Easy peasy revision notes

What is AI in simple words?

1. AI is a branch of computer science focused on creating

technology capable of performing tasks that typically require
human intelligence. These tasks include learning, reasoning,
problem-solving, decision-making, and even understanding language.

2. how machines are perceived as intelligent. According to the

famous Turing Test, proposed by Alan Turing, a machine can be
considered intelligent if it can produce responses during a conversation
that convince a human interviewer that the responses come from
another human.

3. technology dedicated to developing systems and machines

capable of performing tasks that typically require human
intelligence. These tasks include learning, reasoning, problem-
solving, understanding language, and perceiving the environment.

Key elements of AI

1. Technology refers to the underlying systems and tools that drive

AI.

2. Autonomy is another defining feature of AI, highlighting its ability to

operate independently to achieve its predefined goals.

3. Human Involvement is essential in the development and

application of AI. Humans play a crucial role in training AI
systems by providing data, defining the objectives the system
is meant to achieve, and fine-tuning its algorithms to align
with intended outcomes.

4. Output represents the tangible results generated by AI

technology. These outputs can take many forms, including solving
problems, automating tasks, producing content, or providing
recommendations.

Brief History of AI

The Early Days: First AI Summer (Mid-1950s to Mid-1970s)

AI began with great promise, as groundbreaking ideas and tools emerged.
The Dartmouth Conference in 1956 was a defining moment, establishing AI
as a distinct academic field. During this time, AI research labs were set
up at top universities, and pivotal advancements like the first AI
programming language, LISP, developed by John McCarthy, laid the
foundation for future exploration. Notable achievements included ELIZA, an
early natural language processing program developed at MIT,
demonstrating AI’s potential to interact with humans.

First AI Winter (Mid-1970s to Mid-1980s)

This optimism gave way to the first AI winter, as the initial excitement
dwindled. Skepticism grew over AI's capabilities, leading to reduced
funding and significant critiques of its limitations. Many ambitious
goals set during the first summer proved overly optimistic, and progress
slowed dramatically.

Second AI Summer (Mid-1980s to Late 1980s)

Interest in AI resurged during the second AI summer, driven by the rise of

expert systems that emulated human decision-making processes in
specialized fields. These systems demonstrated the practical potential of
AI in industries such as healthcare and engineering. The Japanese
government’s Fifth Generation Computer Systems project sought to
create AI-powered computing, bringing renewed energy and investment into
the field.

Second AI Winter (Late 1980s to Late 1990s)

Despite the renewed interest, this progress was not sustainable. The second
AI winter set in, marked by waning enthusiasm and significant funding cuts.
The high cost of maintaining expert systems and the geopolitical
shift following the end of the Cold War contributed to this downturn.
Once again, AI faced challenges in meeting the high expectations
placed upon it.

Renaissance and the Era of Big Data (Late 1990s to 2011)

The late 1990s marked a turning point for AI, as advances in technology
began to overcome earlier barriers. The era of big data transformed AI,
providing the vast datasets needed to train more effective models.
IBM’s Deep Blue made history in 1997 by defeating world chess
champion Garry Kasparov, signaling a new era of machine capability. The
emergence of the internet, combined with advancements in
computational power and machine learning techniques, enabled AI to
solve increasingly complex problems. Practical applications like
recommendation systems for shopping and voice assistants in
smartphones brought AI into everyday life.

AI Boom (2011 - Present)

Since 2011, AI has experienced an unprecedented surge in development,

driven by advances in deep learning, a subfield of AI focused on training
neural networks with vast amounts of data. This era has been marked
by achievements that were once unimaginable. In 2016, Google’s AlphaGo
defeated the world champion of Go, a game known for its complexity,
showcasing AI’s strategic capabilities. More recently, OpenAI’s GPT-3
demonstrated the immense potential of language models to process and
generate human-like text, reshaping industries like content creation and
customer service.

