AI models penalize Swiss French speakers, experiment finds

Artificial intelligence models are showing a measurable bias against Swiss French speakers, revealing an ignorance of regional linguistic variations that could lead to discrimination. An experiment conducted on four leading large language models (LLMs) demonstrated that these AI systems struggle to understand and process "Romand," the French spoken in Switzerland, compared to standard French.

The study, which partnered with the newsletter Gargant'IA, tested models including Mistral from French startup Mistral AI, Google's Gemma3, Meta's Llama3.1, and Alibaba's Qwen3. These AI chatbots, designed to comprehend and generate text, were subjected to fictional scenarios presented in both standard French (fr_FR) and Swiss French (fr_CH). The results indicated significant differences in how the models responded, with Swiss French variations often leading to less accurate or appropriate outputs.

Researchers noted that the models' performance varied, but the overall trend showed a clear disadvantage for Swiss French. This linguistic blind spot could have practical implications, potentially affecting everything from translation services to customer support bots. The experiment's methodology was inspired by previous research on AI's treatment of American English dialects, aiming for statistically robust findings by running multiple queries for each scenario.

Despite the clear evidence of bias, the article points out a concerning lack of urgency in addressing the issue within Switzerland. The findings suggest that AI systems, while powerful tools, are not yet equipped to handle the nuances of regional languages, raising questions about fairness and inclusivity in AI development and deployment. The researchers emphasize the need for regulation to ensure that AI technologies do not perpetuate or exacerbate linguistic discrimination.