Artificial intelligence

ESG stakes

Discussion paper
INTRODUCTION

The convergence of artificial intelligence (AI) and environmental,

social and governance (ESG) factors is a pivotal development in
today’s rapidly evolving business landscape. In this paper we provide
an in-depth analysis of the symbiotic relationship between AI and
ESG, outlining how AI technologies can be applied to enhance ESG
management and mitigate related risks.

EY teams discuss ESG’s significance in today’s world and its synergy

with AI technologies. It also scrutinizes how AI can be a catalyst in
achieving ESG objectives, from environmental sustainability to social
responsibility and governance. We conclude with a critical evaluation of
AI’s impact on ESG metrics and offer actionable insights for mitigating
potential challenges.

By integrating AI into ESG risk management and strategies,

organizations can not only achieve their sustainability goals, but
also unlock new avenues for driving both sustainable development
and organizational innovation.
AGENDA Background

ESG and AI overview

What is ESG?
What makes ESG so significant today?
How AI systems are related to ESG factors

How does AI enable ESG management?

How AI can make the environment more sustainable
How AI’s potential can be unlocked for social good
How AI can benefit governance

How does AI impact ESG?

Environmental impact of AI models and mitigations
Social impact of AI models and mitigations
Governance impact of AI models and mitigations
Generative AI models and ESG factors
How do AI models contribute to an organization’s
overall carbon footprint?

Appendix
How can AI assist when selecting ESG investments?
Net zero: science, finance and policies in support
of a just transition
Glossary
References
EY contacts
In today’s rapidly evolving business landscape,
CEOs are placing AI and sustainability at the
pinnacle of their strategic agendas
According to EY CEO Outlook Pulse Report, CEOs across the globe are incorporating
AI technology and sustainability into their growth agenda.

• While most CEOs are positively

adopting AI in their future strategy, the
journey towards sustainability is fraught
with challenges, since the benefits of
ESG initiatives tend to lie in the long
term.

88%
of CEOs reported existing or planned
capital investments to AI-driven
products or service innovations.

38%
of CEOs reported they prioritize
sustainability issues when making
capital allocation decisions.

• While the promise of AI is undeniable,

the exciting new technology is bringing The convergence of AI and
along new sustainability challenges. sustainability offers a compelling
solution to both challenges.
Businesses are increasingly using AI

65%
to expedite their sustainability
initiatives, particularly at a time
when there is growing pressure from
investors, regulators and broader
society for greater transparency
of CEOs say more work is needed to in ESG practices. The alignment of
address the social, ethical and criminal AI and sustainability is not just a
risks in the new AI-fueled future. strategic move, but a critical
response to meet the diverse
demands of today’s stakeholders.

4 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Background
The promise of AI offers innovative tools to tackle ESG priorities. The rise of AI
highlights the urgency to examine AI ESG stakes

AI adoption drivers

AI systems are machine-based systems • In recent years, AI models have been increasingly deployed
with varying levels of autonomy that in various domains, such as medicine, finance and education
can, for a given set of objectives, due to:
produce an output (predictions,
recommendations or decisions) using
massive amounts of data sources and
data analytics (big data).
Abundance of Increase in Advancements in
available data computational capacity AI methods and
technology

AI in support of ESG priorities

AI systems are rapidly providing new

benefits and efficiencies to organizations
around the world through new • Climate challenges: AI models can be
automation capabilities, greater ease of used to enhance climate actions through
use and accessibility and a wider variety energy management and climate change
of well-established use cases. monitoring.

Organizations are also applying AI to

tackle far-reaching challenges with
greater social and environmental impact. • Financial crimes: AI-driven technology
For example, organizations are has been extensively used in fraud
addressing skills or labour shortages or detection and anti-money laundering
helping advance ESG-related initiatives system designs.
and reducing their environmental
impact.

AI impact on ESG factors

As the deployment of AI systems • Large AI models such as deep learning (DL) and gen AI
around the world is expected to generally consume a significant amount of energy and
grow in importance in the coming generate large carbon emissions, since the process of
years, the potential challenges and training and operating large AI models requires vast amounts
risks emerging from its application of energy. This results in increased air pollution, water usage
are becoming more concerning. and carbon emissions that can accelerate climate change.

600,000 lbs
The rise in popularity of AI systems
also raises concerns about ESG factors
because of the potential impacts of AI
algorithms on ESG factors. The process of training a single deep learning natural language
processing (NLP) model can lead to approx. 600,000 lbs of
carbon dioxide emissions, similar to the amount produced by
five cars over the cars’ lifetime.[1]

96 tonnes
Google’s AlphaGo Zero generated 96 tonnes of CO2 over
40 days of research training, which amounts to 1,000 hours
of air travel or a carbon footprint of 23 American homes.[2]

5 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

ESG and AI
overview
6 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes
What is ESG?
The term environmental, social and governance (ESG) is often used interchangeably
with sustainability and corporate responsibility. It refers to the three main factors
used to evaluate a company’s sustainability and ethical impact.

ENVIRONMENTAL
Environmental criteria evaluate how sustainable a company’s
operations are. It captures an organization’s overall impact on
the environment and the potential risks and opportunities it
faces because of environmental issues, such as climate
change and measures to protect natural resources.
Examples of environmental factors that can be ESG criteria
include energy consumption and efficiency, carbon footprint
(including greenhouse gas emissions), waste management,
air and water pollution, biodiversity loss, deforestation,
natural resource depletion, clean energy and technologies.

SOCIAL
Social criteria assess how a company treats different
groups of people — its employees, customers, suppliers
and communities — and its efforts to promote diversity,
equity and inclusion.
The criteria used include employee safety, product safety,
human rights, child labour and the diversity agenda.