Tasks AI can do
Bias in AI

Bias in AI systems is a critical challenge that affects fairness, civil liberties,

safety, and economic opportunities. It arises from societal inequalities and
flaws in data, influencing AI systems at all stages of their lifecycle. To
address bias effectively, it is essential to understand its various forms and
implement strategies to mitigate them.

1. Implicit bias occurs when discrimination is unintentionally embedded

in AI models, often reflecting unconscious biases of developers or
decision-makers. For example, a hiring AI system trained on historically
biased resumes may favor male applicants.

2. Sampling bias happens when training data disproportionately

represents a subset of a group, leading the model to favor that subset.
A facial recognition system trained on predominantly light-skinned
individuals may perform poorly on darker-skinned faces, resulting in
inaccuracies.

3. Temporal bias emerges when models fail to adapt to changes over

time. For instance, a recommendation system trained on 2020 data
might struggle with 2024 trends. Projection bias arises when
developers design systems based on their preferences or assumptions,
disregarding diverse user needs. This can result in applications that
cater to only a limited demographic.

4. Overfitting occurs when a model performs exceptionally well on its

training data but poorly on new data, as it becomes too tailored to its
training set. For example, an AI predicting housing prices may excel in
its training market but fail in different regions. Edge cases and outliers
represent unusual or extreme data points that models often struggle to
process. For example, a vehicle detection system might fail to
recognize an unusually shaped car.

5. Noise in data, such as random or erroneous information, negatively

impacts model accuracy. For example, mislabeled data in a speech
recognition dataset can degrade its performance. Outliers, which
deviate significantly from the majority of data, can also skew
predictions. For instance, extremely high or low salaries in a dataset
can distort salary prediction models.

Mitigating bias requires:

  diverse data collection to ensure representation across demographics,

 regular retraining to adapt to new trends

 rigorous testing to identify and address biases during development.

 Handling edge cases and reducing noise in datasets are also crucial for
improving model reliability.

Types of AI

1. Classification 1 : Upon capability

1. Strong/Broad AI (General AI):

Also known as General AI, this is the type of AI capable of understanding,

learning, and applying its intelligence to a wide range of problems,
much like a human being.

Able: to perform any intellectual task that a human can, with

reasoning and understanding that extends across different domains.

However, Strong AI is hypothetical and does not currently exist. 2.

Weak/Narrow AI:

This is the AI we interact with daily. Weak or Narrow AI is designed to

perform specific tasks and operates within a predefined range or
scope. It lacks general intelligence or consciousness, focusing only on
particular applications.

Examples include:

1. Chatbots like those on customer service websites

2. Recommendation systems used by Netflix or Amazon

3. Image recognition software in smartphones

4. Voice assistants like Siri or Alexa

3. Artificial Superintelligence (ASI):

ASI represents a theoretical future where AI surpasses human

intelligence in all fields. This would include self-awareness,
understanding emotions, and experiencing reality similarly to humans.
Like Strong AI, ASI does not yet exist and remains a subject of speculation
and ethical concern.

2. Classification 2: based on how systems are trained

The three primary training methods are Supervised Learning,

Unsupervised Learning, and Reinforcement Learning.

1. Supervised Learning. Supervised learning involves training models on

labeled datasets where each input is paired with a known output. The system
identifies patterns and relationships within the data to make predictions for
unseen inputs. With exposure to more data, the model refines its accuracy
and minimizes errors.

o Categories of supervised learning:

 Classification models. Categorize data into specific

labels (e.g., determining if an image contains a cat or a
dog).

 Regression models. Predict continuous values (e.g.,

forecasting stock prices or estimating house values).

o Examples:

 Image recognition.

 Medical diagnostics.

 Natural language processing.

Supervised learning is highly effective for structured tasks, though it requires

considerable effort to curate accurate labeled datasets.

2. Unsupervised learning

Unsupervised learning operates without labeled datasets, allowing models to

discover patterns, similarities, and relationships independently. While this
approach reduces the need for manual labeling, it can produce less accurate
or unpredictable outcomes due to the absence of guidance.

o Common techniques:
  Clustering. Groups data points with similar features (e.g.,
identifying genetic similarities in DNA samples or
categorizing customers by purchasing behavior).

