13 June 2024

Chapter I :Core Concepts of AI –

- Turing Test: test to determine if the machine is intelligent. Machine is considered intelligent if it fooled
the interviewer to believe that it is human.
- Common elements in various definitions of AI
o Technology
o Autonomy
o Human involvement
o Output
- Process of teaching AI systems is ‘machine learning’. There are 3 types:
o Supervised – Using ‘labelled’ data, e.g. images of cats and dogs classified into categories
o Unsupervised – Using ‘unlabelled’ data, e.g. fraud detection, genetics
o Reinforcement – Uses reward and punishment matrix
- AI is a socio-technical system. AI influences society and society influences AI.
- OECD Framework for classifying AI systems, which has main dimensions:
o People and planet dimensions: Identify the individuals impacted, privacy rights, human
rights, environment and society in general
o Economic context: Sector where the AI system operates, business function, technological
maturity of the sytem
o Data and input: what data was used, whether any expert input was used, who collected the
data (machine or human), structure of the data
o AI model
o Tasks and output: System tasks,
- AI use-case types:
o Recognition (image, speech, face), plagiarism detectors, retailer product matches,
o Detection: fraud detection, event detection, cyber events, systems management
o Forecasting: sales, revenue, demand, weather, ride-sharing pricing
o Personalization: Customer experience,
o Interaction support: chatbots,
o Goal-driven optimization: Supply-chain, driving-routes, idle-time,
o Recommendation: based on predictive analytics, product, viewing, medical diagnosis,

History of AI and the evolution of data science

- Dartmouth conference, 1956: Term AI coined

- Cycle of AI
o 1st AI summer (mid-50s to mid-70s): John McCarthy created the 1st AI programming language
LISP. Joseph Weizenbaum at MIT developed ELIZA
o 1st AI winter (mid-70s to mid-80s):
o 2nd AI summer (mid-80s to late-80s)
o 2nd AI winter (late-80s to late-90s)
o Renaissance and the era of “big data” (late-90s to 2011)
o AI boom (2011 to present): advancements in deep learning
- Emergence and growth of data science
o Foundations (1960s – 80s)
o The age of databases (1980s – 1990s): Advent of Relational Database Management Systems
(RDMS) and Structured Query Language (SQL)
o The advent of the Internet (1990s – 2000s): ‘big data’ and ‘data mining’
o Rise of data science: Hadoop
 o Current trends:

Types of AI

- Artificial narrow intelligence: Weak AI, e.g. chess game

- Artificial general intelligence: Strong, deep or full AI, closely mimics human intelligence
- Artificial super intelligence
- Broad AI – In between ANI and AGI – involves reliance on a group of AI systems capable of working
together, e.g. autonomous driving vehicles

Supervised machine learning model categories:

- Classification models: produce output in the form of specific categorical responses

- Regression Models: predict a continuous value

Unsupervised machine learning model categories:

- Clustering: groups data points that share similar or identical attributes.

- Association rule learning: identifies relationships and associations between data points

Semi-supervised learning models – uses a combination of supervised and unsupervised learning

processes. Uses small amount of labelled data and large amounts of unlabelled data

LLMs – rely on semi-supervised learning models.

Robotics and AI:

- machine perception: systems are trained to process sensory information and mimic human senses.
- Robotic process automation

Expert systems: AI intended to mimic the decision-making abilities of a human expert in a specific field. 3
main elements:

- The knowledge base

- The inference engine
- The user interface

Fuzzy logic

Fuzzy rule e.g. if temperature is hot, then set the fan speed to high. Employ fuzzy inference mechanisms to
make decisions based on fuzzy rules and input data. Standard steps:

- Fuzzification
- Rule evaluation
- Aggregation
- Defuzzification

e.g. climate control systems, traffic management systems

Chapter 2: AI technology stack, platforms, applications and model types:

An AI model typically comprises input data, a pattern-matching algorithm and an output classification

AI platform = cloud platform

AI systems = AI application

Commonly-used AI models:

- Linear and statistical models: model the relationship between two variables,
 - Decision-tree models: predict outcomes based on a flowchart of questions and answers