GOVERNANCE
Governance factors examine how a company polices itself,
focusing on internal controls and practices to maintain
compliance with regulations, industry leading practices
and corporate policies.
Examples include executive compensation policies,
financial transparency and business integrity, regulatory
compliance and risk management initiatives, ethical
business practices and financial reporting.

7 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Why is ESG so important today?
Helpful to achieve the United Nations Sustainable Development Goals

By incorporating ESG factors in the decision-making process, organizations significantly contribute to achieve the
United Nations (UN) Sustainable Development Goals (SDGs). On the corporate side, ESG considerations can be
broadly mapped to SDGs.

What are the SDGs?

The UN 17 SDGs consist of a global call to action to end poverty, protect the planet and ensure that all
people enjoy peace and prosperity.
They cover environmental sustainability, social inclusion and economic growth.
There is also a focus on health, education, gender equality and climate action.

By adopting ESG practices, businesses can contribute towards the SDGs

Ethical supply chain practices

Companies that establish and enforce ethical supply chain practices, ensuring fair labour conditions,
responsible sourcing and transparency, align with SDG 8 - Decent Work and Economic Growth. These
practices promote sustainable economic development, decent working conditions and fair trade.

Waste management
Organizations that implement waste reduction strategies, prioritize recycling and promote circular
economy principles align with SDG 12 - Responsible Consumption and Production.

Water management and conservation

Businesses that implement sustainable water management practices, reduce water consumption and
promote water conservation initiatives align with SDG 6 - Clean Water and Sanitation. Their efforts
help ensure access to clean water and contribute to the overall preservation of water resources.

Carbon footprint
When a company focuses its efforts to reduce carbon emissions and promote renewable energy, this
strategy aligns with SDG 7 - Affordable and Clean Energy and SDG 13 - Climate Action.

Relating ESG Components to the 17 SDGs

1:End Poverty
2: Zero Hunger
3: Good Health and Wellbeing
4:Quality Education
5:Gender Equality
6:Clean Water and Sanitation
7:Afortable and Clean Energy
8:Decent Work and Economic Growth
9:Industry, Innovation and Infrastructure
10: Reduce Inequalities
11: Sustainable Cities and Communities
12:Responsible Consumption and Production
13:Climate Actions
14:Life Below Water
15:Life on Land
16:Peace, Justice and Strong Institutions
17:Partnerships for the Goals

8 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

The Canadian Government has a comprehensive and aggressive 2030
emissions reduction plan
Aggressive sector decarbonization

BUILDING OIL & GAS TRANSPORTATION AGRICULTURE

$150 75% 100% $950

million Target reduction
in emissions below
Light-duty vehicle
(LDV) zero-emission
million
Investment for vehicles (ZEV) sales Investment to climate
2012 levels
net-zero buildings by 2035 mitigation, carbon
strategy by 2050 sequestration, adaption and
agricultural clean technology

Increased use of carbon Clean, reliable, affordable

pricing mechanisms electricity supply
rising Enhance support for the

$15
per tonne
deployment of commercially
ready renewable energy
technologies with over

Carbon price
annually, reaching

$170 Renewable
$850
in 2030 energy million
investment

Canada’s greenhouse gas emissions pathway to 2030 [3] (Mt CO2 eq)

9 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

How AI systems are related to ESG factors
What is AI?

AI models
AI models consist of the application of computational tools to build models from
examples, data and experience, rather than following pre-programmed rules.

1
Programs that attempt to simulate the The adoption of AI models provides the ability
behaviour of the human brain by to treat large amounts of unstructured and
learning from large amounts of data. structured data for better decision-making and
to address sustainability issues (climate issues,
e.g., deep learning
education, health). For example:
Use in medicine to diagnose diseases,
develop drugs faster, improve gene editing,
personalize treatment

2
Programs that allow machines to learn Use for environmental management:
from data and make decisions/predictions climate change modeling, monitoring
on their own. deforestation through satellite imagery
e.g., machine learning techniques analysis, energy management

3
Programs that enable computers to In the absence of proper controls, adoption
understand text and spoken words in of AI may have significant environmental,
much the same way human beings can. social and governance impacts.
e.g., large language models
An AI model is an energy consumer
through its lifecycle, yielding carbon
emissions.
AI models are exposed to regulatory,

4
reputational and business risks (e.g., data
Programs that help computers process,
privacy and transparency issues).
analyze and interpret visual data (e.g.,
digital images or video).
e.g., computer vision

10 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

How AI systems are related to ESG factors
Potential link with ESG factors

Some of the common technical, May affect different ESG criteria as illustrated in the following
regulatory and practical challenges aspects of diagram with few technical, regulatory and
through the AI lifecycle practical considerations when using AI models

AI sustainability
E AI transparency
SG
The importance of assessing the The degree to which a human can
environmental impact of AI throughout its understand the cause of the
lifecycle and its supply chain: meaning the prediction and the model outcome and can
sustainable development and use of the consistently predict the model’s result.
technology by taking into consideration its
environmental impact.

AI for sustainability problems

ES Data privacy
SG
The potential of AI to solve complicated AI systems must prioritize and safeguard
environmental and societal issues and help privacy and data rights and provide explicit
to meet the United Nations SDGs (e.g., assurances about how personal data will be
quality education, reduce inequalities, used and protected.
climate actions, innovation,
infrastructure).

AI resilience SG Compliance SG
The ability of the AI system to continue functioning AI-powered systems must comply with
even when it encounters unexpected inputs, errors
all applicable laws and regulations. For
or other forms of disruptions (environmental). The
idea is to create robust AI systems that can maintain example, the data used to train AI systems
their functionality even in the face of unforeseen should be collected and used legally and
circumstances, such as hardware failures, ethically.
cyberattacks or environmental changes. This is
crucial for safety-critical applications, where system
failures can lead to severe consequences.