 Association rule learning. Finds relationships between

variables (e.g., identifying products frequently bought
together in a supermarket).

o Examples: Detecting anomalies in machinery for predictive

maintenance or fraud detection.

 Segmenting consumers into target groups for marketing.

 Analyzing genetic data to uncover medical insights.

Unsupervised learning is ideal for exploring unknown data structures and

uncovering hidden insights, especially in fields like bioinformatics and market
research.

3. Reinforcement learning

o Reinforcement learning trains models using a reward-and-

punishment system, enabling them to learn through trial and
error by interacting with their environment. Feedback in the form
of rewards for correct actions or penalties for mistakes helps the
model optimize its decisions over time.

Examples:

 A robot navigating a maze learns to avoid obstacles and

find the shortest path through rewards.

 Generative text models improve their responses based on

feedback from interactions.

 Optimizing online ad placement in real-time bidding

environments by rewarding strategies that maximize
conversions.

 Developing autonomous vehicles.

 Creating video game agents capable of competing against

human players.
Reinforcement learning is particularly effective for tasks requiring continuous
adaptation and optimization in dynamic environments.

TECH SIDE: AI Technology stack

The AI technology stack is like the foundation of a building—it

includes all the tools, hardware, and systems needed to make
artificial intelligence work effectively. Think of it as the layers that
support AI, starting with the basics like computers and storage, and
building up to the apps we interact with every day, such as chatbots or
recommendation systems. Each layer plays a specific role, and
together, they allow AI to process data, learn, and solve problems.

elements to know what they mean:

1. compute infrastructure. This is the "brainpower" of AI, provided

by hardware like powerful computers and special chips (like
GPUs) that handle heavy tasks. It also includes cloud computing,
where code can run on multiple devices at once, making AI fast
and scalable. GPUs can handle this by performing thousands of
calculations at the same time, which allows AI models to train
faster and more efficiently. In summary, GPUs are specialized
tools that speed up AI by doing the repetitive, data-heavy
tasks much faster than CPUs, enabling breakthroughs in
everything from gaming and healthcare to autonomous
vehicles and virtual assistants.

2. Storage &Network .

3. AI [Link] are the tools developers use to create, test, and

update AI systems. Popular platforms like Google Cloud or Microsoft
Azure make it easier to manage data, train AI models, and deploy them
into the real world.

What are the benefits of platforms?

1. Automate repetitive tasks in system development

2. Monitor and manage AI models in production

3. Accelerate the deployment of applications by integrating

tools for data preparation, model training, and
deployment
 4. AI Applications

This is the "visible" layer—how we experience AI in our daily lives. It has

become a powerful tool in solving real-world problems, reshaping industries,
and enhancing our daily lives.

Useful terms:

1. Software democratization is a key aspect of compute infrastructure

in AI, enabling users with minimal technical expertise to leverage AI
through low-code or no-code tools. This approach lowers barriers to
entry, making AI accessible across industries. For instance, a business
analyst can use AI software to predict sales trends without needing to
write complex algorithms.

2. Tuning AI systems is another essential process, involving the

adjustment of hyperparameters to optimize model performance. This
process, often requiring trial and error, is vital for generating accurate
outcomes. For example, tuning a model to predict customer
preferences may involve balancing computational efficiency with
accuracy.

3. Data transformation is equally critical, as raw data must often be

reformatted or enhanced to ensure compatibility with AI models. High-
quality transformation improves reliability, while poor transformation
can lead to faulty results. For instance, in healthcare, transforming
patient records into standardized formats allows models to predict
treatment outcomes more effectively.

4. Labeling plays a significant role in training and deploying AI models.

Labels act as annotations that help models understand what they are
learning. For example, in an image recognition task, labels indicate
whether an image contains a cat or a dog. However, challenges such
as scaling labeling operations and maintaining quality control can
affect model performance. Inconsistent or incorrect labels may lead to
unreliable results.

