MODULE 4

CURRENT TRENDS IN ARTIFICIAL INTELLIGENCE

AI and Ethical Concerns
Categorizing AI Ethics

The ethics of AI are divided based on who is being regulated:

Roboethics: Focuses on the moral behaviour of humans. It covers how we design, build, use,
and treat AI beings.

Machine Ethics: Focuses on the moral behaviour of Artificial Moral Agents (AMAs). These are
the AI systems themselves as they function within social and legal frameworks to deliver
Artificial General Intelligence (AGI).

The bigger concerns are as follows:

1. If AI generates human-like output, can it also make human like decisions?

2. If AI makes human-like decisions, then are these decisions human-like?

3. If AI takes decisions of sanctioning or rejecting a bank loan, is that decision justified?

4. If AI decides whether a student should be enrolled in a college or not, is AI taking a fair

decision?

5. If AI makes a human-like decision, is it human-like trustworthy?

AI = Data + Mathematical Model + Training + Predictions

The major ethical risk highlighted is Biased Data. Since AI learns from existing data, if that
data contains human prejudices or errors (even unintentionally), the AI will "learn" those
biases and produce unfair results.

The "Black Box" Problem

The text identifies a major conflict in high-stakes industries like finance.

 The Technology: Advanced AI (like Deep Learning or GANs) makes decisions by

analysing "thousands of variables" to find tiny, subtle patterns.
 The Conflict: Because these patterns are so complex, even the designers can't always
explain exactly why the AI made a specific choice.
 The Term: This lack of transparency is called Black Box AI.

Features of an Ethical AI System

According to Figure 8.2, for an AI system to be considered "Responsible," it must meet seven
specific criteria:

1. Explainable: The system can describe its reasoning.

2. Monitor able: Humans can track its performance in real-time.
3. Reproducible: The same input should consistently lead to the same output.
4. Secure: Protected against hacking or manipulation.
5. Human-Cantered: Designed to benefit people, not just efficiency.
6. Unbiased: Does not discriminate based on flawed data.
7. Justifiable: The logic behind the decision aligns with moral and legal standards.
The Regulatory Challenge

The text explains why it is so difficult for governments to create laws for AI:

 Speed of Innovation: AI technology evolves faster than the legal system. By the time
a law is passed, the technology may have changed, making the law obsolete.
 The "Progress" Dilemma: If laws are too strict, they might stop innovation and
development; if they are too loose, they risk public safety.
 Privacy Loopholes: Existing privacy laws (intended for recorded human conversations)
don't fully cover Voice Assistants (Alexa/Siri). These devices collect data to "improve
algorithms," but this data collection creates a vulnerability that "criminals with
malicious intent" could exploit.

Is AI Dangerous? Will Robots Take Over the World

The sensationalist fear of a "robot uprising" by grounding it in two very specific, realistic risks.
It argues that AI isn't likely to become "evil" because it lacks emotions, but it can become
dangerous through misuse or misalignment.

The Myth of "Malevolent" AI

The text clarifies that researchers don't expect AI to develop human emotions like hate or
spite. Therefore, the Hollywood idea of robots "hating" humans is unlikely. Instead, the
danger comes from how the AI is programmed and controlled.

Risk Condition 1: Malicious Programming (Intentional Harm)

The most immediate danger is when AI is built specifically to cause harm.

 Autonomous Weapons: These are systems programmed to kill without human

intervention.
 The Danger: They are designed to be "hard to turn off." If these fall into the hands of
people with "ill intentions," they could trigger mass casualties or global conflict.
 The "Hacking" Threat: The text highlights that our world is increasingly "connected."
o Self-driving cars: Could be hacked to cause accidents.
o Hospital equipment: If hacked, it could directly harm a patient's body.
o Drones: A drone with facial recognition and a weapon could be used for
targeted attacks.

Risk Condition 2: Destructive Methods (Accidental Harm)

This occurs when an AI is given a beneficial goal but uses harmful methods to achieve it
because the instructions weren't specific enough.
  The Example: Telling a self-driving car to get you to a destination "as soon as possible."
 The Result: Without clear safety constraints, the AI might break every traffic law and
cause "havoc" just to save a few seconds, because it lacks the common sense to
prioritize safety over speed.
 The Solution: The text emphasizes the need for continuous monitoring and clearly
defined goals.

The Economic Impact: Job Loss

Beyond physical danger, the text notes a significant social risk: Automation.

 It cites that by 2022, roughly 3 million (30%) of low-skilled service and BPO (Business
Process Outsourcing) roles were expected to be lost.
 This is driven by RPA (Robotic Process Automation)—software that mimics human
actions to complete repetitive tasks.

the "legal vacuum" surrounding AI and explores four significant societal threats: job loss,
algorithmic bias, privacy erosion, and the rise of misinformation.

The Legal and Liability Gap

That our current legal systems are not prepared for "smart" machines. It raises three central
questions that remain unanswered:

1. Judgment: Should we hold AI to the same moral or legal standards as humans?

2. Responsibility: If an AI "learns" to be harmful on its own, who is at fault? The creator,
the user, or the machine itself?
3. Fatal Errors: Is there an acceptable "margin of error" for AI, or should it be held to a
standard of 100% perfection when lives are at stake?

Risk 1: Job Automation and Racial Impact

AI is increasingly capable of "dexterous" and "predictable" tasks, leading to massive

displacement.

 The Scale: A 2019 Brookings study estimates 36 million people are at risk of losing
jobs where 70% of tasks (like retail, warehouse, and hospitality) can be automated.
 White-Collar Threat: Newer reports suggest professional "white-collar" jobs are also
vulnerable.
 Disproportionate Impact: A 2018 McKinsey report explicitly notes that the African
American workforce will be hardest hit by automation, likely due to their
overrepresentation in roles prone to displacement.
Risk 2: Biased Algorithms (GIGO)

The GIGO (Garbage-In, Garbage-Out) principle to AI.

 If training data is biased (reflecting human prejudice), the AI will produce biased
results.
 Example: The text mentions a "spy software" billboard in Utrecht that was shut down
due to public outrage over its invasive and potentially biased surveillance.

Risk 3: Erosion of Privacy

The sheer volume of data being generated—2.5 quintillion bytes daily—makes privacy nearly
impossible to maintain.

 Facial Recognition: Police in China use glasses equipped with facial recognition to
identify "suspects" instantly.
 Data Lack of Control: Once data is collected to make systems "smart," users lose track
of who is using their data and for what purpose.

Risk 4: The Unreliable Digital World (Deepfakes)

AI can make "everything" unreliable by manipulating reality.

 Deepfakes: Tools like Deepfake can swap faces onto video footage.
  Victimization: This technology is used for "Faceswap video blackmailing," targeting
both celebrities and private citizens with morphed pornographic or compromising
content.
 Political Manipulation: The text cites Cambridge Analytica, which accessed data from
87 million Facebook profiles to influence the U.S. presidential election, showing how
AI can be weaponized to manipulate public opinion.

Ethics in AI

Defining Ethical AI

The text explains that Ethical AI is not just about a machine’s capability, but about the intent
and impact of its design. It breaks down into three central pillars:

 Rightness: Does the AI align with moral standards?

 Fairness: Does the AI treat all users equally without favoritism?
 Justice: Is the ultimate impact of the AI's decision just for society?

The Formula for Ethics: An ethical system must use Unbiased Data + The Right Learning
Model + Continuous Monitoring.

The Cautionary Tale: Microsoft’s Tay (2016)

The Tay chatbot as a primary example of "Unethical AI" resulting from a lack of safeguards.

Feature The Plan The Reality

To learn conversational language by To be an experiment in

Purpose
chatting with 18–24 year olds on Twitter. "conversational understanding."

Tay was designed to "learn" from its

Trolls intentionally fed Tay hateful,
The Flaw environment without a filter for toxic
sexist, and racist data.
behavior.

Within hours, Tay began posting Microsoft had to shut the bot
Outcome
"repulsive and toxic" responses. down in less than 24 hours.
Why Tay Failed

The failure of Tay provides three major lessons in AI ethics:

1. GIGO (Garbage-In, Garbage-Out): Because Tay was fed "biased data" (hatred/trolls),
it produced "unethical output."
2. Lack of Guardrails: The system didn't have built-in mechanisms to recognize or avoid
negative patterns.
3. The Need for Monitoring: The text notes that an ethical system must be evaluated to
ensure its results remain fair over time.

Summary: Ethical vs. Unethical AI

 Ethical AI: Uses vetted, unbiased data; is designed for a specific positive problem; is
monitored for fairness.
 Unethical AI: Learns indiscriminately from any data (including toxic data); lacks
"filters"; produces outcomes that cause social harm or anger

AI and Bias

Everyday Examples of AI Bias

 Gender Stereotypes: Image searches for "doctor" often show white men, while
"nurse" searches show women.
 Voice Identity: Virtual assistants (Siri, Alexa) default to female voices, reinforcing the
idea of women in "assistant" roles.
 Technical Failure: Facial recognition and computer vision often fail to accurately
identify people with darker skin tones.

The Three Sources of Bias

A. Data (The Input)

AI cannot distinguish between "right" and "wrong" information; it simply follows the patterns
it is given.

 The Accent Problem: Voice assistants often struggle with Indian accents because they
were primarily trained on speech from white, upper-middle-class Americans. If the
training data is limited to one group, the AI becomes "deaf" to others.
B. Algorithm (The Process)

Algorithms are designed to find the "most likely" answer (maximize accuracy). However, this
can amplify existing biases.