Explainability/interpretability
G AI ethics
ESG
Can a human understand, challenge and A set of guidelines that advise on the
validate the inner workings and results design and outcomes of AI systems.
produced by the AI system?

The direction of the impact (+/-) of AI on ESG criteria depends on whether the
AI system is responsibly used and sustainable over its lifecycle.

11 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

How does AI
enable ESG
management?
12 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes
How can
AI promote
ESG?
AI has the potential to revolutionize the way we approach
and address global challenges.

With ESG gaining prominence over recent years

and the increasing use of AI in various domains of
society, it is critical to understand how AI can help build
a sustainable future by promoting environmental, social
and governance (ESG) practices.

Energy management Climate change monitoring Deforestation monitoring

Energy use and consumption AI models can assist Satellite imagery can detect
can be monitored through in providing accurate illegal deforestation in
the use of AI models, predictions to assist policy- real time. AI models using
which in turn can provide and decision-makers in image/video annotation can
optimized usage settings implementing more be used in conjunction with
to result in reduced effective strategies to this to identify patterns of
greenhouse emissions. mitigate the impact of forest loss. This will enable
climate change. conservation organizations
to take timely action.

Financial inclusion Health & wellbeing Employment discrimination:

AI can help FinTech AI can help health care AI can assist companies
companies provide providers improve access to in analyzing hiring and
affordable financial services quality health care promotion data and
to unbanked and excluded for underprivileged correcting for any
individuals by performing communities (e.g., use potential biases and ensure
alternative credit checks. of delivery drones). a more inclusive, fair and
objective workforce.

Corporate governance Public sector efficiency Regulatory compliance

AI can assist in analyzing AI can assist in streamlining AI can assist in monitoring
corporate governance public sector processes and large amounts of regulatory
data to assess ultimately improve service data to identify potential
organizations’ ESG delivery. breaches in a timely
performance and identify manner. This will allow
possible enhancements organizations to take
and efficiencies. proactive measures instead
of being reactive.

13 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

How can AI make the environment
more sustainable?

Biodiversity
When paired with satellite imagery, AI can assist in identifying changes in land use, vegetation,
forest cover and the effect of natural disasters. Further, AI can improve waste management
through better AI-enabled sorting across the entire waste management lifecycle.

Energy
Using neural networks, pattern recognition and fuzzy logic models, AI can assist in
reducing consumption of natural resources and energy demands associated with human
activities. For example, Chen et al. (2021) introduced an effective evaluation model
based on AI techniques that can be used for predicting energy efficiency and
conservation. The proposed model exhibits a significant energy efficiency rate of around
97.32% [4].

Water
AI can forecast stream flow and examine water quality. It can assist in predicting droughts,
as well as soil and subsurface water conditions.

Transportation
Computer vision techniques can aid decision-making in traffic management,
public transportation and urban mobility.

Air
AI can collate data from sensors and satellites and assist scientists in mixing climate
models. AI-enhanced purifiers can continually record air quality data and modify their
filtering performance as needed. In addition, AI can be used to better qualify localized
emissions from satellite remote-sensing data.

Agriculture
Farmers can use drones and satellite imagery to assess soil quality and crop productivity.
This can increase efficiency, productivity and yields. AI can also be used to monitor
illegal fishing.

14 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

How can AI potential be unlocked for social good?
New developments in AI can spur democratization of access to services and work
opportunities, which could improve the lives of many around the world. However, these
advances could be used for good or ill. They could result in creating new commercial
opportunities for business, increased fairness, increased access to health care solutions, but
they could also lead to new inequalities, biases and exclusions. Below are some of the potential
benefits that AI can unlock for social good.

Human augmentation

Also known as biohacking , human augmentation technologies can enhance human

performance for good or evil.
When used for good, these AI technologies can improve the lives of people with
disabilities, using AI-powered exoskeletons.
These exoskeletons can allow disabled people to perform physical tasks that were
previously impossible for them.
AI algorithms used with a sign language glove can enable people communicating in sign
language to verbalize their signs by converting the signing patterns to electrical signals
and spoken words.

Sensory imbalance

The five human senses offer rich territory for AI technologies and applications.
AI technology could be used to detect a person’s physical and mental wellbeing by
analyzing pitch, tone, timbre and vocabulary.
AI technology can currently analyze large amounts of data sets to predict melanomas and
be as accurate as dermatologists (see Tri-Cong Pham et al. (2021) for more details)[5].

Geographic tracking

Al technology in conjunction with Google street view can be used to analyze large
amounts of images of a city landscape to identify patterns of inequality and urban
deprivation.
AI technology can analyze and derive results that can be used to complement official
statistics such as government census programs.
AI technology can be used for tracking and controlling infectious diseases.
By analyzing travel data, news reports and other data points, a Canadian-based company
sounded the alarm early around the spread of the coronavirus in the city of Wuhan.

15 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

How can AI benefit governance?

1
FRAUD DETECTION

Financial institutions are required to monitor their customers on an ongoing basis to identify
potentially fraudulent or criminal activity between normal customer review cycles. Using AI-based
solutions, they can construct comprehensive customer profiles by leveraging additional data sources,
which more accurately pinpoint suspicious activities and assess risk across various domains. This
reformed underlying detection logic leads to the enhancement of screening and monitoring tools.
Implementing robust AI modeling techniques, such as unsupervised learning and outlier detection
models, can fine-tune thresholds for rule-based monitoring systems, leading to more resilient
thresholds supported by extensive data. Consequently, this can diminish false positives and enhance
the efficiency of the investigation process[6].

