IEEE Computational Intelligence Magazine - May 2023 - 61

FIGURE 1. The trend of AI-related papers published in different disciplines for the past 10 years. Data
are obtained fromWeb of Science onNov 22, 2021.
the carbon footprint of training a Transformer
model is as high as 626,155 pounds
ofCO2 emission [12]. Other concerns in
AI ethics include fairness and privacy,
which could significantly impede its progress
and usage in safety-critical real-world
applications ifnot appropriately taken into
consideration [13]. In order to support
continual progress and development of
AI, the sustainability challenges must be
takenintoserious considerationand more
efforts should be dedicated to this critical
field.
In order to clearly define what is sustainable
AI, the definition of 'sustainable
development' is first investigated. Mensah
explained the term 'sustainable development'
as " the entire issue of sustainable
development centres around inter- and
intragenerational equity anchored essentially
on three-dimensional distinct but
interconnected pillars, namely, the environment,
economy, and society " [14].
The sustainable development ofAI should
also consider these three key factors: environment,
economy, and society. As discussed
above, AI has significantly
improved our productivity and efficiency,
and leads to great economic impacts.
Despite ofits achievements in economy, it
recently faces great risks in environmental
and societal issues, for example, the significant
carbon emissions and AI ethics offairness
and privacy. Thus, sustainable AI
could be defined as developing the innovation
ofAI to promote economic development,
while simultaneously addressing
its negative impacts on the environment
and ethical concerns to the society, which
refer to environmental sustainability and
social sustainability of AI, respectively.
Under this definition, various sustainability
challenges must be addressed.
In this survey, our aim is to review the
sustainability issue of AI from two major
aspects, namely, examining the recent and
emerging technical research progresses
made in the environmental sustainability,
and secondly, focusing on the social sustainability
ofAI. In particular, we consider
the following sustainability challenges in
terms of the two aforementioned aspects
while conducting the review ofthis work.
❏ Significant carbon emissions for AI
model training and inferences.
❏ Huge efforts on data collection and
annotation.
❏ Fairness and privacy issues ofAI models
❏ Understanding the decision making
process ofAI models
❏ Ensuring the safety ofAI in terms of
adversarial attack and data poisoning.
To date, several surveys that focused
on sustainable AI [10], [11], [15], [16] have
been reported. Sustainable AI was defined
in two perspectives [11], i.e.,AI for sustainability
(e.g., using AI to address climate
change) and the sustainability of AI (e.g.,
reducing carbon emissions from AI models).
In addition to the concepts above, the
author in [11] also discussed possible highlevel
strategies for sustainable AI. Larsson
et al. discussed sustainable AI in terms of
the ethical, social, and legal challenges of
AI [10]. They also summarized some recommendations
for the development of
sustainable AI, including AI governance,
multi-disciplinary collaboration, transparency,
and accountability of AI. Kindylidi
et al. presented sustainable AI from the
standpoint ofcustomer protection, where
themain focus has been onAI-related policies
[15]. It is worth noting that the existing
works have focused on the strategic
views for sustainable AI. None of the
reviews to date, on the other hand, has
considered the technical and algorithmic
progresses towards achieving AI sustainability.
Considering that more and more
attention has been devoted to these key
technologies for achieving sustainable AI,
this technical review attempts to fill this
gap by reviewing the emerging AI technologies
(as shown in Fig. 2), highlighting
how they contribute to the sustainable
development of AI, discussing key challenges
ahead, and presenting potential
future challenges and notable research
directions worthy ofinvesting time in.
The rest of the paper is organized as
follows. The detailed technical reviews
from two perspectives of environmental
sustainability and social sustainability ofAI
are first introduced in Section II and III,
respectively. Then, detailed discussions
and analyses are provided in Section IV,
and the future research challenges and
directions are presented in Section V.
Finally, the conclusions of this paper are
made in SectionVI.
II. Environmental Sustainability of AI
With increasing size and complexity ofAI
models, one major concern is
their significant carbon footprint. For
example, a well-known study by [11]
reported that the process of training a
GPT-3 model, i.e., a natural language
processing model, can lead to 1,212,172
pounds ofCO2 emissions. This is roughly
equivalent to the CO2 emissions produced
by ten cars over their entire lifespans.
Fig. 3 shows a clear trend of
substantial CO2 carbon footprint of
advanced AI models in comparison with
daily CO2 emissions ofhuman activities.
The huge models, like Transformer and
GPT-3, are also called Red AI that pursues
better performance (e.g., accuracy)
while ignoring their huge carbon footprint
[17]. An alternative concept is Green
AI, which considers efficiency as a primary
evaluation metric alongside model
performance [17].The efficiency of AI
models in Green AI research is directly
related to environmental impacts. Thus,
MAY 2023 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 61

IEEE Computational Intelligence Magazine - May 2023

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2023