5. Observability and monitoring are crucial for ensuring AI models

perform as intended throughout their lifecycle. Observability includes
tracking the health of the data ecosystem and monitoring algorithm
performance, such as input data quality and output accuracy. For
instance, observability in customer service chatbots ensures the bots
understand user queries correctly and identifies areas for
 improvement. This process also helps detect issues like data drift and
bias.

6. Data drift, a common challenge, occurs when the data applied to a

model differs from its training data, leading to inaccuracies. For
example, an AI model trained on data from one demographic may
produce unreliable results when used on another. This issue is
particularly problematic in fields like healthcare and requires
continuous monitoring and retraining of the model.

Biases

Risks
Security risks

1. Centralization of Power

AI often concentrates decision-making power in the hands of a few

organizations. Think about tech giants shaping industries, governments, and
even public opinions. What happens when this power is used for surveillance,
targeted advertising, or automated decision-making?

A chilling real-world example is China’s social credit system. It uses AI

to monitor behaviors and assign scores that determine access to jobs, loans,
or travel. Supporters say it encourages good behavior, but critics warn about
restricted freedoms and enforced conformity
2. Overreliance on AI

Here’s a common trap: People assume AI is flawless because it handles

critical tasks like cybersecurity or financial transactions. But is that trust
misplaced?

Consider cybersecurity systems that use AI to detect threats. In 2020,

hackers bypassed AI-based fraud detection in a major bank, leading to
financial losses. It’s a reminder that no system is infallible, and human
oversight is critical.

3. Adversarial attacks

Let’s talk about adversarial attacks, where attackers manipulate input

data to trick AI systems.

Imagine this: A self-driving car encounters a manipulated stop sign with a

strategically placed sticker. The AI misreads it as a speed limit sign and
accelerates instead of stopping. Scary, right? Now, think of similar attacks in
healthcare—altering medical imaging data to cause misdiagnoses, leading to
delayed treatments or wrong interventions.

4) Training data poisoning

Another significant threat is training data poisoning. Malicious actors

insert corrupted data into the training dataset, compromising the AI’s
reliability.

For example, a fraud detection system trained on tampered transaction data

could overlook real fraudulent activities. In healthcare, poisoned datasets
might cause diagnostic tools t

5) Data leakage

Data leakage happens when sensitive information is unintentionally exposed,

often due to weak security measures.

6) Filter bubbles and Echo chambers

Generative AI systems often personalize content delivery based on a user’s

past interactions, creating what are known as filter bubbles or echo
chambers. These phenomena occur when users are repeatedly exposed to
information that aligns with their existing preferences and beliefs, limiting
their exposure to diverse perspectives.
Filter Bubble

Filter Bubble is caused by algorithms that personalize content. A

filter bubble happens when AI systems, like those on social media or
content platforms, show you information based only on your past
preferences and behavior. It happens when the system limits what you
see based on your past preferences.

An echo chamber is resulting environmentwhen you’re surrounded

by ideas and opinions that match your own because the AI system
keeps reinforcing what you already believe.

7) Deepfakes

Deepfakes are AI-generated media that convincingly alter reality, often with
harmful consequences.

o Example: A deepfake video of President Obama delivering a

fabricated speech once went viral. While it was a demonstration,
the technology has been weaponized in scams and
disinformation campaigns.

Operational
 o Hardware сostsю Training advanced AI models requires
specialized processors like GPUs. For instance, GPUs can cut
the time for training a deep learning model from weeks to days
compared to CPUs. But here’s the catch—they’re pricey!

Example: Training a state-of-the-art image recognition model on CPUs might

cost $10,000 in time and energy, while using GPUs could cost $50,000
upfront but save weeks of work.

o Storage Challenges: AI thrives on data—lots of it! But storing

petabytes of data, especially video data from surveillance
systems or IoT devices, is a costly and logistical nightmare. Think
about managing surveillance footage for an entire smart city—
petabytes of data that need high-speed networking just to
keep up.

o skilled professionals. These experts are in short supply and

high demand, commanding six-figure salaries in many cases. For
instance, developing an AI tool to predict equipment failures in
manufacturing would require data scientists to:

 Curate massive amounts of training data.