2
MONITORING AND COMPLIANCE AUTOMATION
AI’s application in monitoring and compliance automation is streamlining processes in financial
institutions. Employing natural language models, firms can efficiently scan regulatory sources,
producing consolidated and relevant summaries for senior management review. By facilitating first
drafts of policy documents, AI solutions offer a foundation for human refinement, reducing costs and
enhancing procedural efficiency. Furthermore, automating tasks with AI allows compliance officers
to focus on strategic matters, exemplifying AI’s role in identifying potential fraud or errors, akin to its
function in safeguarding data against cyber threats.[7]

3
DATA GOVERNANCE
Data governance entails using a set of metrics, standards, policies and processes to ensure
companies use customer data correctly and responsibly. In data governance, AI can be used for
various purposes. Businesses can train an AI-based solutions to help detect anomalies such as
breach in data centres as well as cyberattacks by identifying patterns of cyber threats, ensuring their
customer data is protected 24/7. AI is also useful in secure data transmission through monitoring
data traffic, leveraging advanced encryption methods and anomaly pattern recognition techniques to
safeguard against interception by cybercriminals.[8]

4
BOARD REPORTING AND GOVERNANCE ANALYTICS
Presenting accurate and concise information is paramount for the board’s effective decision-making.
However, the process of preparing reports to the board and making sure all the information is correct
and up to date can be a time-consuming task. AI technologies can be introduced to streamline this
process, linking directly to databases to generate real-time, accurate board reports. Further, AI’s
potential extends to personalizing reporting dashboards per board member, emphasizing distinct key
areas of focus, thereby enhancing efficiency and responsiveness in governance analytics.[9] [10]

16 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

How does AI
impact ESG?
17 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes
Environmental impact of AI models
Negative environmental impact of using AI

The major environmental impacts of AI are primarily related to

energy consumption, greenhouse gas (GHG) emissions and
water consumption, which can occur throughout the model
lifecycle, from the development to the deployment phases.

Model development phases over the AI system lifecycle are illustrated below.

Data Extract features from data

processing • Transform features using weights based on features importance

Design, test and evaluate the quality of:

Experimentation
• The feature selection process
• The modeling techniques
• The model architecture
• Training method to determine the model parameters

Finalize the model selection and refine the model:

Training • Performance comparison
• Refinement of the selected model
• Preparation for deployment

The best model is deployed

Inference
(deployment) • To produce trillions of daily predictions
• To serve billions of users

Each phase of the AI model lifecycle is computationally demanding and will use energy/water, so it will
have some carbon footprint on the environment.
• E.g.: The experimentation phase needs a lot of energy to calibrate the model parameters, similarly in the
training phase to evaluate several candidate models' performance for appropriate model selection.

The quantity of energy used and the carbon footprint will depend on the type of model.
• E.g.: Deep learning, natural language processing (NLP) and generative AI models are strongly
computationally demanding and hence more energy consuming than some classification models.

18 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Environmental impact of AI models
Negative environmental impact of using AI

Two different carbon emissions will be considered throughout the AI system lifecycle:
• Embodied carbon emissions from data processing and model experimentation phases
• Operational carbon emissions from training and inference phases

Embodied carbon emissions Operational carbon emissions

Data processing: Model developers will extract Training phase where several AI models'
features from data during this phase and apply performance is evaluated using extensive
weights to individual features based on feature production data with the aim of selecting an
importance to the model optimization objective. appropriate candidate. The selected model
will then be refined to prepare it for
• This will consume some energy and hence
deployment. The process often requires
produce carbon emissions.
additional hyper-parameter tuning.
• The quantity of energy (carbon emission)
needed (produced) during this phase will largely • This phase is largely computationally based.
depend on the complexity of the available data It demands a significant quantity of energy,
set, the volume and the type of data set. which depends on several factors, including
model complexity, precision of the algorithm,
data complexity and the number of models to
Experimentation: During this phase, model evaluate.
developers design, implement and evaluate
the quality of proposed algorithms, model
architectures, modeling techniques and/or Inference: the best-performing model is deployed,
training methods for determining model producing trillions of daily predictions to serve
parameters. billions of users worldwide.

• The quantity of energy consumed will depend • This phase requires energy during the full model
on the complexity and the type of the use case. deployment timeline.
• e.g., a deep learning/NLP/RL model may • The total compute cycles for inference
consume more energy than a regression-based predictions are expected to exceed the
or classification model. corresponding training cycles for the
deployed model.

19 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Environmental impact of AI models
Quantifying the carbon emissions of AI models: strategies for AI use cases

Before proposing a way to reduce carbon

emissions over an AI system lifecycle, you
should first be able to assess the carbon
emission generated by the models and
understand the factors impacting the carbon
footprint.
The energy used and the carbon emission will
be essential to understand the potential climate
impacts of machine learning models.
Several methods have been proposed to
quantify or to estimate the carbon footprint of
AI models.

The metric to quantify the carbon emission:

CO2-equivalents (CO2eq), a standard metric
used to evaluate the environmental impact.
Factors that could affect the metric:
Carbon emission calculator[11]: an alpha version has been
• The geographical zone of the server (provide proposed by Alexandre et al. (2019).
information about the energy grids used):
the distribution and variation in carbon
emissions depends on the location of the
server.
• The type of graphics processing units
(GPU) (computing infrastructure) and the
training time: models such as NN often use
multiple GPU for several weeks/months,
which requires more energy.
• The calculator uses those factors and
outputs the approximate amount of CO2eq
produced to inform model
developers/users about the model’s
potential environmental impact.

