Sustainability and AI - Lecture 6
What is the concept of "Alignment" in AI-speak?
It is "Al-speak for 'Design for Values'" (Page 3).
Why is it difficult to have AI systems with computational structures that directly correspond to social values like "fairness"?
Because of the conceptual complexities surrounding what "values" are (Page 4).
What is a more realistic goal than embedding "values" into AI systems, according to the IEEE?
To embed explicit "norms" into such systems (Page 4).
How can norms be expressed computationally?
In terms of obligations and prohibitions (Page 4).
What is the "anthropocentric" view of sustainability?
"We have to protect nature for future generations," which sees an instrumental value in nature (Page 5).
What is the "biocentric" view of sustainability?
"We have to protect nature for its own sake," which sees an intrinsic value in nature (Page 5).
What is the "narrow conception of technology"?
The functional "stuff" under the control of engineers and designers (Page 7).
What is the "broad conception of technology"?
The functional "stuff" plus all the users, institutions, and social elements needed to make it work in the real world (Page 7).
What are "sociotechnical systems"?
Hybrid systems involving components from the material world as well as individual people, corporate actors (like businesses), and abstract social entities (like laws and institutions) (Page 8).
What are some of the "far-flung elements" of an AI's sociotechnical system?
Data servers, mines for rare elements, and micro-workers (Page 11).
What are "micro-workers" in the context of AI production?
Little-qualified, low-paid workers who annotate, tag, label, correct, and sort the data used to train and test AI solutions (Page 11).
What are "negative externalities"?
Costs, risks, or harms created by an enterprise that do not appear on its balance sheet and are imposed on others (Page 12).
What are some examples of negative externalities from hosting a large AI model?
Major disruptions to education, uncompensated pollution (like e-waste), and strain on the energy grid (Page 12).
What normative premise can be formulated using the idea of negative externalities?
Enterprises should insure or compensate for their negative externalities (Page 12).
QD: What is the primary meaning of sustainability? A: The acceptability of long-term environmental risks and harms (Page 13).
What is "social sustainability"?
A more recent emphasis on sustainability regarding factors like employment and political stability (Page 13).
What is the "Brundtland" definition of sustainability?
"development that meets the needs of the present without compromising the ability of future generations to meet their own needs" (Page 20).
What is the "Precautionary Principle"?
A principle for dealing with uncertainty, stating that where there are threats to the environment or human health, precautionary measures should be taken even if cause-and-effect is not fully established (Page 22).
What are the four dimensions of the Precautionary Principle?
Threat, uncertainty, action, and prescription (Page 22).
What are the "two directions" of the sustainability dilemma in AI?
"AI for sustainability" (using AI to solve problems) and "Sustainability of AI" (addressing the impacts of AI itself) (Page 25).
What is "data pollution" as defined by Hasselbalch & Van Wynsberghe?
It encompasses tangible and intangible impacts, including environmental damage, data "spills," privacy violations, and discrimination (Page 25).
What are the three levels of "power dynamics" in sustainability ethics for AI?
Micro (an engineer's choices), Meso (company/institution standards), and Macro (the global system's trajectory) (Page 26).
What is Lin's (2022) normative premise regarding climate change?
Limiting global warming to $1.5^{\circ}C$ and reaching zero emissions by mid-century are necessary steps (Page 28).
What are the three policy proposals from Lin (2022) to address AI's carbon footprint?
1. Monitoring emissions impact, 2. Climate-aware guidelines for AI training, 3. Clean energy for data centers (Page 29).
What are the three points of impact of machine learning's "footprint" (Lin 2022)?
1. The impact of training, 2. The impact of building/maintaining data centers, 3. The impact of real-life application (Page 32).
What four factors can be used to estimate the environmental impact of LLMs?
1. Training duration (computation), 2. Energy efficiency of GPUs, 3. Location of data center, 4. Offsets (if any) (Page 33).
What was the key finding in the 2025 Kenyan court case regarding carbon offsets?
A flagship carbon offset project (used by Meta, Netflix, etc.) was halted for unconstitutionally and unlawfully acquiring community land without consent (Page 35).
According to the IEEE, what should A/IS regulation, development, and deployment be based on?
International human rights standards and (in armed conflicts) international humanitarian laws (Page 36).
What are the "Ruggie principles"?
The UN Guiding Principles for Business and Human Rights, which outline a business's obligation to respect international human rights (Page 36).