Both are more explainable, no blackboxes

- Machine learning models: black box capabilities, lack transparency and explainability, e.g. neural
network which is commonly behind technologies such as facial recognition. Different types of neural
networks are:
o Computer vision models
o Speech recognition models
o Language models, such as NLP
o Reinforcement learning models, rewards and penalties,
o Robotics applications

AI Technology Stack

- Compute: Consists of CPU and GPU (Graphical Processing Unit). GPU has specialised chips. As
algorithms have developed so have GPUs similarly. Serverless: not limited to any particular hardware.
Improves scalability. High performance compute. Aids quantum computing (processing data in 3
dimensions). Trusted Execution environments.
- Storage. 4 stages of AI, all of which require storage
o Ingestion
o Preparation
o Training
o Output
Considerations around storage – expense, storage types, structured vs unstructured data, flexible
storage (AI at scale)
- Network: needs to be fast (high-speed), considerations – Edge computing (IoT), need protocols based
on congestion-free design.
- Software: Tuning AI systems – customize AI, Tuning challenges
o No. of hyperparameters, common hyperparameters are : learning rate, number of epochs
(one cycle through the full training dataset) and momentum (the amount of history included
in the equation)
o Varies on model type and complexity
o Scaling – trial and error tuning
- Observability and Modelling. AI observability is a subcomponent of data observability, Provides indices
and metrics for performance. Challenges – Data Integrity, Data Drift, bias and discrimination, outcome
validation. Opensource AI frameworks – e.g. Tensorflow, PyTorch

Chapter 3: AI impact on people and responsible AI principles

- Harms to individuals, based on bias, types:

o Implicit bias
o Sampling bias
o Temporal bias – based on time
o Overfitting to training data: model works for the training data but not for any new data
o Edge cases and outliers

Areas of discrimination in AI:

o Employment and hiring discrimination

o Insurance and social benefit discrimination
o Housing discrimination
o Education discrimination
o Credit discrimination
 Civil and privacy rights:

o Face identification software biases towards gender

o How/ to screen out personal data?
o How to deidentify personal data? Risks of reidentification
o Issue of appropriation of personal data for model training
o Issue of identifying individuals via inference: AI could identify a wrong individual
o Lack of transparency of use
o Inaccurate AI models can cause privacy issues
o Bias and fairness impact privacy

Economic impact – jobs will be lost, AI driven recruiting practices may also be discriminatory

- Group Harms
o Mass surveillance, potential harm for marginalized groups
o Harms to freedom of assembly and protest
o Deepening of social and racial inequalities
o Societal harms – democratic participation and process, disinformation, ideological bubbles
o Deepfakes
o Safety, lack of human oversight – e.g. use of lethal autonomous weapons
- Environmental harms
o Energy consumption for training
- Organizational harms - identification
o First identify the owner of the risks – BU, representative of the customer, AI risk manager, AI
SME, compliance, legal and privacy officer, IT leaders – CISO, CTO, CSO, CIO, execs, Board
o Defining organizational harms – reputational, cultural and societal (wrong assumption of
accurate outcome, bias), acceleration (volume, speed and complexity means all risks can be
anticipated, legal and regulatory risk.
- Characteristics of trustworthy AI systems - Trustworthy AI operates in an expected, legal and fair
manner
o Human-centric
o Accountability (organisations to be responsible)
o Transparent
- Responsible AI practices with risk management framework
o Transparent and explainable
o Fairness
o Human oversight

How to operationalize responsible AI practices:

o AI playbooks
o Technical standards
o Guidelines
- Existing and emerging ethical guidance on AI
o OECD AI Principles in 2019
o White House Office of Science and Technology policy blueprint for AI Bill of Rights
o UNESCO Principles
o Asilomar AI Principles
o The Institute of Electrical and Electronics Engineers Initiative on Ethics of Autonomous and
Intelligent Systems
o CNIL Action Plan
- OECD AI Principles
o Inclusive growth, sustainable development, and well-being
 o Human-centred values and fairness
o Transparency and explainability
o Robustness, security and safety
o Accountability (for proper functioning of AI systems
- Factors involved in creating ethical AI
o Legal and compliance
o Equitable design
o Transparency and interpretability
o Privacy and cybersecurity
o Data governance