Carbon-tracker to track Carbon-tracker is an open-source tool written in Python for tracking and
and predict the energy and predicting the energy consumption and carbon emissions of training
carbon footprint of training DL models.
DL models (see Lasse et It is available through the Python Package Index (PyPi).
al. (2020) .
Carbon-tracker uses several metrics for tracking carbon footprint.
• Power usage effectiveness (PUE): This is defined as the ratio of the
total energy used in a data centre facility to the energy used by the IT
equipment such as the computing, storage and network equipment.
• Energy consumption (E) obtained by combining PUE and average
power consumed and training duration.
• Carbon footprint obtained with E and the carbon intensity.
• The carbon intensity is forecasted using application programming
interfaces (API). It refers to how many grams of carbon dioxide (CO2)
are released to produce a kilowatt hour (kWh) of electricity. This is
specific to each region.

20 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Environmental impact of AI models
Quantifying the carbon emission of AI models: strategies
for AI use cases

Example of the default setup added to training scripts for tracking

and predicting with Carbon-tracker.

Example output of using Carbon-tracker to track and predict

the energy and carbon footprint of training a DL model.

21 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Leading practice to mitigate
AI’s environmental impact
Through developing sustainable AI solutions

1
MULTI-OBJECTIVE OPTIMIZATION
Energy and carbon footprint can be directly incorporated into the cost function as optimization
objectives to enable discovery of environmentally friendly models.

2
REDUCE WASTED RESOURCES
Replacing grid search with random search can significantly accelerate hyperparameter search,
consequently reducing carbon emissions (it reduces training time during the experimental and
training phases).
Also, while failed experiments are a common part of ML research and are sometimes unavoidable, their
number can often be reduced with careful design such as unit tests, integration tests and extensive
and early debugging.

3
DEVELOP EFFICIENT TRAINING ALGORITHMS
Evaluations of optimization methods should account for all experimentation efforts required to tune
optimizer hyperparameters, not just the method performance after tuning. Efficiently scale training
by reducing communication cost via compression, pipelining (the processor performs an instruction
in multiple steps) and shading (database partitioning that separates large databases into smaller,
faster, more easily managed parts). Hyperparameter tuning may be improved by substituting grid
search for random search using Bayesian optimization or other optimization techniques like
Hyperband.

4 CARBON-EFFICIENT SCHEDULING FOR AI COMPUTING AT SCALE

Using carbon-free energy to neutralize operational carbon footprint during training and
experimentation.

5
DATA UTILIZATION EFFICIENCY
Data scaling and sampling should be well designed to improve the competitive analysis of AI
algorithms by affecting the size and quality of the training data (e.g.: Sachdeva et al, 2021):
intelligent data sampling with only 10% of data subsample can significantly reduce training time for
similar performance, leading to significant operating carbon footprint reduction.
Data perishability Understanding the rate at which data loses its predictive value has strong
implications on the resulting carbon footprint (e.g.: natural language data sets can lose half of their
predictive value in less than 7 years). This data value amortization rate will help provide effective
sampling strategies to subset data. By doing so, the resource requirement for the data storage and
ingestion pipeline can be significantly reduced leading to lower training time as well as storage
needs.

22 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Social impact of AI models

AI FAIRNESS AND BIAS

• Public training data can reflect societal biases, resulting in AI systems unintentionally
favouring or discriminating against specific sub-groups.
• Inherent bias in algorithmic design and model selection can perpetuate unfair
outcomes across various business applications.
• AI systems, especially interactive ones, may generate or amplify toxic responses,
necessitating continuous oversight and moderation to ensure respectful user
interactions.

DATA PRIVACY AND SECURITY

• AI’s ability to process large data volumes can lead to potential misuse or mishandling
of sensitive data.
• AI algorithms, particularly in machine learning, often require extensive data, which
can intrude on privacy if not adequately managed.
• AI systems can be vulnerable to cyberattacks, risking data breaches.
• Gen-AI models are subject copyright litigation. While the copyright law of fair use
allows sufficiently differentiating “derivative work,” defining its boundaries is
challenging. In addition, the intricate nature of generative AI makes tracing the
specific training data for outputs difficult.[14]

DATA PRIVACY AND SECURITY

• AI integration disrupts roles with repetitive tasks (e.g., manufacturing, data entry,
customer service).
• AI’s impact is task based rather than job based, indicating a workforce shift towards
higher-value tasks.
• While immediate impacts of AI can lead to substantial job displacement, the long-
term outlook hints at a potential adaptation of the workforce and the emergence of
new job opportunities.

23 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Mitigating AI’s impact
on social

1
DATA SECURITY AND PRIVACY
• Implement robust encryption and
anonymization techniques to safeguard data;
ensure sufficiency of consent management
for privacy and confidentiality; adhere to the
AI Acts and any other regulations that may be
violated by using AI.

2
AI FAIRNESS AND BIAS
• Use diverse and representative data for
training to minimize biased outcomes; promote
transparent algorithms and interpretability to
shed light on the decision-making process.

3
JOB DISPLACEMENT
• Initiate reskilling and upskilling programs
to prepare workers for AI-related job
transformations; establish social safety nets,
including unemployment benefits, to support
displaced workers.

24 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Governance impact of AI models

TRANSPARENCY
AI has a serious transparency problem. In fact, due to the complexity of most AI algorithms,
its outcomes are difficult to explain and its processes impossible for lay users to understand.
Hence, it becomes difficult for end users to have knowledge about and control over what data
is being captured and how it is used. This is due to:
• Limited understanding of bias in training datasets
• Lack of visibility into training datasets
• Lack of visibility into the method of data selection
• Some difficulties to explain algorithms

EXPLAINABILITY
AI models, in general, are highly complex and lack inherent explainability, unlike traditional
mathematical/statistical models. Explainability is crucial to understand the underlying
mechanisms that drive the operation of AI systems useful to produce a trusted AI output.

BIAS ISSUES
The inherent biases arising from the composition of the development team, data and training
methods are difficult to identify due to the structure of the model. So the performance
comparison with alternative models should be done with caution. This will also limit the model
trust since end users do not have clear idea about the importance of the bias.

