Generative AI uses more energy on non-English text (news.umich.edu)

Recent research from the University of Michigan reveals that generative AI models consume significantly more energy when processing non-English text, often producing lower-quality outputs. The study found that energy usage can be up to 180 times higher for languages like Hindi and Bengali compared to English, primarily due to the increased number of tokens required for these scripts. This disparity could reinforce global inequalities, as around 5.2 billion individuals, who might prefer using AI in their native languages, face higher costs and inferior performance. The implications for the AI/ML community are critical. The findings suggest that current AI models are primarily optimized for English, creating barriers for broader global usage. To mitigate these issues, researchers recommend developing more efficient language-specific AI models and integrating translation pipelines to lower energy consumption and improve accuracy. This proactive approach could enhance accessibility for non-English speakers and ensure a more equitable distribution of AI resources across diverse linguistic communities.