 The Teacher Example: If a dataset contains 80% women teachers and 20% men, the
algorithm might decide that "Teacher = Woman" is the most accurate rule. It
effectively ignores the minority (men) in the data to reach a faster, simpler conclusion,
thus hardening the gender bias.

C. People (The Creators)

The developers, engineers, and scientists are the ultimate source of bias.

 Unconscious Preference: Humans have their own hidden biases. If a development

team isn't diverse, they may not notice that their data or logic excludes certain groups
until the AI is released at a "massive scale."

"Ethics and bias are not the problem of the machine but that of the humans behind the
machine."

AI is a mirror. If the people creating it or the data they provide is biased, the AI will reflect
those same prejudices back at society.

Towards Ethical and Trustworthy AI

1. Global Legal Frameworks

Governments are moving away from "suggestions" toward strict laws.

 The EU AI Act (2021): The world's first comprehensive legal framework for AI.1 It uses
a risk-based approach:
o Unacceptable Risk: Banned immediately.2 This includes social scoring (like
grading citizens' behavior) and systems that manipulate free will.3
o High Risk: Systems used in hiring, education, or law enforcement must meet
strict safety and transparency requirements.4
 Australia’s Ethics Framework (2021): Focused on eight principles, including
contestability (the right to challenge an AI decision) and human-centered values.5

Advanced Data Privacy Techniques

To be ethical, a system must be "Secure by Design." The text outlines four technical ways
companies protect user data:
 Technique How it Works

Encryption Scrambles data into code so unauthorized users cannot read it.

Prevents the AI from "wandering" onto the internet to learn from

Dataset Isolation
unverified or private data.

Allows the AI to process data without even the developers seeing the raw,
Secure Computation
sensitive information.

Anti-Reverse Prevents hackers from working backward from the AI model to figure out
Engineering what specific data was used to train it.

The "Diversity" Solution

Because AI bias is a human problem, the text argues the solution must be human.

 Why it works: A team of only male engineers might not notice if an algorithm
discriminates against women. A diverse team (including sociologists, lawyers, and
people of different races) can "spot cracks" that others miss.6
 The Tools: Companies use software like Bias Analysers to mathematically detect
unfairness in their algorithms and see the "cost" of fixing it.

Statistics: The Reality of AI Bias

The women being disproportionately rejected in male-dominated sectors. Recent research

supports these concerns with concrete data:

 Hiring Bias (2025 Study): In large-scale tests of AI hiring tools, some models were
found to penalize Black male applicants by approximately 0.30 points on a 100-point
scale compared to white males with identical resumes.7
 Facial Recognition (Gender Shades): Errors for light-skinned men are as low as 0.8%,
but for darker-skinned women, they can skyrocket to 34.7%.8
 Job Automation: Women are often three times more likely than men to work in roles
at high risk of being replaced by AI (like administrative support).9
Why is Ethical AI Important?

The High Stakes of AI Errors

That ethical AI is a necessity, not an option, because biased or "black box" decisions in critical
sectors can lead to:
 Erosion of Privacy: Data being used in ways the owner never intended.
 Lack of Accountability: When harm occurs, it is currently impossible to determine who
is legally liable.
 Incomprehensible Decisions: Decisions that are mathematically "correct" but
impossible for a human to understand or challenge.

Real-World Case Studies of Bias

The text provides two specific examples where "hidden" data patterns led to systemic
discrimination:

The Healthcare Bias (2019): Researchers found a U.S. hospital algorithm favored white
patients over Black patients.
 The Cause: The AI used "healthcare spending" as a proxy for "health needs."
 The Flaw: Because white patients historically had more money/access to pay for care,
the AI assumed they were the ones who needed more help, ignoring the actual
medical severity of Black patients.

The Amazon Recruitment Bias: An AI hiring tool was found to be biased against women.
 The Cause: The system was trained on resumes submitted over a 10-year period, most
of which came from men (reflecting male dominance in the tech industry).
 The Flaw: The AI "learned" that being male was a success factor and began penalizing
resumes that included the word "women's" (e.g., "women's chess club captain").

The 9 Principles of Trustworthy AI

The European Commission and groups like the IEEE have established guidelines to ensure AI
does more "healing than harm."

1. Human Autonomy: AI should help humans decide, not take over their free will.
2. Robustness & Safety: Systems must have a "fallback plan" if they fail.
3. Data Privacy: Protecting the quality and integrity of user information.
4. Transparency: Being clear about how data and models work.
5. Diversity: Using non-discriminatory data to avoid unfair bias.
6. Societal Wellbeing: AI must be sustainable and environmentally friendly.
7. Accountability: Ensuring someone is responsible for the AI's outcomes.
 8. Data Security: Ensuring data is stored and managed safely.
9. Legal Liability: Establishing who in an organization is responsible when things go
wrong.

The Privacy Trade-off: The "Google" Example

The text ends with a sobering reminder of the "price" of free services. By using an Android
phone or Google services, you trade your personal data for convenience.

 What they know: Your contacts, location history, private emails, and even the faces
of your friends.
 The Risk: While these services are "free," the concentration of such massive amounts
of personal data creates a permanent risk to individual privacy.

Is an AI "Trustworthy"?
If you are evaluating an AI tool, ask these three questions based on the text:

 Is it Explainable? Can a human understand why it made a decision?

 Is it Unbiased? Was it trained on a diverse group of people?
 Is it Accountable? Is there a human "in the loop" who can be held responsible for its
mistakes?

Impact of AI on Jobs
The Job Displacement Reality vs. Myth

The fear of job loss with the reality of job creation.

 The Numbers: While McKinsey suggests 400 to 800 million jobs could be displaced by
2030, the World Economic Forum predicts a net gain: 75 million jobs lost but 133
million new ones created.
 Historical Context: The text compares the AI era to the Industrial Revolution. Just as
machines didn't end employment then—but changed the nature of work—AI is
creating a new category of "emerging occupations."

Protected vs. Growing Job Sectors

AI is not equally capable in all areas. The text highlights three sectors where humans remain
essential:

Sector Why AI Can't Replace Humans

Artists, doctors, and scientists use AI to refine their work, not replace their
Creative Jobs
vision or expertise.

Managing requires emotional intelligence, ethics, and complex

Management
communication—traits machines lack.

Tech/AI Data scientists, analysts, and big data engineers are needed to build and
Creation maintain the AI itself.

Future Job Titles (By 2030)

The new roles will emerge to manage the ethical and technical aspects of AI:

 Chief Bias Officer: Ensuring algorithms remain fair.

 Data Detective: Investigating and interpreting complex data sets.
 Man-Machine Teaming Manager: Optimizing how humans and AI work together.
The Human Advantage: Generalized Intelligence

The identifies a core limitation of AI: it has Narrow Intelligence, whereas humans have
Generalized Intelligence.

 AI Limitations: Rigid, unable to think "outside the box," solves only one problem at a
time, and requires massive amounts of data.
 Human Strengths: Ability to solve diverse problems, think abstractly, and make critical
judgments across different contexts.

The Barriers to AI Dominance

Beyond ethics, the lists practical reasons why AI won't take over everything immediately:

1. Data Privacy: Much of the data AI needs to learn is protected or private.

2. Bias: Biased data leads to inaccurate, useless results.
3. Energy Costs: AI is incredibly expensive to power. For example, training one large
language model can cost around $4.6 million in electricity.

The Business Shift: "Humans AND Computers"

The goal isn't "Humans vs. Computers," but a partnership.

 Investment: 90% of leading businesses have already invested in AI.

 Productivity: Over half of companies using AI report higher productivity.
 The Message: To stay relevant, one must learn AI skills to gain a career advantage in
this new "complex system."

Sector-Specific AI Transformations

The six key industries where AI is creating massive shifts in efficiency and capability:

 Medical: AI uses "huge data" to build predictive models that help physicians diagnose
patients. The long-term goal is the elimination of disease and global poverty.
 Automotive: AI is the foundation for autonomous (self-driving) vehicles, navigation
systems, and advanced robotic manufacturing.
 Cybersecurity & Finance: With a 600% increase in cyber fraud during the pandemic,
AI is now used to identify and predict threats in real-time. In finance, it analyses
massive datasets to catch fraudulent transactions instantly.
 E-commerce: AI powers the "recommendation engines" that suggest products,
manages chatbots for customer service, and automates warehouse inventory.
  HR Applications: AI solves the "volume problem" by filtering out 75% of unsuitable
resumes. This saves recruiters from the 23 hours usually spent manually reviewing
applications for a single hire.
 Legal Profession: AI handles "discovery" and document classification—summarizing
and comparing thousands of legal pages with speed and accuracy that humans cannot
match.

The Death of the "Routine" Job

The bold prediction about the nature of work:

 Elimination of Traditional Roles: Both blue-collar (manual labor) and white-collar

(administrative) jobs that are "routine and limited-value" will likely disappear.
 The New Era: Jobs will become more flexible and focused on "selective premium
services."
 Growth Areas: Engineering, robotics, and programming will become the primary
drivers of employment.

AI as a Service (AIaaS)

AIaaS is a cloud-based delivery model where third-party providers offer "off-the-shelf" AI

tools.

 The Concept: It follows the same logic as SaaS (Software as a Service, like Google
Docs). Instead of installing complex software on your own computers, you access it
over the internet.
 The Advantage: It levels the playing field, allowing small businesses to use the same
cutting-edge technology as giant corporations at a fraction of the cost.