PERFORMANCE ISSUES
AI models tend to have higher complexity, higher data consumption and dependency, lower
explainability and lower stability than traditional models. As a result, an appropriate model
performance monitoring plan is needed (model performance monitoring challenges) to ensure
the performance is compatible with end users’ expectations.

25 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Mitigating AI’s impact on governance
Through responsible development of AI solutions in organizations using the
following steps

KEY ACTIVITIES OUTCOMES

Build an AI governance framework to enable Consistent definition of AI and AI system

AI adoption and manage emerging risks. risk tiering
Set up AI policies, procedures and guidelines Clearly defined roles and responsibilities
BUILD

to streamline people, process and technology across the AI model lifecycle (RACI matrix)
pillars.
Model inventory framework
Ethics and privacy assessments framework
AI standards for model development and
validation

Enable the technology supporting AI model Operationalization of the AI governance

OPERATIONALIZE

inventory, privacy and ethics questionnaires. framework across pillars

Log the institution’s AI systems in the model Vetting of AI systems across principles of
inventory. trusted AI to ensure performance, lack of bias,
transparency, resilience and explainability
Set up cadences for executive committees,
escalation forums. Remediation of risks and issues identified in
validation against standards
Vet current AI systems, including
remediation plans for issues identified.

Support institution-wide enablement of Consistent, streamlined adoption of

trusted AI principles. trusted AI principles across functions (e.g.
fairness, ongoing monitoring)
For example, deploy fairness toolkit or
ongoing monitoring toolkit. Acceleration of trusted AI enablement
through technology
SCALE

From a technology point of view, support

enablement of trusted AI through an end-
to-end XOps framework.

26 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Generative AI models and ESG factors
AI drivers – emergence of generative AI models

Generative AI (gen AI) is a branch of AI that

uses ML techniques to generate a wide variety of
new contents in the form of images, text, audio,
etc., instead of simply analyzing and regenerating
existing data.
Gen AI uses sophisticated algorithms to assess
data and derive novel and unique insights, thereby
improving decision-making and streamlining
operations.
Gen AI is becoming very popular and is increasingly
GENERATIVE used in various domains because of the multiple
AI benefits associated with these models in terms
of creating new contents, process automation,
efficiency improvement, etc.
However, the adoption of gen AI models gives rise to
various challenges and risks that could significantly
affect ESG factors.

A few examples of how gen AI can impact ESG factors

ENVIRONMENTAL IMPACTS
• These models require substantial computing power and energy consumption compared to other AI models
(classification, regression-based models).
• Hence, they significantly contribute to global carbon emissions, exacerbating climate change concerns.
• A study by Strubell et al. 2019 among others estimates that training a LLM can emit over 626,000 lbs of
carbon dioxide, similar to the amount of carbon dioxide emissions produced by five cars over the cars’ lifetime.

SOCIAL IMPACT
• Gen AI may produce biased or discriminatory output, perpetuating stereotypes or promoting harmful narratives.
• Such incidents can harm an organization’s reputation, violate ethical standards and impact social harmony.
• Gen AI such as LLMs relies on large amounts of data for training and generating outputs and may
expose an organization to data privacy and security (cybersecurity exposure) issues.
• Data breaches or unauthorized access to AI models can lead to severe reputational damage and legal
repercussions.
• Moreover, the integration of gen AI can reshape the workforce landscape, potentially leading to job displacement
and socioeconomic challenges.
• Due to their ability to automate repetitive tasks and generate content, specific job roles may become
obsolete or require significant reskilling and upskilling efforts.

GOVERNANCE CHALLENGES
• Explainability issue: The large number of model parameters (~100b) makes most gen AI models such as LLMs a
black box lacking explainability.
• This makes it difficult for model users to understand the logic behind certain decisions and may affect trust in
these decisions.
• Cyberattack and adversarial attack: Training data and trained LLMs may be leaked out of the institution or
vendor platform due to cyberattack or adversarial prompt engineering.

To mitigate these risks, organizations should develop a responsible and sustainable Gen AI

27 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

How do AI models contribute to an
organization’s overall carbon footprint?
AI models emission impact

Scope 2 Scope 3
Scope 2 emissions are indirect Scope 3 emissions are other indirect
emissions from energy purchased. emissions upstream and downstream
AI model training and deployment Hardware production
Training often involves iterating over AI hardware such as GPUs and TPUs
large datasets multiple times, requiring have an energy-intensive production
vast computational resources with high process
energy usage.
End user impact
Deployment environments, such as
AI-driven features, especially those
cloud servers or edge devices, also
backed by resource-heavy algorithms,
require power to host and run these AI
can lead to increased device workloads,
models.
consuming more power.
Data storage
Inefficient AI software frameworks
Data for AI modelling purposes is stored or algorithms might require end
in data centres, consuming vast amount users to run their devices for longer
of energy. periods or at higher intensities,
indirectly leading to increased energy
use.

AI models emission quantification

1
Start with a comprehensive look
DATA CENTRE
at emissions on the data
A physical facility organizations use to centre’s aggregated scale.
house their critical applications and data

2
TENANT
Dive into container-specific energy
An individual client or organization uses to further break down the
accessing specific parts of the cloud emission composition of data centres.
environment
CONTAINER

3
Packages of software that contain all of the Deriving from the aggregated power
necessary elements to run in any environment consumption at hardware levels,
identify the granular energy usage at
WORKLOAD the software/workload level for specific
AI-related tasks.
A resource running on the cloud consuming
compute and possibly storage

28 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

How can AI assist when selecting ESG
investments?