 Fine-tune models for accuracy.

 Ensure the system integrates seamlessly with existing

operations.

o c)Environmental impact. systems have a substantial

environmental footprint. Training large models, such as GPT-3,
consumes vast amounts of electricity, generating carbon
emissions equivalent to several cars over their lifetimes.
This raises sustainability concerns. To counteract this,
organizations may adopt green technologies, like renewable
energy or energy-efficient hardware. However, these initiatives
require significant upfront investment, such as building solar-
powered data centers to offset the environmental cost of running
AI systems.

d)Data Integrity Risks. Data integrity is critical to the success of AI

models. Compromised data can lead to catastrophic outcomes. Data
corruption happens when data is inadvertently altered, leading to
inaccuracies.

Privacy

-Data persistence happens when information outlives its purpose—or even

the person it’s about! Often, data is kept indefinitely, creating risks without
clear justification.

Example: A social media platform retaining someone’s photos or posts long

after their death. While families might value digital memories, retaining this
data without consent can lead to privacy breaches or legal issues.

Data repurposing occurs when information is used beyond its original

intent—sometimes in ways users never expected.

Example: A fitness app collects health data to recommend workouts but later
uses the same data to sell targeted ads. This violates purpose limitation, a
core principle of privacy laws like GDPR.

Spillover data refers to collecting information about individuals who

weren’t the intended targets of data collection. Example: A surveillance
camera records passersby unrelated to its purpose. Worse, AI models might
infer sensitive details about these individuals based on correlations in data.
 6. Informed [Link] individuals know what data is being
collected and how it will be used is critical.

2. Transparency. Organizations must clearly explain:

o What data is collected.

o How it’s used.

o What derived data might be created.

o How individuals can access, correct, or delete their data.

Business

Ethical -Bias in AI systems arises from poor-quality training data, inadequate

labeling, or flawed algorithms. This can lead to discriminatory outcomes,
damaging trust and exposing businesses to lawsuits.

Job displacement. AI automates tasks, improving efficiency but potentially

displacing workers, which can impact morale and workforce stability.

Dependence on vendors. Relying on external vendors for AI solutions can

lead to vendor lock-in and data security risks, especially if vendors go
bankrupt or are acquired.
Unclear accountability. Many AI systems operate as "black boxes," making it
hard to trace accountability for their decisions.

Intellectual property. AI systems trained on unlicensed data can infringe on

intellectual property rights, risking legal and reputational damage.

Trustworthy AI

11. Human-centric. Trustworthy AI prioritizes the needs, values,

and rights of humans above all else, ensuring that
technology serves as a tool for human empowerment and
welfare.

o Enhancing human well-being:

 AI systems should be designed to augment human

capabilities, solve real-world problems, and improve quality
of life.

 They must not replace human judgment in areas where

empathy, ethics, and contextual understanding are critical.

o Safety and user-friendly design:

 Trustworthy AI systems must be safe to use, with minimal

risks to users and society.

 They should be intuitive, accessible, and inclusive, catering

to diverse user groups, including those with disabilities.

o Avoiding harm:

 Human-centric AI ensures that systems do not cause

physical, emotional, or psychological harm to individuals or
communities.

 Developers must assess potential risks and impacts on

vulnerable populations and take measures to mitigate
them.

2. Transparent. Transparency is a critical feature of Trustworthy AI,

enabling stakeholders to understand, trust, and interact
with AI systems confidently.

o Explainability:
  AI systems must provide clear and understandable
explanations for their decisions and actions.

 Users should be able to comprehend why an AI system

made a specific recommendation or decision, especially in
high-stakes scenarios like healthcare or legal judgments.

o Disclosures:

 Organizations should disclose the presence and purpose of

AI systems, particularly when interacting with the public.

