AI's Environmental Footprint

Overview

AI systems require substantial computational resources, which translates to significant energy consumption. For clinicians evaluating AI tools, understanding these environmental costs is part of responsible technology adoption—particularly as healthcare organizations increasingly prioritize sustainability.

This appendix provides context on AI's environmental footprint without advocating for or against AI adoption. The goal is informed decision-making.

The Scale of AI Energy Use

Training vs. Inference

AI energy costs come in two forms:

• Training: The one-time (or periodic) process of building the model. Training large language models like GPT-4 requires massive computational resources—estimates suggest training runs can consume energy equivalent to hundreds of homes' annual usage.
• Inference: The ongoing cost of running the model to generate outputs. Each query to ChatGPT, each image analyzed by a diagnostic AI, consumes energy. At scale, inference often exceeds training costs.

Data Center Impact

AI workloads are driving rapid expansion of data centers, which require:

• Electricity: For computation and cooling systems
• Water: Many data centers use evaporative cooling, consuming millions of gallons annually
• Hardware: GPUs and specialized AI chips have their own manufacturing footprint

Major tech companies have seen their carbon emissions rise despite renewable energy commitments, largely due to AI infrastructure expansion.

Healthcare-Specific Considerations

Imaging AI

Radiology and pathology AI systems process large image files, which is computationally intensive. A hospital running AI analysis on every chest X-ray or CT scan generates continuous inference costs. The benefit (faster diagnosis, reduced workload) must be weighed against resource use.

Ambient Documentation

AI scribes that transcribe and summarize clinical encounters require continuous speech-to-text processing plus LLM summarization. For a health system with thousands of daily encounters, this represents substantial ongoing computation.

Clinical Decision Support

Real-time CDS tools that query AI models for every patient interaction (alerts, suggestions, risk scores) generate high inference volumes. The clinical value per query varies—some alerts are ignored, others change care.

Frameworks for Thinking About This

Net Benefit Analysis

Consider AI's environmental cost in context:

• Does the AI reduce other resource use? (e.g., fewer unnecessary tests, shorter hospital stays, reduced travel for telehealth)
• What's the alternative's footprint? (e.g., flying pathology slides across the country vs. digital analysis)
• Does improved diagnosis prevent downstream resource-intensive care?

Proportionality

Not all AI use cases are equal:

• High value: AI that catches cancers, prevents sepsis deaths, or enables care access may justify significant resource use
• Lower value: AI for convenience features with marginal clinical benefit deserves more scrutiny

Institutional Responsibility

Individual clinicians have limited control over infrastructure choices. But institutions can:

• Select vendors with strong sustainability practices
• Choose cloud providers powered by renewable energy
• Right-size AI deployments (not everything needs the largest model)
• Consider on-premise vs. cloud trade-offs

What Clinicians Can Do

Ask Questions

When evaluating AI tools, consider asking vendors:

• Where does computation happen? What's the data center's energy source?
• Is this the right-sized model for the task, or is it using a general-purpose LLM when a smaller specialized model would suffice?
• What's the per-query resource footprint?

Use Thoughtfully

This isn't about avoiding AI—it's about intentional use:

• Use AI when it adds clinical value, not just because it's available
• Consider whether the task requires the most powerful model or if simpler tools suffice
• Batch queries when possible rather than multiple small requests

Advocate Institutionally

Push for sustainability to be part of AI procurement criteria. Health systems increasingly have sustainability officers and carbon reduction commitments—AI infrastructure should be part of those conversations.

Perspective

Healthcare itself has a substantial environmental footprint—estimated at 8-10% of U.S. greenhouse gas emissions. AI is one factor among many, including building operations, supply chains, and pharmaceutical manufacturing.

The question isn't whether AI has environmental costs (it does) but whether its benefits justify those costs and whether we're deploying it thoughtfully. The same critical thinking you apply to clinical AI effectiveness should extend to its broader impacts.

Further Reading