Chapter 3: AI Development Life Cycle

- Planning
o Business objectives and requirements
 Classification business problem
 Regression business problem: Predicting future based on past data
 Recommendation business problem
How to define the problem?
 What type of results are you expecting from the use of AI system?
 What processes do you have to solve the problem?
 What resources are available to use to solve the business problem?
 User interviews, market research, identify AI use-cases, finding the right
data to use
 Scope of the project?
 Impact of the AI system on the problem
 Effort needed
 Fitment
 Governance structure
- Design: Implementing a data strategy
o Gathering data –
 What data is required?
 What type of data is required?
 How much data is required?
 How it is collected and where it is stored
 Use pre-trained data?
 Internal data or external data
 Quality of data
 Format of the data
 Structured data
 Unstructured data
 Static data
 Streaming data
o Wrangling the data, preparing the data (constitutes 80% of AI lifecycle) – 5 Vs
 Volume of the data
 Velocity of the data
 Variety of the data, structured v unstructured
 Veracity of the data
 Value of the data
o Cleansing – remove erroneous or inaccurate data
o Labelling the data
 - Development:
o Build the model – Use same features for training and testing
 Feature engineering: transforming raw data into features, use SMEs. Purpose of
feature engineering are:
 Improving model performance
 Reducing costs, improving effectiveness
 Boosting model explainability
o Model training – train, test, evaluate and retrain
o Model testing and evaluation – metrics should be determined earlier. Also test on new data
- Implement
o Perform a readiness assessment of the model
o Continuous monitoring
o Baseline to measure future iterations
o Maintain the model
o Avoid model drift

Chapter 4: Implementing Responsible AI Governance and Risk Management

- Risks within AI algorithms and models:

o Security and operational. For GenAI
 Hallucinations
 Deepfakes
 Training data poisoning – use AI to hack other AI
 Filter bubbles – echo chambers – GenAI just repeats what we told it.

For general AI

 Eats into individual freedom

 Overreliance of AI – false sense of security, DDoS using AI only
 Misuse of AI

Operational risk

 High costs, hardware cost, CPUs, GPUs, environmental cost

 Data corruption and poisoning
o Privacy
 Data persistence
 Data repurposing
 Data spillover - data collected from unintended subjects
 Data derived from AI itself
o Business
- Principles of AI risk management
o AI Risk management – organizational considerations
 Adopt a pro-innovation mindset
 Ensure planning and design is consensus driven
 Ensure that the team is outcome-focused
 Ensure that the framework is law, technology and industry agnostic
 Adopt a non-prescriptive approach to allow for intelligent self-management
 Ensure that governance is risk-centric
 Create policies to manage third-party risk, ensuring end-to-end accountability
o EU model classification – unacceptable, high, limited, minimal
 Unacceptable – social credit scoring system, facial recognition systems: ban
  High – AI used in the field of education: conduct a conformity assessment. Needs to
be listed in a publicly accessible database
 Limited – use of chatbots: be transparent
 Minimal – spam filters: no requirements
o Establishing AI governance and strategy
 AI Governance models – centralized, federated (hybrid), decentralized


Chapter 5: Implementing AI Projects and Systems

- Mapping, planning and scoping the AI project

o Huderia risk index
o Probability and severity matrix
o Confusion matrix
- Testing and validating the AI system during deployment – accuracy, privacy, interpretability, bias,
integrity, reliability, robustness
o Types of bias – computational, societal, cognitive
o Conduct adversarial testing and threat modelling
o Establish multiple levels of mitigations
- Managing and monitoring AI systems after deployment