 For example, chatbot users should be informed that they

are engaging with an AI rather than a human.

 Traceability:

 AI development and deployment processes must be

traceable, with clear documentation of how systems
were built, trained, and deployed.

 Traceability supports auditing, debugging, and

improving AI systems while ensuring accountability.

3. Accountable. Accountability is a cornerstone of Trustworthy AI, ensuring

that there is clarity and responsibility for the actions and decisions
of AI systems.

a)Clear Ownership. Developers, deployers, and users of AI systems

must have clearly defined roles and responsibilities..Organizations must
establish accountability frameworks to address potential issues, such as
errors, misuse, or harm caused by AI systems.

b) Auditable Systems. Trustworthy AI systems must be auditable, with

mechanisms in place to trace and evaluate their decision-making processes.
This includes maintaining logs, version controls, and records of model
training and deployment activities.

c) Redress Mechanisms. Organizations must provide channels for

individuals or groups to challenge AI-driven decisions or outcomes.
Having a human in the loop during critical decision-making processes
ensures that there is always someone accountable for the final outcome.

HOW TO REACH TRUSTWORTHY AI

[Link] foundations.

o embed trustworthiness into the operating model: Integrate

responsible AI practices as a fundamental part of the
organization’s workflows.

o practice responsible AI processes. Develop AI systems with

fairness, transparency, and explainability in mind.

o operationalize risk management frameworks. Use structured

frameworks to address potential risks, including privacy and
security concerns.

o design for scalability and resilience. Ensure AI systems can

handle increased users, data, or operational scale without
compromising performance or safety.

o guarantee safety and security. Confirm that AI systems are

robust against cyber threats and operational vulnerabilities.

o ensure integrity and explainability. Build systems that are

transparent and understandable, allowing stakeholders to trust
their outputs and decisions.

o uphold fairness and non-discrimination. Implement

measures to prevent bias in AI decision-making and promote
equitable outcomes.

o promote human oversight and values. Design AI to

complement human judgment, ensuring human values and
ethics remain central to its operations.

2. Managerial effort

o roles are assigned & organizational structure is updated

o team has pro-innovation mindset.

o team is focused on result that will change the product/industry.

o no biases towards some industry/ law

o risk management is implemented

o playbooks &policies &tech standards are created + policies for 3r

party risks

o team follows the documentation

 o team works taking into account all stakeholders

o team is trained

[Link] effort.

o the culture of trustworthy AI & accountable AI is present

o leaders champion for responsible AI

o leaders control implementation of documentation to life practices

Key Frameworks for AI and its risk management

ISO 31000:2018 Risk Management – Guidelines

comprehensive approach to risk management across various

domains, including AI.
8 risk management principles:

o Inclusive,

o dynamic,

o best available information,

o human and cultural factors,

o continuous improvement, integration,

o structured and comprehensive,

o customized

Six principles to guide leadership and commitment

The IEEE 7000-21 Standard Model Process for Addressing Ethical
Concerns during System Design

Goal: to ensure that ethical considerations, such as transparency,

sustainability, privacy, fairness, and accountability, are addressed alongside
traditional system engineering values like efficiency and effectiveness.

applicable to organizations of any size and operating across various

sectors

Ethically-aligned processes are performed during 2 critical stages of the

system life cycle: concept exploration and development.

o In the concept exploration phase, ethical risks and values

are identified and incorporated into the design vision.

o During the development phase, these values are

operationalized and implemented into the system’s
architecture, ensuring alignment with the defined ethical
goals.

The Asilomar AI Principles

a significant set of guidelines aimed at assuring ethical and

responsible use of artificial intelligence and related technologies.
Council of Europe’s Human Rights, Democracy, and Rule of Law
Assurance Framework for AI Systems (HUDERIA)

focuses on aligning AI technologies with fundamental rights and

democratic values. It provides a robust foundation for evaluating AI
systems' compliance with human rights and ethical considerations,
particularly in governance and public decision-making.