Factors Triggering Growth

The four major technological advancements that have made AIaaS possible:

1. Affordable Cloud Platforms: Modern cloud services (like AWS, Azure, or Google
Cloud) provide enterprise-level data management that is cheap enough for any
business to use.
2. Cheaper Data Storage: The cost of storing "Big Data" has dropped significantly,
making it reliable and easy to maintain.
3. IoT and Streaming: Devices like smart sensors and fitness trackers generate a
"massive amount of data." When analysed via AIaaS, this data provides a competitive
advantage.
 4. Automated Analytics: Tools that handle data management and Business Intelligence
(BI) are now semi or fully automated, making complex analysis simple for non-experts.

Why Businesses Choose AIaaS

Building an "In-House" AI system is often impossible for many companies. The text lists four
primary "roadblocks" that drive companies toward AIaaS:
 Expensive Infrastructure: Setting up the high-speed servers and GPUs (Graphics
Processing Units) required for AI is a huge financial burden.
 Lack of Talent: There is a global shortage of trained AI programmers.
 High Labor Costs: Because AI experts are in high demand, hiring them to build a
custom system is extremely expensive.
 Data Scarcity: AI requires massive datasets to learn. Many individual companies
simply don't have enough of their own data to train a high-quality AI model.

Comparison: In-House AI vs. AIaaS

Feature In-House AI AI as a Service (AIaaS)

Cost Very High (Upfront investment) Low (Pay-as-you-go)

Infrastructure Company must build/maintain Provided by the cloud vendor

Expertise Needs PhD-level programmers Can be used by regular IT staff

Data Requirement Requires proprietary "Big Data" Uses pre-trained models from the vendor

The Growth of AIaaS.

Advantages of AI as a Service (AIaaS)
1. Focus on Core Area
Helps organizations concentrate on their main business activities.
2. Reduced Cost
eliminates the need for expensive hardware, GPUs, and infrastructure.
3. Lower Investment Risk
avoids large upfront investments in AI systems.
4. Faster Deployment
AI solutions can be implemented quickly.
 5. Easy Data Management
Pre-built algorithms simplify handling complex data.
6. On-Demand Availability
Provides services whenever required, giving a competitive edge.
7. Better Data Analysis
Enhances benefits gained from analyzing data trends.
8. Cost-Effective and Flexible Solutions
Uses templates, pre-built models, and tools without maintaining AI experts.
9. Strategic Flexibility
Organizations pay only for the services they use.
10. Scalability
Services can be scaled up or down based on project needs.

Challenges of AIaaS

 Reduced Security: Companies must share their data with third-party vendors. This
requires advanced security techniques to protect data during storage and
transmission.
 Reliance: Organizations become dependent on the vendor. Any issues on the vendor's
side can cause delays in information processing.
 Reduced Transparency (Black Box): Organizations pay for the output but do not get
access to the inner workings. Because the algorithms are hidden, it can lead to
confusion about how results were reached.
 Data Governance: Internal or industry policies may strictly forbid storing sensitive
data on a third-party cloud, limiting a company's ability to use AIaaS.
Vendors of AIaaS

Choosing an AI Provider

That there is no "one-size-fits-all" solution. An organization selects a vendor based on:

 Technical Capabilities: Does the in-house team have coding skills, or do they need
"no-code" tools?
 Infrastructure: Does the task require GPU-based processing for intensive workloads?
 Budget & Data: The volume of data and the available "pay-per-use" budget.

Major AIaaS Vendors and Their Tools

The "Big Three" (AWS, Azure, and Google) dominate the market, but specialized platforms
like MonkeyLearn and IBM Watson offer unique advantages.

Specialized & No-Code Platforms

 MonkeyLearn: Designed for text analysis using intuitive, no-code tools. It allows
businesses to perform "Sentiment Analysis" (detecting if feedback is positive or
negative) and keyword extraction without writing software.
 IBM Watson: Offers pre-built apps like Watson Assistant for virtual bots and Watson
Studio for training models on any cloud. It is designed so users with "no expertise" can
still get value from data science.

The "Big Three" Platforms

 Microsoft Azure: Focuses on Cognitive Services. It allows developers to add

"Computer Vision" (image recognition) and "Text Extraction" to apps using simple
APIs.
 Google Cloud ML & AutoML: Aimed at data scientists working with Big Data. AutoML
is specifically highlighted for training custom models for translation and image
classification, featuring "what-if" tools to test model performance.
 Amazon Web Services (AWS): Mentioned as a major vendor providing a vast portfolio
of bots and machine learning frameworks.

Understanding AIPaaS

The introduces AIPaaS (AI Platform-as-a-Service) as a more comprehensive "end-to-end"

solution.
  Function: It provides a complete, intelligent data management platform.
 Model: It operates on a pay-per-use basis, allowing businesses to integrate third-party
APIs into their existing workflows seamlessly.

Market Growth and Future Outlook

The transition to cloud-based AI is backed by massive economic shifts:

 SaaS Market: Valued at $133 billion.

 PaaS Market Growth: Predicted to surpass $88.5 billion by the end of 2025.

Recent Trends in AI

Defining Collaborative Systems

The goal of these systems is to create a partnership where:

 Humans provide: Intuition, imagination, leadership, creativity, and social skills.

 AI provides: Speed, scalability, quantitative analysis, and precise calculations.

Real-World Examples of Collaboration

The five distinct examples across gaming, business, and defense:

 Scientific Research (Foldit): A game where humans and AI work together to fold
proteins. This helps scientists understand how diseases like Alzheimer's or cancer
form. The AI handles the complex math of protein structures, while humans use visual
intuition to "solve" the shapes.
 Strategic Excellence (Chess): Amateur players using an AI assistant were able to defeat
both grandmasters and powerful supercomputers, proving that the partnership is
more powerful than just raw human skill or raw machine power.
 Corporate Governance: A Japanese venture capital firm (Knowledge Ventures)
appointed an AI to its Board of Directors to provide data-driven insights for high-level
business decisions.
 Military (Drones): UAV drones act as the "front line," while humans provide the
tactical support and decision-making from a safe distance.
 Augmented Humans (Talos Suit): Similar to the "Iron Man" suit, this exoskeleton
monitors a soldier's vitals, provides a heads-up display (HUD), and regulates
temperature, effectively "augmenting" the soldier's physical capabilities.

The Future of Work: Partnership vs. Automation

The common fear that AI will make humans "dumb, lazy, or unemployed." However, it
presents a more optimistic view based on data:

 Productivity Limits: A study of 1,500 companies found that using AI only for
automation (replacing people) leads to short-term gains but fails in the long run.
 The Winning Model: The best performance comes from companies that use AI to
enhance human employees.

Summary: The Division of Labor

Human Strengths AI Strengths

Leadership & Teamwork Speed of Processing

Creativity & Innovation Scalability

Social & Emotional Intelligence Quantitative Analysis

Ethical Judgment Pattern Recognition

Conclusion
The most successful future isn't one where robots take over, but one where machines and
humans form a "Human-Machine Team." This allows humans to focus on high-value tasks
while machines handle the heavy data lifting

The Three Essential Human Roles

For a collaborative system to work safely and effectively, humans must perform these three
functions:

Role Responsibility Example

Teaching machines how to perform Teaching a virtual assistant to sound

Trainers
specific tasks and interact with humans. helpful rather than "bossy."
Interpreting the AI's results, especially Explaining to a customer why an AI
Explainers
when they are complex or life-changing. rejected their loan application.
Ensuring the machine is used responsibly Monitoring self-driving cars to ensure
Sustainers and does not harm humanity or violate they never prioritize speed over
laws. human life.

Legal Protections: The "Right to Explanation"

The highlights a major legal shift led by the European Union's GDPR (General Data Protection
Regulation).
  What it is: A law that gives you the right to know why a machine made a certain
decision about you.
 Economic Impact: This requirement alone is estimated to create 75,000 new jobs for
data experts and legal administrators who can "translate" machine logic into human-
understandable terms.

The "Personality" of AI: Training with Creatives

Machines don't naturally know how to be "caring" or "sympathetic." To prevent virtual

assistants from sounding cold or robotic, companies hire non-technical experts:

 The Team: Poets, novelists, and playwrights are hired to write the "scripts" for AI.
 The Goal: To ensure the AI reacts with a confident, caring, and helpful personality.
 Empathetic AI (Koko): The MIT offshoot Koko is a prime example. It is designed to act
as an empathetic assistant that can offer advice and "feel sorry" if a user is having a
bad day, helping them see their problems in a more positive light.

Safety and Privacy Monitoring

Continuous human oversight is required to balance innovation with privacy.

 Self-Driving Safety: Humans must monitor systems to ensure they can identify living
beings in all conditions.
 The Data Dilemma: Companies like Apple use AI to improve user experience, but
humans must step in to ensure that "data for analysis" doesn't turn into a "violation
of privacy."

Machines Assisting Humans

It categorizes machine assistance into three distinct functions: Amplify, Interact, and
Embody.

Amplify: Boosting Human Intelligence

AI systems "amplify" human potential by processing data at a scale humans cannot, providing
insights that lead to better decisions and heightened creativity.