Chapter 6 – Current laws that apply to AI systems

- SR-7 (a regulatory standard set out by the U.S. Federal Reserve Bank that provides guidance on model
risk management)
- AI conformity assessment mandated by the EU AI Act – required whether or not personal information
is being processed
- 2 proposals in the EU for affixing AI product liability issues- both operate in parallel
o Reform of the 1985 product liability directive
 AI and other software will fall within the scope (product)
 Strict liability regime
 Damage – injury, death, psychological harm, property/ financial damage, data loss/
corruption
 Extraterritorial
o EU AI Liability directive
 Fault liability regime
 Reinforces the EU AI Act
 Courts are empowered to order disclosure in relation to high-risk AI systems
 Courts can presume a casual link between non-compliance with AI act and harm
- US Approach – product liability laws are determined at the state level – 3 types of liability principles
o Strict liability
o Negligence
o Breach of warranty
- Not clear in the US if AI is a product
o Rodgers vs Christie (New Jersey) – AI does not qualify as product.
o Connecticut Fair Housing vs CoreLogic Rental Property Solutions. Rejection of housing
application due to disability. AI vendor may be made liable
o White House Blueprint of AI Bill of Rights
 Safe and effective systems
 Algorithmic discrimination protections
 Data privacy
 Notice and explanation
 Human alternatives, consideration and fallback.
 o Presidential Executive Order on advancing racial equality
o Executive Order on the Safe, Secure, and Trustworthy Development and Use of AI
 Artificial Intelligence must be safe and secure
 Promoting responsible innovation, competition and collaboration
 Responsible development and use of AI require a commitment to support American
workers
 AI policies must be consistent with the dedication to advancing equity and civil
rights
 The interests of Americans who increasingly use, interact with, or purchase AI and
AI-enabled products in their daily lives must be protected.
 Americans’ privacy and civil liberties must be protected as AI continues advancing.
 It is important to manage the risks from the Federal Government’s own use of AI
and increase its internal capacity to regulate, govern, and support responsible use of
AI to deliver better results for Americans.
 The Federal Government should lead the way to global societal, economic, and
technological progress,

Chapter 6 - Existing and emerging AI laws and standards

- EU AI Act: Extraterritorial scope

o Provider – carries majority of the obligations
o User
- Exemptions – Military context, R&D
- Classifications of AI risks
o Unacceptable
o High, 2 types
 Listed in Annex III of Article 6 – 8 areas listed
 Biometric identification and categorization of natural
persons
 Management and operation of critical infrastructure (such as
gas and electricity)
 Education and vocational training
 Employment, worker management and access to self-
employment
 Access to and enjoyment of essential private services and
public services and benefits (e.g., emergency services
dispatching)
 Law enforcement
 Migration, asylum and border control management
 Assistance in legal interpretation and application of the law

