GAO-25-107172, "Artificial Intelligence: Generative AI's Environmental and Human Effects," is a U.S. Government Accountability Office report published April 22, 2025 and authored by Brian Bothwell and Kevin Walsh. It examines the environmental effects of generative AI (energy and water use) and a set of human effects (safety, privacy, cybersecurity, bias, and accountability), and presents policy options for each. Its central finding is that the environmental effects are not well understood because companies generally do not report the relevant data.
Authors: Brian Bothwell, Kevin Walsh Publisher: U.S. Government Accountability Office Document ID: GAO-25-107172 Published: April 22, 2025
Environmental findings
GAO finds that the IT equipment powering generative AI consumes significant energy and water resources, but that companies are generally not reporting details of these uses. Training of large language models consumes electricity "often equivalent to the energy use of entire towns" and relies heavily on water-intensive cooling systems inside data centers. The report's main finding on environmental effects is that "environmental effects are uncertain and not well understood due to insufficient data and information," which GAO frames as a core data gap.
Human-effects findings
GAO identifies five human-effects risks of generative AI:
- Unsafe systems — potential malfunctions or dangerous outputs.
- Lack of data privacy — inadequate protection of user information.
- Cybersecurity concerns — exposure to attacks and breaches.
- Unintentional bias — discriminatory outcomes in AI decisions.
- Lack of accountability — insufficient mechanisms for responsibility.
The report states that these risks span four application areas: public services, labor markets, education, and R&D.
Policy options
For environmental effects, GAO presents three options: maintain the status quo of current efforts in academia, industry, and government; expand efforts to improve data collection and reporting; and encourage innovation. The data-collection-and-reporting option aligns with Sen. Markey's transparency-mandate advocacy.
For human effects, GAO presents three options: maintain the status quo of current policy efforts; encourage the use of available AI frameworks (implicitly including NIST AI Risk Management Framework (AI RMF 1.0)); and continue to expand efforts to share best practices and establish standards.
Reception and context
The report is described as the first GAO product framing generative AI's effects as a government-accountability question rather than solely a technology-capability question. It provides a formal GAO basis for congressional appropriators and committees to scrutinize AI energy use and to demand company disclosure, backing Sen. Markey's position on AI energy-disclosure legislation. By naming the "companies aren't reporting" problem explicitly, it positions state-level disclosure laws such as AB 2013 — Training Data Documentation (California) as a corrective.
Provenance
WebFetch returned 403 errors on both the GAO HTML and PDF endpoints; the summary was reconstructed from a third-party summary (Akin Gump law firm tracker) and corroborating press coverage (FedScoop, Sen. Markey statement). The PDF is available at https://www.gao.gov/assets/gao-25-107172.pdf and can be pulled directly via browser if verbatim quotes are needed.
Relationships
- supports: AI Environmental Impact, NERC 2025 Long-Term Reliability Assessment (LTRA), RAND — AI's Power Requirements Under Exponential Growth (2025), Bain & Company — 6th Annual Global Technology Report (2025), AB 2013 — Training Data Documentation (California).
- related: Epoch AI — How Much Power Will Frontier AI Training Demand in 2030?, Epoch AI — Can AI Scaling Continue Through 2030?, Data Center Siting / AI Power Politics.
- depends-on: AI Compliance Industry / Regulatory Fragmentation.
Sources
- Primary: GAO-25-107172 Generative AI Environmental and Human Effects (https://www.gao.gov/products/gao-25-107172)