Environmental harms from operating LMs
LMs (and AI more broadly) can have an environmental impact at different levels, including: (1) direct impacts from the energy used to train or operate the LM, (2) secondary impacts due to emissions from LM-based applications, (3) system-level impacts as LM-based applications influence human behaviour (e.g. increasing environmental awareness or consumption), and (4) resource impacts on precious metals and other materials required to build hardware on which the computations are run e.g. data centres, chips, or devices. Some evidence exists on (1), but (2) and (3) will likely be more significant for overall CO2 emissions, and harder to measure [96]. (4) may become more significant if LM-based applications lead to more computations being run on mobile devices, increasing overall demand, and is modulated by life-cycles of hardware.
ENTITY
2 - AI
INTENT
2 - Unintentional
TIMING
3 - Other
Risk ID
mit227
Domain lineage
6. Socioeconomic and Environmental
6.6 > Environmental harm
Mitigation strategy
- Prioritize the implementation of energy-efficient computing and cooling solutions for data centers, such as optimizing Large Language Model (LM) inference efficiency and utilizing advanced low-water or waterless cooling systems (e.g., immersion cooling) to directly reduce electricity consumption and associated carbon emissions. - Mandate the procurement of 100% renewable energy for all LM training and operational (inference) workloads to eliminate the carbon emissions generated from reliance on fossil fuels. - Establish policies to extend the operational life of AI-specific hardware (servers and chips) through refurbishment, reuse, and third-party maintenance to minimize the embodied carbon and electronic waste associated with frequent device replacement.
ADDITIONAL EVIDENCE
LMs and other large machine learning models create significant energy demands during training and op- eration [15, 148, 176], and correspondingly high carbon emissions when energy is procured from fossil fuels [141]. They require sig- nificant amounts of fresh water to cool the data centres where computations are run, impacting surrounding ecosystems [132]. Some companies today spend more energy on operating deep neu- ral network models than on training them: Amazon Web Services claimed that 90% of cloud ML demand is for inference and Nvidia claimed that 80-90% of the total ML workload is for inference [141]. This may be indicative that emissions from operating LMs may be higher than for training them.