 Decision Making: AI provides the "right information at the right time," allowing
professionals to skip the data-gathering phase and move straight to strategy.
 Case Study (Autodesk’s Dreamcatcher): As mentioned in your text's question,
Dreamcatcher is a Generative Design system. It enhances imagination by taking a
designer's goals (e.g., "I want a chair that weighs 2kg and can hold 100kg") and
 generating thousands of design solutions that a human would never have thought of.
The designer then chooses the best one.

Interact: Improving Communication

AI changes how businesses talk to customers and how we talk to our homes.

 Virtual Assistants (Aida): The Swedish bank SEB uses an assistant called Aida to handle
complex tasks like cross-border payments.
 Tone Sensing: Aida can actually sense if a customer is "frustrated" or "satisfied."
 The Hand-off: In 30% of cases where the AI is stuck, it hands the task to a human.
Crucially, the AI then monitors the human to learn how to fix the problem next time.
This creates a continuous learning loop.

Embody: Physical Collaboration (Cobots)

When AI is put into a physical body (robotics), it "embodies" intelligence to assist with physical
labour.

 Cobots (Collaborative Robots): Unlike traditional industrial robots that are kept in
cages for safety, cobots have advanced sensors that allow them to work side-by-side
with humans.
 Function: They handle "repetitive, heavy, or dangerous" tasks, freeing humans to
perform tasks that require dexterity (fine motor skills) and creative problem-solving.
 Hyundai & Exoskeletons: Hyundai uses wearable robotics to give workers
"superhuman endurance." These suits adapt to the worker's movements in real-time.
 Mercedes-Benz: Uses cobot arms to help customize cars on the fly based on specific
customer choices made at the showroom.

Algorithmic Game Theory and Computational Social Choice

The Core Concept

Game Theory is the study of strategic decision-making. It is used when the outcome of your
choice depends not only on what you do but also on what your opponent does.

 The Goal: To help AI choose the best possible option while anticipating the moves of
others.
 The Origin: Attributed to John von Neumann, who developed the framework to
analyse conflict and cooperation.
Five Categories of Games

The strategic interactions into five pairs:

Cooperative vs. Non-cooperative:

 Cooperative: Players can form alliances to win.

 Non-cooperative: Players cannot form alliances and must act independently.

Symmetric vs. Asymmetric:

 Symmetric: All players have the same goal but may use different strategies.
 Asymmetric: Players have different or conflicting goals.

Perfect vs. Imperfect Information:

 Perfect: All moves are visible to everyone (e.g., Chess).

 Imperfect: Opponents' moves or resources are hidden (e.g., Poker/Card games).

Simultaneous vs. Sequential:

 Simultaneous: Players act at the same time.

 Sequential: Players take turns and are aware of the previous moves of others (e.g.,
Board games).

Zero-sum vs. Non-zero Sum:

 Zero-sum: One player's gain is exactly equal to another player's loss.

 Non-zero Sum: Multiple players can gain advantages at the same time; one player's
gain does not necessarily mean another's loss

Concepts for AI

When applying Game Theory to AI, two advanced concepts are vital:

A. Nash Equilibrium

Named after John Nash, this is a state where no player has anything to gain by changing only
their own strategy. * In an AI context, if a system reaches Nash Equilibrium, it has found a
"stable" strategy. It assumes everyone else is also acting rationally to maximize their own
benefit.

 The "No Regrets" Outcome: Once the decision is made, a player wouldn't change it
even if they knew what the others had chosen.
B. Inverse Game Theory (Mechanism Design)

While standard Game Theory asks "Given the rules, how will players behave?", Inverse Game
Theory asks:

"What rules should I create to make players behave in a specific way?"

 Example: Designing an auction (like Google Ads) so that the most honest bidders win,
or creating a traffic system that prevents gridlock by incentivizing drivers to take
different routes.

Why it matters for AI

As AI moves into "multi-agent" environments (like self-driving cars on a road or trading bots
in a stock market), it must be able to:

1. Predict the selfish behaviour of others.

2. Act Ethically within pre-defined rules.
3. Optimize outcomes for the "social choice" (the best result for the whole group).

Nash Equilibrium

The Nash Equilibrium is a foundational concept in game theory that describes a "stable state"
where no player has an incentive to change their strategy, assuming everyone else keeps
theirs the same.

In the Prisoner’s Dilemma, the Nash Equilibrium is particularly famous because it leads to a
"terrible but stable" result where both players act in their own self-interest, but both end up
worse off than if they had cooperated.

Why Mutual Betrayal is the Nash Equilibrium

To find the equilibrium, you must look at the decision from the perspective of one player at
a time:

 If Prisoner A thinks Prisoner B will stay silent: Prisoner A is better off confessing (0
years vs. 1 year).
 If Prisoner A thinks Prisoner B will confess: Prisoner A is still better off confessing (5
years vs. 10 years).

Because confessing is the "best response" regardless of what the other person does, it is the
dominant strategy. Since both prisoners follow this logic, they both confess and receive 5
years.
Comparison of Outcomes
Prisoner A's Prisoner B's
Result (A, B) Type of Outcome
Choice Choice
Nash Equilibrium (Stable but
Confess Confess 5 yrs, 5 yrs
suboptimal)
Social Optimum (Best for both, but
Silent Silent 1 yr, 1 yr
unstable)
Confess Silent 0 yrs, 10 yrs Unfair/Unilateral betrayal
Silent Confess 10 yrs, 0 yrs Unfair/Unilateral betrayal

Why it's a "Dilemma"

The dilemma lies in the fact that the Nash Equilibrium (5, 5) is worse for both people than
the Social Optimum (1, 1). However, (1, 1) is unstable because if you are at (1, 1), both players
are constantly tempted to "cheat" and confess to get 0 years, which inevitably pulls the game
back toward the (5, 5) equilibrium.

Importance for AI

For AI developers, understanding Nash Equilibrium is critical for:

1. Safety: Predicting how autonomous cars or trading bots will behave when they
encounter "selfish" agents.
2. Stability: Designing systems (like auctions or traffic routes) where the "best" outcome
for the group is also the most stable outcome for the individuals.

Inverse Game Theory

Inverse Game Theory is the science of designing the "rules of the game" to achieve a specific
result. While standard Game Theory asks "How will players act?", Inverse Game Theory (also
known as Mechanism Design) asks "How do I build the system so that players’ selfish choices
lead to the goal I want?"

In the world of AI, this is the foundational logic behind Generative Adversarial Networks
(GANs).

The GAN Architecture: A Strategic Competition

A GAN is essentially a mathematical "game" played between two competing neural networks.
By setting them against each other, the AI "teaches itself" to create incredibly realistic data.
The two "players" in this game are:

 The Generative Model (The Counterfeiter): Its goal is to create fake data (like an
image of a human face) that looks exactly like the real thing. It studies data
distributions to "guess" how real samples are made.
 The Discriminative Model (The Detective): Its goal is to look at a sample and correctly
identify if it is "Real" (from the original dataset) or "Fake" (created by the Generator).

Reaching Nash Equilibrium in GANs

The "game" is played iteratively (over and over again).

1. The Generator gets better at making fakes to fool the Detective.

2. The Discriminator gets better at spotting the tiny flaws in the fakes.
3. The Result: Eventually, they reach a Nash Equilibrium. At this point, the Generator
produces such perfect fakes that the Discriminator can only "guess" with 50% accuracy
whether an image is real or fake. Neither can improve their strategy further.

Practical Applications: Creating Reality

GANs are most famous for generating hyper-realistic images of things that don't exist.

 Image Generation: Creating faces of people who have never existed.

 Art & Design: Generating new textures or patterns for fashion and architecture.
 Data Augmentation: Creating "fake" medical data to train other AIs when real patient
data is private or scarce.

Symmetric vs. Asymmetric Reality

 Symmetric Games: Easier to reach equilibrium because everyone has the same
goals/starting points.
 Asymmetric Games: Much more common in the real world. In a GAN, the game is
asymmetric because one player is creating while the other is evaluating. Designing
environments for these complex interactions is why Inverse Game Theory is so vital
for AI researchers.

Multi-Agents Reinforcement Learning (MARL)

Multi-Agent Reinforcement Learning (MARL) represents the next level of AI complexity.

While standard Reinforcement Learning (RL) focuses on a single "agent" learning from its
surroundings, MARL deals with a community of agents that must learn to survive or succeed
while interacting with each other.
From Single Agent to Multi-Agent

To understand MARL, we have to look at how the learning process changes:

 Standard RL (Single Agent): Uses Markov Decision Processes. An agent performs an

action in an environment, receives a reward (positive) or punishment (negative), and
adjusts its "policy" (strategy) to maximize rewards. It is like a mouse learning a maze;
the maze doesn't move or change its own strategy.
 MARL (Multiple Agents): The environment is no longer "stationary." Because other
agents are also moving and learning, the "best" move for Agent A depends on what
Agent B, C, and D are doing at that exact moment.

The Complexity Problem: "Mean Field Scenarios"

The modelling MARL is not "trivial." If you have 2 cars, the interactions are simple. If you have
2,000 cars, the number of possible interactions increases exponentially.

To solve this, researchers use Mean Field Scenarios (MFS):

 The Logic: Instead of tracking every single interaction between every single agent, the
AI looks at the "mean" (average) behaviour of the crowd.
 The Assumption: It assumes all agents have similar reward functions (e.g., all cars
want to reach their destination safely and quickly). This simplifies the math
significantly.

The Traffic Flow Example

The text provides a perfect real-world application: Self-driving cars.