 AI systems which are safety components of other products or the product is a AI

system itself
 Conformity assessment should follow the existing procedures of product
safety laws
 High risk AI system should keep a log of events automatically. Functioning should be
traceable. Transparency and provision of information. Human oversight
requirements, accuracy, robustness and cybersecurity. Quality management system,
conformity assessment
o Responsibilities of users/ deployers of high risk AI systems
 Use in accordance with the instructions of use document
 Monitor the system and suspend in case of serious issues
 Keep the provider or distributor of the system updated about serious incidents or
malfunctioning
  Keep logs that are automatically generated
 Assign human oversight to appropriate individuals
 Cooperate with the regulator
o Importers/ distributors to ensure that the conformity assessment has been completed and CE
mark is displayed – they cannot place the product on the market
o Registration and notification – High risk AI system
 EU wide database for High Risk AI system- public website
 Registration required –
 Notification requirement – providers have to establish and document the post
market monitoring system – report serious incident or malfunctioning which
threatens the breach of the obligations under EU law to protect fundamental rights
(similar to data breach notification requirements). Must be reported within 15 days
o GPAI – General Purpose AI
 An AI model that displays significant generality and can perform a wide range of
distinct tasks, regardless of how the model is released on the market
 Can be integrated into a variety of downstream systems or applications
 For "all other" GPAI, obligations include:
 Maintaining technical documentation
 Making information available to downstream providers who integrate the
GPAI model into their AI systems
 Complying with EU copyright law
 Providing summaries of training data
 For GPAI "with systemic risk," where the definition is based on computing power and
substantial compliance requirements, the obligations are the four items listed above,
plus:
 Assessing model performance
 Assessing and mitigating systemic risks
 Documenting and reporting serious incidents and action(s) taken
 Conducting adversarial training of the model (also known as "red teaming")
 Ensuring security and physical protections are in place
 Reporting the model's energy consumption
o European AI Office and AI Board established
o Sectoral regulators will enforce AI Act for their sector
o Penalties
 7% of global turnover or EUR 35 Mill for using prohibited AI
 3% of global turnover or EUR 15 Mill for noncompliance with use of non-prohibited
AI.
 1% of global turnover or EUR 7.5 Mill for supplying incorrect information to
authorities
o Timeframe – 2 years by August 2026 with full application by August 2027
o Top 5 tips for professionals for implementation
 Identify high risk AI systems
 Those systems are within the territorial scope of the AI act
 Are you a provider or deployer?
 AI procurement processes and policies
 Undertake a gap analysis for compliance
 Keep up with technical standards of EU and international laws
o Only EU has banned certain AI systems. Many laws include transparency. Audits and audit
results.
o Canada (Digital Charter Implementation Act of 2022) adds specific requirements in the use of
AI in public sector. Artificial Intelligence and Data Act takes a broad definition of AI in the
 private sector and identifies many activities that will be regulated. Canada (High-Impact
systems). Proposed creating a federal AI and data commissioner.
o Singapore – AI governance framework (2019), toolkit released in 2023 called AI Verify
o China – takes a rights-based approach vs risk-based approach, provide an overview of end
user rights, right to opt out of end user personalized recommendation, prohibition on price
differentiation, municipalities have passed additional AI governance requirements such as
auditing, banning.
o SA – PIPA – broad definition of PI, Brazil – risks-based approach, requires human oversight if
fundamental rights may be implicated,
o Japan – non-binding AI guidelines, created a national AI strategy including governance
principles, included contract guidelines and model clauses, they already have well developed
ML management guidelines and provide an indepth discussion of QMS guidance, also issued
guidelines governing cloud services and AI use,
o South Korea – non-binding guidance, people/ human-centric approach,
- Standards and risk management frameworks
o Identify, evaluate, assess and treat
o ISO 31000:2018 – ISO risk management guidelines
 Breaks down the process into principles, process and framework.
 Eight risk management principles:
 Inclusive, dynamic, best available information, human and cultural factors,
continuous improvement, integration, structured and comprehensive,
customized
 Six areas of focus:
 Leadership, integration, design, implementation, evaluation and
improvement
 Process:
 Identify risks
 Evaluate the probability of a risk event occurring
 Determine the severity of the effects of a risk event occurring
 Not intended to eliminate risks, but identify risks and implement mitigation
strategies and processes.

o ISO 42001: 2023 – ISO/IEC AI Management Systems

 Provides guidance for using AI responsibly and effectively, including the various
aspects of artificial intelligence and the different applications an organization may
use. It takes an integrated approach to manage AI projects, from risk assessment to
effective treatment of these risks.
 Applies to organizations of any size and industry involved in developing, providing or
using AI-based products or services.
 Process:
 Integrate the AI management system into the organization’s processes and
overall management structure.
 Consider specific issues related to AI in the design of processes, information
systems and controls, such as:
o Determining organizational objectives, involvement of interested
parties and organizational policy
o Managing risks and opportunities
o Processes for managing concerns related to the trustworthiness of
AI systems
 o Processes to manage suppliers, partners and third parties that
provide or develop AI systems for the organization
o Singapore’s Model AI Framework
o Huderia – Council of Europe’s Human Rights, Democracy and the Rule of Law Assurance
Framework for AI systems

General guidance:

 Develop impact assessment models that incorporate human rights with AI-centered approaches

 Apply a risk-based approach based on specific principles

 Formulate a methodology of impact assessments that follow the proportionality principle

 Develop a method for assessing and grading the likelihood and extent of risks associated with an AI
system

o Consider use, such as contexts and purposes, underlying technology, stage of development
and stakeholders

Principles:

1. Human dignity: Humans should be treated as moral subjects and not as objects to be algorithmically
manipulated

2. Human freedom and autonomy: Humans should be informed and empowered to act; systems should
enrich humans, not control or condition them

3. Prevention of harm: AI systems must not be permitted to adversely impact humans’ mental or physical
health, or planetary health

4. Non-discrimination: AI systems must be fair, equitable and inclusive in their beneficial impacts and in
the distribution of risks

5. Transparency and explainability: AI use and the rationale for this use must be made clear to affected
individuals

6. Data protection and the right to privacy: Informed, freely given and unambiguous consent for AI use
when personal information is involved

7. Democracy: Transparent and inclusive oversight mechanisms to ensure the safeguarding of the above
principles

8. Rule of Law: AI systems must not undermine judicial independence, due process, etc.

Process:

 Identify relevant human rights that could be adversely impacted

 Assess the impact on those rights

 Assess governance mechanisms to ensure the mitigation of risks, stakeholder involvement, effective
remedy, accountability and transparency

 Monitor and evaluate the system continuously for sufficient response to changes in context and
operation.
 o NIST AI Risk Management Framework

A guide for risk and incorporating trustworthiness considerations into AI design, development, use and
evaluation.