Principles and priorities to address specific AI-related challenges:

• Human dignity

• Human freedom and autonomy

• Prevention of harm

• Non-discrimination

• Transparency and explainability

• Data protection and the right to privacy

• Democracy

• Rule of law
The steps of assessment:

• 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

The OECD Principles on Artificial Intelligence

offer an international perspective, emphasizing inclusive growth,

sustainability, and human-centric AI.
National Institute of Standards and Technology’s AI Risk
Management Framework (NIST AI RMF)

provides practical tools for assessing, managing, and mitigating AI-

related risks. Its iterative approach makes it suitable for organizations
seeking to adapt their risk management processes over time as their AI
systems and environments evolve.

Trustworthy AI characteristics:

46. Fair,

47. with harmful bias mitigated;

48. privacy-enhanced;

49. accountable and transparent;

50. safe;

51. valid and reliable;

52. secure and resilient;

53. explainable and interpretable.

The framework separates “governing of AI risks” and the actual

management :

54. Map. Identify use and risks related to use

55. Measure. Assess, analyze and track risks

56. Manage. Prioritize risks and act based on projected impact

key steps that ensure AI systems are robust, trustworthy, and aligned
with organizational objectives.

57. Test

58. evaluate

59. verify

60. validate
ISO GUIDE 51

Safety aspects — Guidelines for their inclusion in standards

Basically guidance for guidance creators (regulators).

This Guide provides practical guidance to drafters of standards to assist

them in including safety aspects in standards. The underlying principles of
this Guide can also be used wherever safety aspects require consideration,
and as a useful reference for other stakeholders such as designers,
manufacturers, service providers, policy makers and regulators.

Elements of risks:

a) the severity of harm that can result from the considered hazard;

b) the probability of occurrence of that harm, which is a function of:

— the exposure to the hazard;

— the occurrence of a hazardous event;

— the possibilities of avoiding or limiting the harm.

Types of safety:

61. basic safety standard, comprising fundamental concepts,

principles and requirements with regard to general safety
aspects applicable to a wide range of products and systems;

62. group safety standard, comprising safety aspects applicable to

several products or systems, or a family of similar products or
systems, dealt with by more than one committee, making
reference, as far as possible, to basic safety standards;

63. product safety standard, comprising safety aspects for a

specific product or system, or a family of products or systems,
within the scope of a single committee, making reference, as far
as possible, to basic safety standards and group safety
standards;

64. —standards containing safety aspects, but which do not deal

exclusively with safety aspects, making reference as far as
possible to basic safety standards and group safety standards.
How to assess risks?

AI risk assessment is a critical process in the governance and management

of artificial intelligence systems. It scores risk

AI Risk = Severity of the harm x Probability of the occurrence

Severity of Harm- the potential impact of a risk event, including financial,

reputational, and societal consequences.

Probability of Occurrence- the likelihood that the risk event will occur based
on historical data, system behavior, and contextual factors.

ISO/IEC 22989:2022Information technology — Artificial intelligence

— Artificial intelligence concepts and terminology

ISO/IEC 42001:2023Information technology — Artificial intelligence

— Management system
Resources you HAVE to read before AIGP:
 65. NIST AI RMF

66. ISO 22989

67. ISO 42000

What is AI Governance?

AI Governance refers to the structured framework of processes,

policies, and tools designed to ensure that artificial intelligence (AI)
systems are developed, deployed, and managed responsibly.

By bringing together stakeholders from diverse fields such as data

science, engineering, compliance, legal, and business operations, AI
Governance aims to maximize the benefits of AI while minimizing
potential risks and harm.

It helps organizations align their AI initiatives with ethical principles,

legal requirements, and business goals across the entire machine
learning (ML) lifecycle.