 The Conflict: Imagine two AI cars at an intersection. Both calculate that the "fastest"
route is via Main Street. If they both take it, they create a traffic jam.
 The Solution via Game Theory: By using MARL combined with Game Theory, the cars
can reach a Nash Equilibrium.
 Collaboration: The agents learn that to achieve the best flow for the group, some
individuals might need to take a slightly longer route so that no one gets stuck in a
total deadlock.

Neuromorphic Computing

Neuromorphic computing is the design of computer hardware and software inspired by the
biological architecture of the human brain and nervous system. Unlike standard computers
that follow strict logic, these systems are designed to be "cognition machines."
  Multidisciplinary: It combines biology, physics, math, electronic engineering, and
computer science.
 The Goal: To enable machines to learn, retain information, and make logical decisions
with the same flexibility as a human.

The Efficiency Gap: Brain vs. Laptop

 The Human Brain: Uses only about 20 watts of power (half of a standard laptop).
 The Challenge: Traditional AI requires massive energy-hungry servers. Neuromorphic
computing aims to bridge this gap by performing fast computation with low power.

Five Key Features of Neuromorphic Systems

These systems resemble the human brain through five specific characteristics:

1. Massively Parallel: They handle many tasks simultaneously rather than one by one.
2. Event-Driven: They only activate when there is a specific "event" or change in the
environment (saving energy).
3. Economical: Extreme power efficiency.
4. Flexible/Adaptable: They can generalize and learn from new situations.
5. Fault-Tolerant: The system continues to work even if some parts (neurons) fail.

How It Works: Spiking Neural Networks (SNN)

Traditional AI uses digital signals (0s and 1s). Neuromorphic systems use analog circuitry to
mimic the "electrical pulses" of the brain.

 Neurons and Synapses: The hardware uses "nodes" (neurons) and "connections"
(synapses).
 Varying Strength: Just like in a real brain, the system controls how much electricity
passes through a connection, mimicking the strength of a memory or a thought.
 Spiking Neural Networks (SNN): This is the name for the system that transmits these
electrical pulses.

Breaking the Von Neumann Bottleneck

This is the most significant technical [Link] "Von Neumann" Architecture

In a normal computer, the CPU (processor) and the Memory (storage) are separate. Data
must travel back and forth between them constantly. This travel takes time and energy,
creating a bottleneck.
Neuromorphic Architecture

Neuromorphic chips collocate (place together) memory and processing on the individual
neuron.

 There is no separate data path.

 Information is processed exactly where it is stored.
 The Result: Massive increases in speed and energy efficiency.

Neuromorphic Computing and Artificial General Intelligence (AGI)

Neuromorphic Computing to the ultimate goal of AI research: Artificial General Intelligence

(AGI). It suggests that moving away from traditional computer architecture is the only way to
replicate the versatile, efficient, and self-aware nature of a human.

The Path to AGI

Artificial General Intelligence (AGI) refers to a machine that possesses intelligence equal to a
human.1 while traditional "Narrow AI" (like a chess bot) can only do one thing, an AGI would
be a "jack-of-all-trades."2

The neuromorphic Computing is the most promising path to AGI because it mimics the brain's
physical ability to:

 Reason in uncertain situations.

 Plan and Learn from experience.
 Communicate naturally and Represent Knowledge.

Testing for AGI

How do we know if a machine has reached "human-level" intelligence? The text mentions
two famous benchmarks:

1. The Turing Test: Can the machine carry on a conversation so well that a human cannot
tell if they are talking to a computer or a person?
2. The Robot College Student Test: Can the machine enroll in college, attend classes,
take exams, and earn a degree just like a human?

Human Brain vs. Supercomputers: The Gap

To understand why we need Neuromorphic chips, we must look at how inefficient even our
best "traditional" supercomputers are compared to the biological brain.
Feature Human Brain Fugaku Supercomputer
Power
20 Watts (Low) 28 Megawatts (Extremely High)
Consumption
Needs elaborate, massive cooling
Cooling Sits at 37°C inside the skull
systems
Architecture Massively Parallel Mostly Serial (von Neumann)
Can write poetry, drive, and High-speed math, but struggles with
Versatility
recognize faces simultaneously cross-domain adaptability
High (can function if some neurons
Fault Tolerance Low (errors often crash the system)
die)

Current Research: SpiNNaker

The SpiNNaker (Spiking Neural Network Architecture) as a "hot" research project within the
Human Brain Project. It is a neuromorphic supercomputer designed to simulate the
functioning of the human brain's neural networks in real-time, moving us closer to a machine
that can truly "think."3

The Ethical & Legal Debate

If a machine eventually develops consciousness or self-awareness, how should we treat it?

 The Animal Rights View: Some argue that a conscious AGI should have legal
protections similar to non-human animals.4
 The "Machine" View: Others argue that no matter how smart it gets, it lacks true
biological consciousness and doesn't require "rights."

Uses of Neuromorphic Systems

Neuromorphic systems represent a shift from Cloud-centric AI (where your phone sends data
to a server) to Edge-intelligent AI (where your phone processes data itself). This section
outlines how this technology is moving from theory to real-world hardware.

The Shift to Edge Computing

In traditional AI, devices like smartphones act as "mailmen"—they collect data and send it to
a massive cloud server for processing.

 The Problem: This creates latency (delay) and requires a constant internet connection.
 The Neuromorphic Solution: Processing happens on-device. This makes AI faster,
more private, and capable of working in remote areas.
Solving "Brain-Like" Problems

Current AI is excellent at following rules and patterns. Neuromorphic AI, however, is being
designed to handle Constraint Satisfaction Problems.

 What it is: Finding the best possible answer when you have a hundred different "rules"
that conflict with each other.
 Handling "Noise": Unlike digital computers that can crash if data is slightly "blurry" or
"noisy," neuromorphic systems can find meaning in uncertain or messy information,
just as you can recognize a friend's voice in a crowded room.

Major Systems in Use Today

The several "heavy hitters" in the neuromorphic world that are already operational as of 2026.

Intel’s Loihi & Pohoiki Beach

 Hardware: Intel's Loihi chips are the building blocks. 64 of these chips were used to
create Pohoiki Beach (8 million neurons).
 The Goal: As of late 2025/2026, researchers have scaled this to Pohoiki Springs, which
delivers 100 million neurons.
 Applications: Creating artificial skin that can "feel" and developing powered
prosthetic limbs that respond to a user's intent with biological speed.

IBM’s TrueNorth & Blue Raven

 Hardware: TrueNorth consists of 1 million neurons per chip; the supercomputer

version scales to 64 million neurons.
 Military Use: In partnership with the Air Force, IBM developed Blue Raven.
 The Goal: To create drones that are smarter and lighter. Because Blue Raven uses so
little energy (about 40 watts for 64 million neurons), these drones can stay in the air
longer and process images locally without needing a bulky cooling system.

The Human Brain Project (SpiNNaker & BrainScaleS)

The EU's 10-year initiative concluded its primary phase but left behind two massive platforms:

 SpiNNaker: A "million-core" digital supercomputer at the University of Manchester. It

is designed to simulate a billion neurons to help doctors understand brain diseases.
 BrainScaleS: An analog system. While SpiNNaker is digital, BrainScaleS uses physical
electronics to mimic the actual physical behavior of neurons, allowing it to run
simulations much faster than real-time.
System Developer Scale (Neurons) Primary Use Case

Pohoiki Beach Intel 8 Million Prosthetics, Tactile Sensing

Blue Raven IBM / Air Force 64 Million Autonomous, Low-power Drones

SpiNNaker University of Manchester 1 Billion (Projected) Large-scale Brain Simulation

BrainScaleS Human Brain Project Various Fast-than-life biological emulation

Challenges to Using Neuromorphic Systems

Implementing neuromorphic systems represents a "wholesale change" to the computing

world. It is not a simple upgrade but a fundamental shift that creates several massive hurdles
for engineers and developers today.

The Processing Paradigm Shift

In a traditional Von Neumann system, visual data is processed as a sequence of still images
(frames). If you record a 60 FPS video, the computer processes 60 complete snapshots every
second, even if nothing in the scene is moving.

In contrast, a Neuromorphic processor uses Event-Based Processing.

 It only records "spikes" when a pixel detects a change (like a bird flying across a still
sky).
 This is extremely efficient but requires a complete rethink of how we record and
analyze data.

The Software and Programming Gap

For over 70 years, almost every line of code has been written for Von Neumann's "step-by-
step" logic. Neuromorphic systems are asynchronous and probabilistic, meaning they don't
follow a central clock.

 New Languages: Standard languages like C++ or Python are designed for sequential
tasks. We need new "Spiking" languages to program the timing and weights of
individual artificial synapses.
  Lack of Standards: Currently, there is no "Windows" or "Linux" for neuromorphic
chips. Each company (Intel, IBM) has its own unique architecture, making it difficult
for a single software ecosystem to grow.

Hardware Re-Engineering

Because neuromorphic chips collocate memory and processing, our existing "building blocks"
of hardware are no longer sufficient.

 New Sensors: We need "neuromorphic cameras" (Dynamic Vision Sensors) and

"tactile sensors" that output spikes instead of digital signals.
 New Memory: Emerging technologies like Memristors and MRAM are being
developed to act as "artificial synapses" that can store and process data
simultaneously.
 Manufacturing Complexity: Integrating these biological-style components into
standard silicon manufacturing (CMOS) is technically difficult and expensive.

Integration with Legacy Systems

As of 2026, the biggest challenge is Interoperability. Most of the world's data centers and
apps are built on traditional hardware.