AI
allows investors to collect and analyze large amounts of
information and can help sustainable investors process
data that contains essential ESG information.
By nature, ESG data is very different from traditional
financial data since there is no standardization and the data
is highly unstructured. This makes data retrieval difficult
and impossible to retrieve via traditional NLP approaches.
AI models can assist with retrieval of ESG data.
Al models can be trained to analyze tone (sentiment
analysis) and content and digest all information to form a
holistic view point on an organization's commitment and
approach to ESG initiatives.

1 Allows detailed analysis

of large amounts of data,
which previously was
very time consuming and
2 Using improved
algorithms and tools such
as sentiment analysis and
natural language
resource heavy. processing, AI can
unearth key data and
filter out irrelevant data.

4 Allows companies to
view their current ESG
position and make
strategic decisions to
3 Consumers and portfolio
managers can now have
access to companies'
ESG investing statistics
create improvements and make informed
where needed. decisions when selecting
their investments.

29 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Net zero: science, finance and policies in
support of a just transition
Why commit to net zero and how do we pave the path towards it?

Climate science Financial institutions

The world is rapidly warming, leading to a rise in sea Financial institutions play a pivotal role in driving sustainability
levels and increased droughts, floods and wildfires. through their funding and investment decisions. Economic
Alongside these tangible consequences, there is activities (assets and revenues) are classified based on their
also the destruction and alteration of natural alignment with the 1.5°C scenario to be able to access green
habitats and drastic changes in local temperatures. finance or transition finance.
Climate science shows these effects will become
Green: Provide dedicated financial support for economic
significantly worse if we continue on this path.
activities that are aligned with the 1.5°C scenario by
According to climate science, compared to a positively contributing to eco-friendly initiatives.
1.5°C world, in a 2°C world we would experience Financial institutions can design and promote financial
1.7 billion more people facing heatwaves annually, instruments that cater to green projects, including green
several hundred million more people would be bonds and green loans.
exposed to climate-related risks and poverty, global
Transition: Recognizing that not all industries can switch to
fishery catches could decline by another 1.5 million
green alternatives immediately, financial institutions can
tonnes, and there would be a drastic increase in
provide transition financing. This supports companies in the
drought risk for the Mediterranean region (see
interim period to gradually shift from brown to green
Amir et al. (2019)).
practices.
Limiting global warming to 1.5 °C will require the
Brown: To actively discourage non-sustainable practices,
global economy to release zero greenhouse gas
financial institutions can divest from, or reduce, their
emissions by 2050, and nearly half of those
exposure to brown practices, which are often plagued by high
reductions must have happened before 2030.
transition costs, limited clean energy access or high exposure
This highlights the imperative to achieve net-zero
to physical risk.
carbon emission for the global temperature to
stabilize and provide a guideline of actions to
be undertaken by policymakers and organizations
around the world.
Policymakers

To ensure a seamless and orderly transition towards As global economies gear up for a sustainable transition in
a sustainable future, the role of policymakers line with the net-zero initiative, regulators’ role becomes
and financial institutions is paramount. Their increasingly indispensable. Their involvement can shape the
support, guidance and initiatives not only lay the trajectory of this transition by regulating the financial
foundation for a greener and more sustainable markets, providing incentives and facilitating collaborations.
tomorrow, but also to establish the framework Regulation of financial markets: Policymakers
within which businesses and individuals can set guidelines for the financial sector to promote
operate. Their proactive involvement makes the sustainability. This includes crafting clear green
path to sustainability not only aspirational but taxonomies for financial products, imposing disclosure
also actionable, creating a balance between the requirements for transparency and laying down risk
organization’s current needs and the future it management parameters for key financial players like
aims to shape. banks, insurers and asset managers.

Public sector incentives and enablers: Harnessing the

public sector’s influence, regulators drive the green
transition by investing in sustainable infrastructure and
promoting green innovation. They use subsidies and tax
incentives to encourage sustainable practices, while also
introducing disincentives for environmentally detrimental
activities.

Facilitate collaborations: Policymakers prioritize

collaboration across government departments. By
combining insights and weighing both environmental and
societal factors, they develop comprehensive policies
aiming for a balanced transition.

30 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

Glossary
CARBON REDUCTION
Carbon reduction refers to the act of decreasing or mitigating the emission of greenhouse gases, particularly
carbon dioxide (CO2), into the atmosphere. It is a crucial strategy to combat climate change and achieve a more
sustainable and low-carbon future.

DISTILLATION
Distillation in AI refers to the process of transferring knowledge from a large model to a smaller one. The objective
of distillation is to reduce the size of the training dataset to improve accuracy.

MODEL PRUNING
AI model pruning refers to the process of removing unneeded parameters or connections from a model in order to
simplify it. This results in performance improvement by reducing its complexity and making it easier to train and
deploy. In addition, pruning can help prevent overfitting by reducing the number of parameters that can be
tuned.

QUANTIZATION
AI quantization refers to the method of reducing computational demands. It is a model size reduction technique
that converts model weights from high-precision floating point to low-precision floating point. This results in
improved performance and power efficiency by reducing memory access and increasing computing efficiency.

GRID SEARCH
Grid search is an exhaustive search technique used to identify the optimal hyperparameters for a given model. The
process involves evaluating the model performance for every combination of specified hyperparameters.

RANDOM SEARCH
Random search is a strategy in machine learning that employs random combinations of hyperparameters to identify
the optimal solution for a given model. Unlike grid search, which exhaustively explores all possible combinations,
random search samples points from a bounded domain of hyperparameter values.

NET ZERO
Net zero refers to a state in which the greenhouse gases going into the atmosphere are balanced by removal out
of the atmosphere. To "go net zero" is to reduce greenhouse gas emissions and/or to ensure that any ongoing
emissions are balanced by removals.

31 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

 References
1 Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and Policy Considerations for Deep Learning in NLP.
In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3645–3650, Florence, Italy.
Association for Computational Linguistics.