To be effective, AI Governance must address key considerations:

[Link] management and quality. The foundation of any AI system lies in

its data. It is critical to understand where this data originates, how it was
collected, and whether its use is lawful. Data may come from first-party
sources (collected directly by the organization), public sources (such as
government records), or third-party providers (like data brokers). Ensuring
transparency and auditability in data collection and processing is vital, as is
maintaining high data quality to produce reliable AI outputs.
2. Legal and ethical compliance. AI systems often intersect with privacy
laws, data protection regulations, and intellectual property rights.
Governance frameworks must proactively address potential conflicts
between AI's data dependency and privacy obligations. For instance,
mechanisms like data inventories can identify sensitive, personal, or
copyrighted data, enabling organizations to evaluate its use responsibly.

[Link] and fairness. Biases in AI systems can arise from the data used, the
algorithms themselves, or human involvement in their design. Effective
governance includes mechanisms to detect and mitigate such biases to
prevent discriminatory or unfair outcomes.
[Link] and accountability. Understanding how AI systems make
decisions is essential for building trust. Transparent systems allow
stakeholders to comprehend the logic behind AI outputs, while accountability
ensures that clear lines of responsibility exist for AI-related actions and
outcomes.

5. Security and safety. AI systems must be secure against threats such as

data breaches or model manipulation. Additionally, they should be safe for
employees and end-users, with rigorous testing to ensure reliability.

The key takeaway is clear: AI Governance is not just a technical

necessity—it is a leadership imperative. Organizations that embed
governance into their culture and processes can harness AI’s transformative
power while safeguarding ethical and legal standards. The time to act is now,
ensuring AI becomes a tool for positive impact rather than a source of
unchecked risk.

WHICH APPROACH TO CHOOSE:

When considering AI governance within an organization, the structure of the

governance model is deeply influenced by the organization's existing
structure and culture. To establish effective AI governance from the outset, it
is essential to engage leadership and existing governance teams early. Their
involvement ensures alignment with the organization’s broader strategic
goals, leverages existing governance frameworks, and fosters support across
all levels of the organization.

1. highly centralized governance model

AI governance models can take several forms, each tailored to different

organizational needs and structures. In some cases, a highly centralized
governance model may be most appropriate. This approach assigns
responsibility for all AI-related affairs to a single team or individual.
All decisions, processes, and oversight flow through this central
authority, ensuring consistency and clear accountability. This model is
particularly effective in organizations that prioritize stringent
control over AI activities or operate in highly regulated
environments where uniformity is critical.
 2. decentralized governance model

Alternatively, a decentralized governance model may be the most

effective for organizations with a highly technical focus or where
decision-making is integrated into development processes. Often
referred to as "local governance," this model delegates decision-making
authority to lower levels within the organization. It reduces hierarchical
tiers, enabling a broader span of control and fostering a bottom-to-
top flow of ideas and decisions. This structure is well-suited for
organizations that emphasize agility, innovation, and empowering technical
teams to make informed decisions without excessive oversight.

[Link] governance model

In other scenarios, a hybrid governance model may be more suitable. This

structure combines elements of centralized and decentralized
governance. Typically, a central authority sets overarching policies
and guidelines, while local entities or teams take on the role of
implementing and supporting these directives. This model is often seen
in large organizations where central oversight ensures alignment with
organizational values and compliance requirements, while the flexibility at
the local level allows for context-specific adaptations and innovations.

Ultimately, the choice of governance model depends on the

organization’s size, culture, strategic goals, and the complexity of its AI
applications. Each model has distinct advantages and challenges, but
engaging leadership early and fostering collaboration across governance
teams ensures a solid foundation for the chosen approach.

Tailoring AI governance to the entity

Effective AI governance is not a one-size-fits-all approach.

The process begins with a deep understanding of the organization’s

specific characteristics, including its operations, incentive structures,
and the sector it belongs to—be it technology, finance, healthcare, or
another field. These factors significantly influence the risks,
opportunities, and regulatory considerations associated with AI use.

68. Understanding your role in the AI ecosystem: developer,

deployer, user?

69. What your business does?

 70. Legal & sector-specific considerations

ROLES in AI ecosytem

+GUIDE "Steps of AI governance" crafted for you