The Challenge: How do you plug a brain-like, low-power chip into a high-power, frame-based
data center?

Even if a neuromorphic chip is 1,000x more efficient, it will take years to build the "bridges"
(adapters and translators) required to make it work with our existing digital world.

EXPERT SYSTEM

An Expert System (ES) is one of the most established branches of Artificial Intelligence. It
emulates the decision-making ability of a human expert by using a "knowledge-heavy"
approach rather than just raw processing power.

Unlike a regular computer program that follows a fixed path (if A then B), an expert system
uses Reasoning. It doesn't just calculate; it "thinks" through a problem using the same logic a
doctor, scientist, or engineer would use.

The Architecture: How It Works

The working of an expert system, as referenced in your Fig. 9.1, relies on three primary
components that separate the "knowledge" from the "thinking."
 1. Knowledge Base (KB): This is the heart of the system. It contains high-quality, domain-
specific data.
o Facts: Proven, objective information (e.g., "Water boils at 100°C").
o Heuristics: These are "rules of thumb" or educated guesses based on
experience (e.g., "If a car won't start and the lights are dim, it's likely a battery
issue").
2. Inference Engine: This is the "brain" or reasoning component. It doesn't store data;
instead, it applies the rules from the Knowledge Base to the user's current problem to
reach a conclusion.
3. User Interface: This is the communication bridge. It allows a non-expert to ask a
question (like "Why is my plant leaves turning yellow?") and receive an expert-level
answer.

History and Evolution

 The Origin: The field began in the mid-1960s and 1970s. One of the first major systems
was DENDRAL (1965), which helped chemists identify molecular structures, followed
by MYCIN (1972), which diagnosed blood infections.
 Modern Example: You interact with an expert system daily via Google Search. When
you start typing and Google suggests a correction or a specific search term, it is using
an expert system based on linguistic rules and massive data to "predict" the correct
expertise needed for your query.

Characteristics

 Domain-Specific: An expert system for heart surgery cannot help you with a legal
contract. It is highly specialized.
 Knowledge-Dependent: Its performance is directly tied to the KB. As the text states:
More Knowledge = Better Performance.
 Assistant, Not Replacement: Its goal is to provide a "second opinion" or handle
complex data so that humans can make the final, informed call.
Human Expert vs. Expert System

Feature Human Expert Expert System

Availability Available during work hours Available 24/7

Consistency Can be affected by mood/fatigue Always consistent and logical

Transferability Knowledge is hard to transfer Knowledge is easily copied/stored

Cost High (Salary/Training) Low (After initial development)

Popular Examples of the Expert System

Scientific & Engineering Systems

These systems were designed to handle complex calculations and structural configurations
that would take humans much longer to complete.

 DENDRAL (Chemical Analysis):

o Focus: Organic Chemistry.
o Function: It was the first "knowledge-intensive" system. It analyzed "mass
spectra" (the fingerprint of a molecule) to figure out the 3D structure of
unknown chemicals.
 R1/XCON (System Configuration):
o Focus: Computer Engineering.
o Function: Developed for the Digital Equipment Corporation (DEC), it helped
sales reps order the correct parts for complex computer systems. It ensured
that all chosen software and hardware were compatible, acting as a
"customizer."

Medical Diagnosis Systems

Medical Expert Systems are the most well-known because they save lives by providing doctors
with a "digital second opinion."

 MYCIN (Infectious Disease):

o Focus: Blood infections (Bacteraemia/Meningitis).
o Function: It is famous for using backward chaining (starting with a goal, like a
specific disease, and working backward to see if the symptoms fit). It didn't just
 name the bacteria; it calculated the exact dosage of antibiotics needed based
on the patient's body weight and health.
 PXDES (Lung Cancer Identification):
o Focus: Radiology/Oncology.
o Function: It specializes in Image Analysis. By looking at scans of the upper
body, it can identify the specific stage and type of lung cancer.
 CaDeT (Early Cancer Detection):
o Focus: Early Diagnostics.
o Function: While PXDES identifies the type of cancer, CaDeT acts as a "Support
System" to catch cancer at its earliest, most treatable stages.
 DXplain (Clinical Decision Support):
o Focus: General Diagnosis.
o Function: This is a broad tool for doctors. A physician enters clinical "findings"
(symptoms, lab results, etc.), and the system provides a list of potential
diseases that match those findings.

Comparison of Medical Systems

System Goal Input

MYCIN Recommend Treatment Blood test results & Patient data

PXDES Classify Cancer Stage Medical images/pictures

DXplain Suggest Potential Diseases General clinical findings

Why "Backward Chaining"?

The text mentions MYCIN uses backward chaining. In medicine, this is very effective because:

1. The doctor starts with a Hypothesis (e.g., "Does this patient have Meningitis?").
2. The AI looks through the Knowledge Base for the evidence required for that disease.
3. It then asks the doctor for specific data (e.g., "Is there a high white blood cell count?")
to confirm or deny the goal.
Characteristics of an Expert System
1. High performance: solving any type of complex problem of a specific domain with high
efficiency and accuracy.
2. Understandable: The system takes input in human language and provides the output
in the same language.
3. Reliable: An expert system generates efficient and accurate output. This makes it a
reliable system.
4. Highly responsive: An expert system generates results for any complex query within a
very short period of time, making it highly responsive. The total time by an expert
system should be less than that taken by a human expert.
5. No memory limitations: An expert system can store and exploit huge amounts of data.
6. Expertise in a domain: the knowledge gained from these human experts into the
expert system, the results are provided by utilizing all the facts and knowledge
7. Not affected by emotions: emotions such as fatigue, anger, depression, anxiety, etc.
This helps them to give a consistent performance every time they are used.
8. Not biased: To respond to any query, the expert systems check and consider all the
available facts.
9. Reduce cost: Human experts charge a heavy amount but with an expert system,
multiple experts can be consulted at much cheaper rates.
10. Non-perishable: Human experts are perishable as they have a limited life span but an
expert system is permanent.
11. Intelligent: These systems use knowledge base and inference engine to solve complex
problems.

Components of an Expert System

User Interface: The Communication Bridge

The User Interface (UI) is designed to make complex AI accessible to everyday people. It acts
as the "translator" between the user's natural language and the system's technical logic.

Functions:

 Query Handling: It accepts questions from the user and displays the final solution.
 Transparency (The "Why" Module): A unique feature of Expert Systems is their ability
to explain their reasoning. The UI can provide text, verbal narrations, or a list of rules
to show exactly how it reached a conclusion. This builds credibility and trust.
 Efficiency: A good UI focuses on helping the user reach their goal quickly while making
the best use of the data the user provides.
Inference Engine: The "Brain"

The Inference Engine is the active processing unit. It doesn't just store data; it reasons with
it. By looking at the facts provided by the user and the rules stored in the Knowledge Base, it
"infer" or deduces a result.

Types of Inference Engines:

1. Deterministic Inference Engine:

o Logic: Uses absolute facts and "If-Then" rules.
o Result: The conclusion is treated as 100% true (e.g., "If the temperature is
$>100$°C, then the water is boiling").
2. Probabilistic Inference Engine:
o Logic: Based on the likelihood of an event.
o Result: The conclusion includes a degree of uncertainty (e.g., "Based on these
symptoms, there is an 80% probability of a bacterial infection").

Advanced Capabilities:

 Conflict Resolution: If two different rules apply to the same situation, the engine is
smart enough to select the "best" rule for that specific context.
 Debugging: It can trace through its own logic to find errors or explain its path to the
user through the Explanation Module.

Inference Strategies: Forward vs. Backward Chaining

The inference engine uses two different paths to connect data with a result.

Forward Chaining (Data-Driven)

 How it works: It starts with known facts and moves toward a conclusion. It follows
the logic: "If these facts exist, what will the result be?"
 Purpose: To predict future outcomes.
 Example: If interest rates rise (fact), the system predicts that stock prices will fall
(conclusion).

Backward Chaining (Goal-Driven)

 How it works: It starts with a goal (an event that happened) and works backward to
find the facts that caused it. It asks: "Since this happened, what must have caused it?"
 Purpose: To diagnose or find a "root cause."
 Example: Starting with a diagnosis like blood cancer (goal) and looking back through
medical history and symptoms to find the origin.
The Knowledge Base (KB)

The knowledge as the combination of Data + Information + Past Experience.

Types of Knowledge

The Knowledge Base is split into two distinct categories to mimic human expertise:

1. Factual Knowledge: Objective, widely accepted facts within a field (e.g., "The human
body typically has 206 bones").
2. Heuristic Knowledge: "Rules of thumb" based on practice and experience. It is less
about certainty and more about probability and evaluation (e.g., "Usually, when a
patient has a high fever and a rash, it indicates an infection").

Case Study: The 5 Steps of MYCIN

The MYCIN to illustrate the lifecycle of an expert system in action:

1. Creation: Human experts (doctors/scientists) provide knowledge about bacteria and

symptoms to build the initial KB.
2. Input: A doctor provides a specific patient's symptoms and medical history to the
system.
3. Refinement: The system interacts with the patient/user to get general data like age
and gender to narrow down the rules.
4. Processing: The Inference Engine applies "If-Then" rules to the KB. (In MYCIN's case,
it uses backward chaining to see if the symptoms match a specific bacterial profile).
5. Outcome: The system provides a diagnosis or treatment plan through the User
Interface.

