Generative AI Has A Hidden Energy Problem That Could Penalize Billions Of Non-English Users

The Hidden Energy Tax of Global AI Accessibility

The global rollout of generative artificial intelligence is masking a quiet, resource-intensive inequity. A new study reveals that AI models consume significantly more energy when processing non-English languages compared to English, creating a “double penalty” that threatens to alienate billions of users and widen the digital divide.

Research presented at the Association for Computational Linguistics’ conference on Empirical Methods in Natural Language Processing highlights that for roughly 5.2 billion people, interacting with AI in their native language demands more than double the energy required by English speakers. This discrepancy is not merely an environmental concern but a potential economic barrier that could discourage companies from supporting diverse linguistic outputs.

“The languages that typically consume more energy also tend to have lower accuracy in the model output,” explains Naihao Deng, a doctoral student at the University of Michigan AI Lab and lead author of the study. “Most people will want to use AI models by asking questions in their native languages, and they may feel pressured to use it as it’s incorporated into more parts of society, but the double penalty makes it more likely that certain people will be priced out of AI.”

Why Non-Latin Scripts Demand More Processing Power

To quantify these disparities, researchers analyzed three prominent open-source models—Alibaba’s Qwen 3, Google’s Gemma 3, and Meta’s Llama 3.1—using the ML.Energy framework developed at the University of Michigan. By evaluating reading comprehension and math reasoning across various languages, the team found that while Western European languages generally maintained energy costs within a twofold increase, other languages faced extreme spikes. Specifically, Hindi and Bengali required up to 7.9 and 15.3 times more energy, respectively, than English prompts.

The root of this efficiency gap lies in how AI models handle “tokens,” the fundamental units of data used to process information. Because many languages—particularly those using non-Latin scripts like Devanagari or Bengali—require more complex encoding, they necessitate a higher volume of tokens to represent the same amount of information. This increased tokenization places a heavier burden on computer memory, forcing models to process information in smaller, less efficient batches, which compounds energy consumption.

A Path Toward Equitable AI

The energy inefficiency often translates into direct financial costs for users, as service providers typically charge based on token volume. To prevent these technical limitations from cementing existing inequalities, the study authors suggest several strategic shifts for AI developers:

• Language-Specific Routing: Developers could route queries to specialized models optimized for specific linguistic structures rather than forcing a generalized model to handle high-token-count inputs.
• Strategic Translation: Where reliable, systems could translate non-English prompts into English before processing, potentially reducing the computational load.
• Resource Expansion: There is a critical need to increase training data for underrepresented languages. The researchers suggest crowdsourcing native speakers to help build datasets, which would improve both the efficiency and the accuracy of models serving these populations.

Rada Mihalcea, director of the U-M AI Lab and corresponding author of the study, emphasizes that the burden of adjustment lies with the developers. “AI developers should be creating more resources for those languages with an energy and cost disparity,” she notes.

The research, which utilized the University of Michigan’s Great Lakes high-performance computing cluster, was supported by funding from the National Science Foundation, DARPA, Ford, the Laude Institute, OpenAI, and the Survival and Flourishing Fund. As AI becomes deeply embedded in global infrastructure, these findings underscore that true accessibility requires more than just language support—it demands a fundamental rethink of the underlying computational economics.