An AI fix for an AI problem? - Nature

Artificial intelligence (AI) agents can be used to calculate the carbon footprint of electronic devices, potentially helping to address the increasing environmental impact of electronics and AI.

More than 60 million tonnes of electronic waste is produced every year, and less than a quarter of this is reported as being collected and recycled1. Electronics also now accounts for a notable proportion of global greenhouse gas emissions, with estimates suggesting the information and communication technology sector could be responsible for up to around 4% of global emissions2. This environmental impact is currently only getting worse3, and thus the need to develop more sustainable electronics — and to carefully consider the way we design, build, use and discard electronic devices — is clear.

Accurately quantifying the carbon emissions associated with a particular electronic device is an important step in this direction. Such information is typically estimated using a life-cycle assessment, a methodology that attempts to trace the environmental impact of a product across the supply chain. An analysis of an electronic device — and its numerous materials and components — can though take weeks or months, as data have to be manually gathered (often by life-cycle assessment professionals) from different proprietary supplier documentation and disparate organizational sources.

In an Article in this issue of Nature Electronics, Zhihan Zhang, Adriana Schulz, Vikram Iyer and colleagues show how artificial intelligence (AI) agents can be used to calculate the carbon footprint of electronic devices. The researchers — who are based at the University of Washington, the University of Notre Dame, and Northeastern University in Boston — develop a multi-agent AI system that relies on both large language models and vision–language models, and can emulate the life-cycle assessment process of human experts. With the help of structured data abstraction and software tools that mine information from the Internet, the AI agents iteratively build a complete life-cycle inventory, reducing the time it takes to collect such data from weeks or months of expert time to under a minute.

The approach provides carbon footprint estimates — on product categories such as smartphones and graphics processing units — that are within 19% of expert life-cycle assessments. This level of variation is, in fact, similar to that seen between human expert assessments, where methodological differences can lead to comparable error margins.

As Zhang and Iyer from the research team explain in the accompanying Research Briefing, “Our approach can guide low-emission design choices for components and materials early in the product development cycle and enables rapid environmental labelling of existing products.” They also note that, “Beyond electronics, our approach can be adapted to other manufactured goods with complex and fragmented supply chains.”

The increasing environmental impact of electronics is a problem that has been exacerbated by the recent advance of AI, and large language models, in particular. Thus, suggestions of an AI fix for an AI problem do merit some scepticism. Here, at least, attempts are made to minimize energy consumption, with the use of existing pre-trained models and custom lightweight algorithms. But Zhang and Iyer still caution that “the environmental cost should be considered if such AI tools are deployed at scale.”

References

- Baldé, C. P. et al. Global E-waste Monitor 2024 (International Telecommunication Union and United Nations Institute for Training and Research, 2024).

- Bieser, J. C. T., Hintemann, R., Hilty, L. M. & Beucker, S. Environ. Impact Assess. Rev. 99, 107033 (2023).

- Freitag, C. et al. Patterns 2, 100340 (2021).

About this article

Cite this article

An AI fix for an AI problem?. Nat Electron 9, 713 (2026). https://doi.org/10.1038/s41928-026-01679-0

- Published:

- Version of record:

- Issue date:

- DOI: https://doi.org/10.1038/s41928-026-01679-0