Participants in the Development of Expert System

The Expert (The Source)

The Expert is a human being with "extra-ordinary" knowledge in a specific field (like a
cardiologist, a structural engineer, or a senior chemist).

 Role: They provide the "raw material"—the facts, the "rules of thumb" (heuristics),
and the logic they use to solve problems.
 Importance: Since the system’s performance is directly tied to the quality of its

The Knowledge Engineer (The Translator)

The Knowledge Engineer acts as the architect and the bridge between the human expert and
the machine.

 Role: They interview the Expert to "extract" their thought processes. They then
translate this messy human logic into a structured format that a computer can
understand (usually If-Then rules).
 Task: They are responsible for "Knowledge Representation"—organizing the
information so the Inference Engine can process it efficiently.

The End User (The Beneficiary)

The End User is the person seeking an answer.

 Role: They interact with the completed system through the User Interface.
 Background: They often lack deep expertise in the subject. For example, a junior
doctor might use a medical expert system to confirm a diagnosis, or a customer might
use a troubleshooting bot to fix a router.

The Workflow: How the Participants Connect

Participant Contribution Goal
Expert Domain Wisdom To share "How" and "Why" they solve problems.
Knowledge
Technical Structure To convert wisdom into a digital Knowledge Base.
Engineer
End User Complex Queries To receive expert-level advice or solutions.

Capabilities of the Expert System

1. Advise

An expert system provides expert-level advice to users for any query within its defined
domain, similar to consulting a human expert.

2. Decision-Making

It can make complex decisions by analyzing rules, facts, and data, commonly used in finance,
medicine, engineering, and management.

3. Demonstrate Devices

Expert systems can demonstrate product features, specifications, and correct usage, helping
users understand new or complex devices.

4. Problem-Solving
They are capable of solving complex domain-specific problems using stored knowledge and
logical reasoning.

5. Explain a Problem

An expert system can explain the given problem in detail, breaking it down into
understandable components for the user.

6. Interpret Input

It interprets user input (data, symptoms, conditions, queries) and converts it into a form
suitable for internal processing.

7. Predict Results

By analyzing historical data and patterns, expert systems can predict outcomes such as trends,
failures, or future conditions.

8. Assist in Decision-Making

Rather than replacing humans, expert systems support humans by offering recommendations
and insights for better decision-making.

9. Diagnose

Widely used in the medical field, expert systems help diagnose diseases by comparing
symptoms with stored medical knowledge.

10. Justify Conclusions

A key strength! The system can justify its conclusions by explaining the rules, facts, and logic
used to reach a decision.

11. Suggest Alternatives

If multiple solutions exist, the expert system can suggest alternative approaches, helping
users choose the most suitable option

However, expert systems cannot do the following:

1. Substitute human decision-makers

2. Possess human capabilities
3. Produce accurate output for inadequate knowledge base
4. Refining knowledge base 5. Use emotions to make decisions

Advantages of Expert Systems

1. Results obtained by an expert system are highly reproducible.

 2. They can be used in risky places where it is unsafe for humans to work.
3. They provide accurate results when the knowledge base is regularly updated.
4. They offer consistent performance at high speed and are not affected by emotions,
tension, or fatigue.
5. They are easily available due to mass production of software.
6. They are affordable because production costs are reasonable.
7. They operate at great speed and reduce human effort.
8. They can explain how a solution was obtained.
9. They improve the quality of decisions.
10. They collect scarce expert knowledge and use it efficiently.

Limitations of Expert Systems

1. Results may be incorrect if the knowledge base contains wrong or outdated

information.
2. Solutions may differ from those given by human experts since each problem is unique.
3. They lack creativity, especially in extraordinary situations.
4. They have high development and maintenance costs.
5. Knowledge acquisition is difficult.
6. Their success depends on knowledge obtained from human experts.
7. They cannot learn on their own and require manual updates.
8. They cannot produce correct results with insufficient knowledge.
9. They require extensive training.
10. They do not possess emotions.
11. They lack common sense.
12. They are developed for a specific domain only.
13. They cannot make decisions in extraordinary situations.
14. They follow the Garbage-In Garbage-Out (GIGO) principle.
15. They have high maintenance costs.

Applications of Expert Systems

1. Design & Manufacturing

VLSI systems, camera lenses, automobiles.
2. Knowledge Publishing
Tax advisors and other advisory systems.
3. Finance
Fraud detection, suspicious activity identification, stock trading, loan approvals.
4. Medical Diagnosis
Diagnosis of diseases and treatment support.
 5. Planning & Scheduling
Project planning, airline scheduling, cargo management, manufacturing schedules.
6. Software Development
Tracking project progress.
7. Monitoring Systems
Detecting pipeline leaks and continuous data comparison.
8. Process Control
Monitoring and controlling physical processes.
9. Automobile & Electronics
Fault detection in vehicles and computers.
10. Complex System Diagnosis
Electronic, electromechanical, and diesel-electric systems.

Expert System Technology

1. Expert System Development Environment

The development environment includes hardware and tools such as:

1. Workstations, minicomputers, and mainframes

2. High-level symbolic programming languages like
o LISP (LISt Programming)
o PROLOG (PROgrammation en LOGique)
3. Large databases

2. Tools

Tools help reduce the cost of expert system development by providing:

1. Powerful editors
2. Debugging capabilities with multi-window support
3. Rapid prototyping
4. In-built definitions for model design, knowledge representation, and inference design

3. Shells

A shell is an expert system without a knowledge base. It provides modules for:

 Knowledge acquisition
 Inference engine
 User interface
 Explanation facility
Popular expert system shells include:

 JESS (Java Expert System Shell) – provides Java API for creating expert systems
 Vidwan – enables knowledge encoding using IF–THEN rules

Development of Expert Systems

The development of an expert system is an iterative process. The steps involved are:

1. Identify Problem Domain

Identify the problem to be solved and ensure it is suitable for an expert system. Identify
domain experts, analyze cost effectiveness, and define system characteristics through
coordination between knowledge engineers and domain experts.

2. Design the System

Select appropriate technology including hardware, tools, and shells. Plan system integration
with existing systems and databases, and design a framework for representing domain
knowledge.

3. Develop the Prototype

Acquire domain knowledge from human experts and represent it using IF–THEN–ELSE rules.
Determine how heuristic knowledge and explanation mechanisms will be incorporated.

4. Test and Refine the Prototype

Test the prototype using sample cases and with end users to identify errors or performance
issues. Use test results to refine the system.

5. Develop and Complete the System

Test the system with users, databases, and other information systems to ensure smooth
integration and correct functionality. Document the system and train users.
Expert systems are also used for information management and help desk management.

6. Maintain the System

Maintenance is continuous. Regularly review and update the knowledge base, and
incorporate new user requirements and interface changes as needed.
Internet of Things

Defining IoT

IoT is a massive network of "interrelated" things. The core requirement for an object to be
part of the IoT is:

 Unique Identifier: It must have a way to be identified (usually an IP Address).

 Data Transfer: It must be able to send or receive data over a network without needing
a human to type or click a button.
What Can Be a "Thing"?

The beauty of IoT is its versatility. The text provides several examples of how "things" can be
biological or mechanical:

 People: A person with a heart monitor implant that sends data to a doctor.
 Animals: Farm animals with biochip transponders to track health and location.
 Automobiles: Cars with tyre pressure sensors that alert the driver automatically.
 Everyday Objects: Lamps, coffee makers, and washing machines that can be
controlled remotely.

The Role of IPv6

A critical technical enabler for IoT is IPv6.

 The Problem: The old system (IPv4) had a limited number of addresses.
 The Solution: IPv6 provides a "huge availability of address space," essentially giving us
enough addresses to assign one to every atom on Earth. This allows billions of devices
to have their own unique "ID card" on the internet.

History: The Carnegie Mellon Coke Machine

The first "Smart Appliance" wasn't a fancy refrigerator, but a Coke machine in the early 1980s.
 Programmers at Carnegie Mellon University connected the machine to the internet so
they could check two things from their desks:
1. Is the machine empty?
2. Is the soda cold?
 This saved them the trip of walking to the machine only to find it empty or warm—the
very first example of IoT increasing efficiency.

Future Growth: A World of Connections

The text highlights a massive growth trajectory:

 2022: 14.4 billion active connections.

 2025: Expected to reach 27 billion connected devices.

This creates three types of relationships in our modern world:

1. People-People (e.g., Video calls)
2. People-Things (e.g., Controlling your home AC from your phone)
 3. Things-Things (e.g., Your car communicating with a smart garage door to open
automatically)

Examples of Applications of IoT

1. Smart Car Navigation

Suggests best routes, provides traffic updates, and sends delay notifications.
2. Smart Home Automation
Automatically triggers alarms and starts devices like coffee machines.
3. Wearable Fitness Trackers
Tracks activity, productivity, and provides health insights.
4. Self-Driving Cars
Uses sensors to detect objects and ensure safe navigation.
5. Smart Football
Tracks distance, speed, and records training statistics.
6. Smart Refrigerator
Alerts when groceries run out or food expires.
7. Smart Air Conditioner
Allows remote temperature control through smartphones.
8. Smart Cities
Helps reduce traffic congestion, manage resources efficiently, and lower pollution
and crime.

