AI systems influence real people and real communities, so technical performance alone is not enough. This section introduces the ethical risks of bias and opacity, along with the growing environmental costs of large-scale AI infrastructure.
Subsection3.6.1The Black Box Problem and Algorithmic Bias
As AI transitions from a theoretical computer science discipline into part of society’s infrastructure, we must look beyond optimization and confront a critical reality: AI is a reflection of human choices.
One of the most immediate technical and ethical challenges in modern AI is the Black Box Problem. While traditional programs follow clear, human-readable logic blocks, a deep learning network might alter millions of internal numerical parameters to reach a decision.
If an autonomous system or automated screening tool denies a bank loan or rejects a job applicant, it can be mathematically difficult for engineers to explain exactly why the model reached that conclusion. This lack of transparency becomes dangerous when systems amplify systemic prejudices, a phenomenon known as algorithmic bias. Because machine learning models derive internal rules from historical human data, they can inherit, automate, and scale societal bias.
Consider a large corporation using a supervised model to screen resumes. Engineers train the model on historical submissions from the last 10 years, with labels indicating who was hired and promoted.
If engineering and leadership roles were historically male-dominated due to systemic factors, the dataset may contain more male profiles marked with positive outcomes. The model, lacking social context, searches for correlations only. It may overvalue phrasing patterns or activities that correlate with past hiring labels and underweight equally strong candidates whose resumes use different language or affiliations.
In 2014, Amazon deployed such a system, that acted as an automated mirror of past hiring patterns rather than a fair evaluator of potential. As a result of this kind of bias risk, Amazon discontinued use of the internal AI hiring tool in early 2017. All code written by humans can encode human bias.
Subsection3.6.2The Environmental Cost of AI: Energy and Water Footprints
While social impacts of bias are widely discussed, the physical infrastructure supporting AI introduces an escalating environmental footprint. Training and running modern generative models requires large-scale computation that consumes significant resources.
According to the International Energy Agency (IEA), global data center electricity use was approximately 460 TWh in 2022. Driven by AI inference and large-model training, projected use ranges from 650 TWh to 1,050 TWh, approximately doubling the baseline.
A traditional web search is often around 0.3 Wh. A single generative AI query can be around 0.3 to 0.34 kWh, roughly a 1000x increase in energy per query.
In the United States, which represents a major share of the global data center market, data centers consumed about 4.4% of total electricity in 2023 and are projected to rise as AI infrastructure scales.
AI hardware runs hot and requires active cooling. Data center cooling is a major operating cost and often depends on large evaporative cooling systems.
United Nations University reporting highlights rapid growth in cooling-water withdrawals for next-generation server clusters. In the United States alone, facilities have consumed billions of gallons annually for cooling. Global AI-related water demand is projected in the billions of cubic meters per year.
Solving these overlapping challenges, social equity in algorithm behavior and physical resource efficiency in infrastructure, is a core task for the next generation of computer scientists.
Recycled Water + Renewable Energy Co-location
---
Air-Cooled (Waterless) + Fossil-Fuel Grid Power
---
Direct Aquifer Cooling + Mixed Grid Power
---
Direct Aquifer Cooling + Dedicated Fossil-Fuel Peaker Plant