2 Karen Hao: Training a single AI model can emit as much carbon as five cars in their lifetimes (2019).

3 2030 Emissions Reduction Plan – Sector-by-sector overview - Canada.ca

6 How AI will affect compliance organizations | EY – US

7 Lin Chen, Zhonghao Chen, Yubing Zhang, Yunfei Liu, Ahmed I. Osman, Mohamed Farghali, Jianmin Hua, Ahmed Al-Fatesh, Ikko
Ihara, David W. Rooney & Pow-Seng Yap: Artificial intelligence-based solutions for climate change: a review (2023).

8 Tri-Cong Pham, Chi-Mai Luong, Van-Dung Hoang & Antoine Doucet: AI outperformed every dermatologist in dermoscopic melanoma
diagnosis, using an optimized deep-CNN architecture with custom mini-batch logic and loss function (2021)

9 Brian Benjet, Bennett Borden, Katrina Hausfeld, Gregory Ferroni: AI tools are already here: How they can help compliance officers,
and four general principles.

10 Ronald Van Loon: AI Is Transforming Real-Time Data Governance (2023)

11 Diligent Corporation: Crack the code on your cyber risk reporting.

12 permutable.ai: The Transformative Role of Data and AI in Corporate Governance (2023).

13 Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, Thomas Dandres: Quantifying the Carbon Emissions of Machine Learning
(2019).

14 Peter Henderson, Jieru Hu, Joshua Romoff, Emma Brunskill, Dan Jurafsky, Joelle Pineau: Towards the Systematic Reporting of the
Energy and Carbon Footprints of Machine Learning (2020).

15 Lasse Heje Pedersen, Shaun Fitzgibbons, Lukasz Pomorski: Responsible Investing: The ESG-Efficient Frontier (2019).

14 Gil Appel, Juliana Neelbauer, David A. Schweidel: Generative AI Has an Intellectual Property Problem (2023).

15 Lasse Heje Pedersen , Shaun Fitzgibbons , Lukasz Pomorski: Responsible Investing: The ESG-Efficient Frontier (2020)

16 Amir Bazaz, Paolo Bertoldi, Marcos Buckeridge, Anton Cartwright, Heleen de Coninck, François Engelbrecht, Daniela Jacob, Jean-
Charles Hourcade, Ian Klaus, Kiane de Kleijne, Shauib Lwasa, Claire Markgraf, Peter Newman, Aromar Revi, Joeri Rogelj, Seth
Schultz, Drew Shindell, Chandni Singh, William Solecki, Linda Steg, Henri Waisman: WHAT THE IPCC SPECIAL REPORT ON GLOBAL
WARMING OF 1.5°C MEANS FOR CITIES November 2019.

17 Dr. A. Shaji George, A S. Hovan George, A. S. Gabrio Martin: ChatGPT and the Future of Work: A Comprehensive Analysis of AI's
Impact on Jobs and Employment (2023).

32 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

 Contacts
To discuss further, please contact your Canadian EY advisor:

Mario Schlener
Canada FSRM Lead and Global FS Risk
Technology/Alliance/Innovation Lead,
Partner
mario.schlener@ca.ey.com

Katerina Kindyni
Canada FSRM Senior Manager,
ESG Risk Lead
katerina.kindyni@ca.ey.com

Yara Elias
Canada FSRM Senior Manager,
AI Risk Lead
yara.elias@ca.ey.com

Yara Mohajerani
Canada FSRM Manager,
ESG Risk
Yara.Mohajerani@ca.ey.com

Cindy Ning
Canada FSRM Senior Consultant
ESG Risk
cindy.ning1@ca.ey.com

N’Golo Kone
Canada FSRM Senior Consultant
AI Risk
ngolo.kone@ca.ey.com

Aslam Mayet
Canada FSRM Manager
Non-Financial Risk
aslam.mayet@ca.ey.com

33 Artificial Intelligence (AI) Environmental, Social, Governance (ESG) Stakes

EY | Building a better working world

EY exists to build a better working world, helping to create long-

term value for clients, people and society and build trust in the
capital markets.

Enabled by data and technology, diverse EY teams in over 150

countries provide trust through assurance and help clients grow,
transform and operate.

Working across assurance, consulting, law, strategy, tax and

transactions, EY teams ask better questions to find new answers
for the complex issues facing our world today.

EY refers to the global organization and may refer to one or more, of the
member firms of Ernst & Young Global Limited, each of which is a separate
legal entity. Ernst & Young Global Limited, a UK company limited by
guarantee, does not provide services to clients. Information about how EY
collects and uses personal data and a description of the rights individuals
have under data protection legislation are available via ey.com/privacy. EY
member firms do not practice law where prohibited by local laws. For more
information about our organization, please visit ey.com.

This material has been prepared for general informational and

educational purposes only and is not intended, and should not be relied
upon, as accounting, tax, legal or other professional advice. Please refer
to your advisors for specific advice.

Neither the EY organization nor any of its member firm thereof shall
bear any responsibility whatsoever for the content, accuracy, or security
of any third-party websites that are either linked (by way of hyperlink or
otherwise) or referred to in this presentation.

The views of third parties set out in this webcast are not necessarily the
views of the global EY organization or its member firms. Moreover, they
should be seen in the context of the time they were made.

© 2023 Ernst & Young LLP.

All Rights Reserved.

4313357/4261712
ED None

This publication contains information in summary form, current as of

the date of publication, and is intended for general guidance only. It
should not be regarded as comprehensive or a substitute for
professional advice. Before taking any particular course of action,
contact Ernst & Young or another professional advisor to discuss these
matters in the context of your particular circumstances. We accept no
responsibility for any loss or damage occasioned by your reliance on
information contained in this publication.

ey.com