best understood as an initial benchmark rather than a definitive assessment. Further validation of the judging pipeline and deeper analysis of trade-offs such as over-refusal would strengthen confidence in the conclusions and support broader application of the framework.

In relation to Criteria 1, it was not immediately clear to me how the findings could actually be used or deployed in a real world context to limit disinformation during elections. Generally, the project articulated quite clearly the problem, the test / results, and the limitations which were all clear to follow from a non-technical perspective. Something I did not quite understand was how Constitutional AI actually fixed the problem (I'd liked to have understood how Constitutional AI worked in a bit more detail).

The paper's most valuable contribution is the unexpected reframe, discovering that the locale gap hypothesis is superseded by a universal baseline failure and this is communicated clearly and is immediately actionable for low-resource African civic-tech practitioners. The most significant gap is the omission of the Constitutional AI system prompt text from the paper entirely, since the mitigation result is the second major finding and not reproducing it makes independent replication impossible

## Strengths

The risk this work identifies is real and timely, which makes the contribution highly relevant. Election disinformation is a live threat to African democratic processes, and the finding that open-weight models will generate it on demand at a near-total rate is exactly the kind of deployment-side concern the field needs to surface now rather than later.

The methodology is clever. Adapting an existing, published election-disinformation eval into a locale-controlled design is an efficient way to get to a rigorous result in a weekend, and holding the prompt language constant so that only the locale varies is the right move to isolate the effect being studied. The honest reframing of the result is also a strength. The team predicted a locale gap and instead found that compliance is near-ceiling everywhere, then drew the more important conclusion: the guardrails are essentially absent, not unevenly applied.

## Weaknesses

The main weakness is the ceiling effect. Because compliance sits so close to 100% across every arm, there is little headroom to disentangle a locale effect from the overall absence of guardrails, and the locale comparison is underpowered as a result. This is an honest finding rather than a flaw in the design, but it does limit what the current experiment can conclude about regional differences specifically. The scope is also narrow on the regional axis, covering two anglophone African countries, which constrains claims about African contexts more broadly.

## Recommendations for the authors

The clear next step is to **scale this work up**. First, **extend the evaluation to frontier models**. Unlike the open-weight models tested here, frontier systems have meaningful guardrails, so their compliance should fall away from the ceiling and finally create the headroom needed to detect whether locale matters. Second, **add more African country variations** so that genuine regional differences can be isolated. If both are done, this becomes a powerful and decision-relevant resource that could guide real democratic institutions in Africa as they defend against disinformation. On presentation, the paper is clear and easy to follow, and expanding some of the briefer bulleted sections into fuller prose would strengthen the structure and help it read as a complete paper.