IoT Products

1. Amazon Echo (Alexa)

 Voice assistant for smart homes

 Performs tasks like playing music, giving weather updates, and ordering services

2. Fitbit One (Wearables)

 Tracks steps, calories, and sleep

 Syncs with smartphones and computers

3. Barcelona – Smart City

 Uses IoT for smart parking and environmental improvements

 Recognized as a leading smart city

4. Popular IoT Platforms

 Amazon Web Services (AWS)

  Microsoft Azure
 IBM Watson
 Cisco IoT Cloud Connect
 Salesforce IoT Cloud
 Oracle Integrated Cloud
 GE Predix

5. AT&T – Connected Car

 Added 1.3 million connected cars in 2016

 Total reached 9.5 million connected cars

Challenges in IoT

1. Security

 Privacy and data-sharing concerns

 Connected devices are targets for hackers (cars, homes, etc.)
 Lack of encryption causes security vulnerabilities

2. Scalability

 Large volumes of data require big data analytics and cloud storage
 Devices must scale to handle increasing data

3. Interoperability

 Lack of common standards for IoT devices

 Difficulty integrating with legacy systems

4. Lack of Standardization

 Compatibility issues among IoT devices

 Prevents fully connected systems

5. Lack of Government Support

 Limited IoT-specific safety and security regulations

 Insufficient regulatory oversight

6. IoT Malware and Ransomware

 Devices vulnerable to malware and ransomware attacks

  Hackers can take control and demand ransom

7. Connectivity

 Poor internet connectivity affects performance

 Stable connections are required for monitoring and control

8. Privacy Policies

 Unclear regulations on private and public data

 Example: IoT toys collecting children’s data without consent

9. Bandwidth Strain

 Increasing IoT devices strain network bandwidth

 Requires higher bandwidth and powerful servers

10. IoT Skills Gap

 Shortage of trained IoT professionals

 Difficulty training staff due to rapid IoT growth

Sensors
Types of Sensors with Description and Examples

 Flame Detector
Senses fire presence and quality.
Example: Combustion control systems.
 Electrical Sensor
Detects electrical parameters like current and voltage.
Example: Monitoring electrical systems.
 Contact Sensor
Detects physical touch or contact.
Example: Alarm systems for doors and windows.
 Non-contact Sensor
Does not require physical touch.
Example: Radar guns for speed detection.
 Speed Sensor
Measures speed of an object or vehicle.
Example: Vehicle monitoring.
 Ultrasonic Sensor
Measures distance using high-frequency sound waves.
Example: Object distance measurement in robotics.
 Particle Sensor
Detects airborne particulates.
Example: Bin and baghouse monitoring.
 Motion Sensor
Senses movement or stoppage.
Example: Conveyor detection, stalling bearings.
 Metal Detector
Senses the presence of metal.
Example: Security checks.
 Level Sensor
Determines the height of substances in tanks.
Example: Control and display systems.
 Leak Sensor
Detects unwanted discharge of liquids or gases.
Example: Monitoring gas leaks.
 Humidity Sensor
Measures water content in the air.
Example: Control systems for temperature regulation.
 Gas and Chemical Sensor
Detects gases or chemicals.
Example: Environmental monitoring.
 Force Sensor
Measures forces like weight or torque.
Example: Control systems for load measurement.
 Flow Sensor
Detects movement of gases, liquids, or solids.
Example: Industrial processing systems.
 Flaw Sensor
Detects inconsistencies on surfaces.
Example: Manufacturing processes like welding.
 Oxygen Sensor
Detects the gasoline/oxygen ratio.
Example: Car emission control systems.
 Motion Sensor (Energy Interruption)
Detects interruption in energy flow.
Example: Home security lights, automatic doors.
  Vision and Imaging Sensor
Detects objects or colors in view.
Example: Cameras for object detection.
 Temperature Sensor
Measures thermal parameters.
Example: Industrial processes for gases, liquids, and solids.
 Radiation Sensor
Detects alpha, beta, or gamma radiation.
Example: Surveys and sample counting.
 Proximity Sensor
Detects presence of objects without contact.
Example: Manufacturing operations.
 Pressure Sensor
Detects force per unit area in gases or liquids.
Example: Control and display devices.
 Position Sensor
Senses position of objects like valves or doors.
Example: Control systems for machinery.
 Photoelectric Sensor
Detects objects and senses color.
Example: Manufacturing and material handling automation.

Artificial Intelligence of Things (AIoT)

 Artificial Intelligence of Things (AIoT) combines Artificial Intelligence (AI) with the
Internet of Things (IoT)
 Purpose of AIoT:
o Enhance efficiency
o Improve decision-making
o Enable better data management
 Artificial Intelligence (AI):
o Simulates human intelligence
 Internet of Things (IoT):
o Connects devices to transfer data over networks
o Operates without human intervention
 AIoT improves traditional IoT by:
o Adding machine learning algorithms
o Enhancing intelligent decision-making
How Does AIoT Work?

 AI is embedded directly into IoT devices

 Devices are connected through IoT networks
 Working process:
o Devices collect data
o Data is analyzed using AI techniques
o Results improve productivity and efficiency
 Edge AI:
o Processes data closer to the source
o Reduces bandwidth usage
o Minimizes delays in data processing

Where Does AI Unlock IoT?

 AI enhances the analysis step in the IoT process

 Improved analysis directly increases system efficiency

AIoT helps to:

 Manage large volumes of data

 Analyze data efficiently
 Extract meaningful insights
 Provide fast and accurate data analysis
 Balance:
o Localized intelligence (edge devices)
o Centralized intelligence (cloud systems)
 Ensure:
o Data privacy
o Data security

Applications and Examples of AIoT (Artificial Intelligence of Things)

1. Smart Cities

 Use of smart sensors, lights, and meters

 Data collected is processed using AI
 Improves:
o Operational efficiency
o Economic growth
o Quality of life of residents
2. Smart Retail

 Smart cameras identify customers using self-checkout counters

 Cameras and sensors:
o Track customer movement
o Predict arrival at checkout lines
 Benefits:
o Dynamic staffing
o Increased cashier productivity
o Reduced checkout waiting time
 AIoT enables:
o Customer behavior analysis
o Proximity detection
o Personalized marketing campaigns
o Increased footfall in physical stores

3. Smart Home

 Smart appliances collect data from user interactions

 AI analyzes data to:
o Understand user habits
o Provide customized support

4. Manufacturing

 Use of smart chips in equipment

 Helps to:
o Detect malfunctioning equipment
o Predict correct time for replacement

5. Hiring

 AIoT integrates data from:

o Social media platforms
o HR-related systems
 Helps in:
o Identifying the right talent
o Reducing bias in hiring

6. Autonomous Vehicles

 Collect real-time data using:

o Video cameras
 o Sensors
 AI analyzes data to:
o Monitor driving conditions
o Make instant driving decisions

7. Robots

 Autonomous robots use multiple sensors

 AI processes environmental data to:
o Make intelligent decisions
o Perform optimized movements

8. Healthcare

 Medical devices and wearables collect:

o Real-time health data (e.g., heart rate)
 Healthcare systems generate large volumes of data:
o Patient records
o Medical images
o Test results
 AIoT helps to:
o Process data quickly
o Support diagnosis and treatment decisions
o Improve diagnostic accuracy
o Enable telemedicine and remote care
o Reduce administrative workload

9. Smart Thermostat Solution

 Smartphone integration allows:

o Remote temperature monitoring
o Temperature control from anywhere
 Users can:
o Set temperature schedules
o Customize preferences
 Example:
o Nest smart thermostat uses AI-powered IoT

10. Drone Traffic Monitoring

 Drones collect real-time traffic data

 AI analyzes data to:
 o Reduce congestion
o Make traffic diversions
o Adjust speed limits
o Control traffic signal timing automatically
 Example:
o ET City Brain by Alibaba Cloud
o Detects accidents and illegal parking
o Alters traffic lights to help ambulances reach hospitals faster

11. Office Buildings

 Smart office buildings use:

o Environmental sensors
 Sensors detect:
o Number of people present
 AI adjusts:
o Lighting
o Temperature
 Benefits:
o Improved energy efficiency
 Facial recognition technology:
o Controls access
o Maintains employee attendance

12. Fleet Management

 AIoT is used to:

o Monitor vehicles
o Reduce fuel costs
o Track vehicle maintenance
o Identify unsafe driving behavior
 Uses:
o GPS
o Sensors
o AI systems
 Helps companies manage fleets efficiently
Benefits and Challenges of AIoT

Benefits of AIoT

 Increases operational efficiency by analyzing and adjusting system operations in real

time
 Enables on-the-fly decision-making using real-time data analysis
 Reduces workload and saves time by automating data analysis
 Provides scalable systems that can connect more devices as needed
 Improves risk management by predicting potential problems in advance
 Enhances products and services through real-time data insights

Challenges of AIoT

 Security: Protecting sensitive and confidential data is a major concern

 Scalability: Managing and processing large volumes of data from multiple devices
 Interoperability: Lack of standardization among different IoT devices and platforms
 Privacy: Handling user data responsibly and ensuring privacy compliance

Future of AIoT

 Growth with 5G:

o 5G technology will enhance AIoT by enabling faster data transfer
o Lower latency will improve real-time decision-making
 Revolutionizing Industries:
o AIoT will improve supply chains
o Increase operational efficiency
o Enable smarter and faster decision-making
 Exponential Data Growth:
o Increasing number of connected devices will generate massive data
o AI will help manage, process, and analyze this data efficiently
 Future Projection:
o By 2025, the number of IoT devices is expected to reach 42 billion
o AI will play a crucial role in handling and analyzing this growing data volume