 Seven characteristics of trustworthy AI: Valid and reliable; safe; secure and resilient; accountable and
transparent; explainable and interpretable, privacy-enhanced; fair, with harmful bias mitigated.

 Four core functions:

o Govern: Cultivate and implement a culture of risk management. A cross-cutting function

infused throughout AI risk management enabling the other functions.

o Map: Identify use and risks related to use.

o Measure: Assess, analyze and track risks.

o Manage: Prioritize risks and act based on projected impact.

 Organizational processes and activities to assess and manage risk.

o Key steps: Test, evaluate, verify and validate.

 There is also a NIST AI RMF Playbook with suggested actions to accomplish outcomes in AI RMF core
functions.

 Allocating roles, responsibilities and authority to relevant stakeholders is essential.

o Oversight should review and hold accountable the risk management structures —declaring
risk tolerances for development or use.

o Determine and document roles and responsibilities.

o Support AI risk management and executives by ensuring appropriate authority and resources
to perform risk management. Equip the right people with the right tools.

o Delegate authority to staff involved in AI design, development, deployment, assessment and

monitoring.

o EU AI Act
o IEEE 7000-21: Standard Model Process for Addressing Ethical Concerns during System Design

Outlines how organizations can consider ethical values throughout system design.

Purpose: Enabling organizations to design systems with explicit consideration of individual and societal ethical
values, such as transparency, sustainability, privacy, fairness and accountability, as well as values typically
considered in system engineering, such as efficiency and effectiveness.

 Projects conforming to this standard balance management commitments for time and budget
constraints with long-term values of social responsiveness and accountability

 Emphasizes management and engineering communication with stakeholders

 Describes processes for traceability of ethical values in the concept of operations, ethical
requirements and ethical risk-based design
  Relevant to all sizes and types of organizations using their own life cycle models

 Defines the term "ethical" as supporting the realization of positive values or reduction of negative
values

The ethically-aligned processes described are performed during two stages in the system life cycle:

1. Concept exploration

2. Development

o ISO/ IEC Guide 51 Safety Aspects

 Intended for those who draft standards

 Provides guidance on including safety aspects and serves as a reference for other stakeholders

Goals:

 Reduce risk from product or system use

 Reduce risk in design, production, distribution, use and destruction or disposal of systems or products

 Achieve tolerable risk for people, property and the environment

 EU AI Act uses similar terminology and processes in identifying risk factors

There are other ISO/IEC guidelines addressing aspects of AI safety and security, including ISO/IEC 23894:2023.
Familiarize yourself with those that apply to your industry.

Challenges in implementing frameworks

- Change Management

Chapter 8 – Ongoing AI issues and concerns: awareness of legal issues

- ISACA’s COBIT 2019 framework

- Committee of Sponsoring Organizations (COSO) Enterprise Risk Management Framework
- Institute of Internal Auditors (IIA) Artificial Intelligence Auditing Framework
- Singapore PDPC Model AI Governance Framework
- US Govt. Accountability Office (GAO) Framework

Other Materials

Definition of AI under the EU AI Act – cumulative criteria

 Be machine-based.

 Be designed to operate with varying levels of autonomy.

 Have the ability to infer how to generate outputs from inputs received for explicit or implicit
objectives and make decisions that can influence physical or virtual environments.

 Exhibit adaptiveness after deployment.

General purpose AI (GPAI) = Foundational models = GenAI

"an AI model, including where such an AI model is trained with a large amount of data using self-
supervision at scale, that displays significant generality and is capable of competently performing
a wide range of distinct tasks regardless of the way the model is placed on the market and that
can be integrated into a variety of downstream systems or applications, except AI models that are
used for research, development or prototyping activities before they are released on the market."