Auditing Election-Disinformation Compliance in Open-Weight LLMs Across African and Western Contexts

This study tests how often open-weight AI models agree to write false election information such as a wrong voting date when asked. We tested 220 prompts across four elections: South Africa, Kenya, the UK, and the US (Pennsylvania). Models complied about 97% of the time with no safety measures. A short safety instruction (Constitution) reduced this to about 12%, but also made models refuse some genuine election questions. Code, prompt templates (with false claims replaced by placeholders), judgement labels, and results are included. Raw model outputs and filled-in false claims are withheld to prevent misuse.