RV Institute of Technology and
Management®
Chapter 8 – Ethical Considerations in AI
8.1 Ethics of AI deals with the moral obligations and duties specific to robots and other artificially
intelligent beings. This field is divided into two main groups:
• Roboethics: This considers the moral behavior of humans as they design, construct, use, and treat
artificially intelligent beings.
• Machine ethics: This manages the moral behavior of artificial moral agents (AMAs) that are crucial for
developing artificial general intelligence (AGIs) within existing legal and social frameworks.
Key Concerns in AI Ethics Several critical questions arise regarding AI's capabilities and implications:
• Can AI make human-like decisions, and are these decisions justified or fair?
• Is AI trustworthy if it makes human-like decisions?
• What happens if the data used for AI training is unintentionally wrong or biased?
Features of an Ethical AI System An ideal AI system should be:
1. Explainable: It must be able to explain how decisions were made, especially in regulated
industries like finance, where "black box AI" (unexplainable decision-making) is a challenge.
2. Monitorable: Its performance and decisions should be continuously traceable.
3. Reproducible: Its results should be consistent and repeatable
4. Secure: It must protect data and prevent malicious access or misuse.
5. Human-centered: Designed with human well-being and control in mind.
6. Unbiased: It should not produce results systematically prejudiced towards any group.
7. Justifiable: Its decisions should be defensible and align with ethical standards.
Dangers and Risks of AI While super-intelligent AI is generally believed to be unlikely to exhibit human
emotions, risks can arise under specific conditions:
• Programmed for devastation: If AI systems, like autonomous weapons, are programmed to kill and fall
into the wrong hands, they could lead to AI wars and mass casualties. Such systems can be difficult to
"turn off", leading to a rapid loss of human control.
• Unclear goals: If AI is programmed for a beneficial task but with vaguely defined goals (e.g., a self-
driving car instructed to reach a destination "as soon as possible" without considering safety), it could
cause significant harm . Continuous monitoring and control are crucial.
RV Institute of Technology and
Management®
Additional Threats/Risks:
• Job automation: AI robots are increasingly replacing human jobs, especially predictable and repetitive
tasks in sectors like retail, hospitality and administrative roles, potentially affecting millions. However, it
is also predicted that AI will create more new jobs than it displaces, leading to a "paradigm shift" similar
to the Industrial Revolution . Jobs in creative fields, management (with AI tools as assistance), and tech
(programmers, data scientists) are expected to grow .
• Biased algorithms: AI systems can produce biased results if trained with biased or skewed data (e.g.,
voice assistants trained predominantly on American accents struggling with Indian accents, or hiring
algorithms biased against women due to historical data) . Humans developing the AI systems are also a
source of bias.
• Privacy erosion: Smart systems collect vast amounts of user data, raising concerns about privacy and
how this data is used. Facial recognition, for example, can capture detailed personal information.
• Unreliability and manipulation: AI can create fake news, filter bubbles, manipulate images (e.g.,
Deepfake), clone voices, and generate highly realistic faces of non-existent people, making it difficult to
discern reality .
Towards Ethical and Trustworthy AI Efforts are underway globally to ensure ethical AI development:
• Regulation: Countries like the EU and Australia are developing legal frameworks and ethics
frameworks to guide responsible AI design, development, and use. These frameworks aim to ban AI
systems that manipulate human behavior or threaten rights.
• Organizational engagement: Companies are building trust through cultural and educational
programmes, risk assessments, and third-party audits.
• Rights and activist groups: These groups challenge discourse and advocate for human-centric AI
systems, questioning the necessity of technology and ensuring benefits outweigh harm.
• Data privacy: Measures include encrypting data, restricting algorithms from learning outside their
datasets, using secure computation, and preventing reverse engineering of AI models .
• Diverse teams: Building diverse teams of data scientists, business leaders, lawyers, accountants,
sociologists, and ethicists can help mitigate bias .
8.1.6 Importance of Ethical AI Ethical AI is crucial because biased or incorrect AI outputs in critical
areas (medicine, law enforcement, recruiting, self-driving vehicles) can:
• Erode people's privacy through misuse of data.
RV Institute of Technology and
Management®
• Lead to decisions that are difficult to understand.
• Complicate assigning liability for damages when harm occurs .
Guiding Principles for Trustworthy AI (European Commission High-Level Expert Group):
1. Human autonomy: AI systems should support human decision-making.
2. Technical robustness and safety: Systems must be robust, safe, and have fallback plans.
3. Data privacy and governance: Protect user data, ensure quality, integrity, and responsible
collection/storage.
4. Transparency: Be clear and explicable about underlying data and models.
5. Fairness: Trained with diverse, non-discriminatory data to avoid bias.
6. Societal and environmental well-being: Designed to benefit everyone and be sustainable.
7. Accountability: Someone must be responsible for decisions.
8.2 AI as a Service (AIaaS)
Definition: AIaaS refers to companies utilising off-the-shelf AI tools to implement and scale AI
techniques at a significantly lower cost compared to building a complete in-house AI system . It typically
leverages cloud computing to provide AI services over a network .
8.2.1 Factors Triggering Growth:
• Availability of cloud platforms: Affordable options for enterprise data management.
• Improved data storage: Cheaper and more reliable technologies.
• Massive data generation: Streaming devices and IOT technologies produce large amounts of data that
can be analysed for competitive advantage.
• Automated data management: Advancements in analytics and Business Intelligence (BI) products.
• High costs of in-house AI: Expensive infrastructure, lack of trained programmers, and high charges for
AI implementation.
• Insufficient data: Many companies lack enough data to train AI models themselves.
8.2.2 Benefits of AIaaS:
• Focus on core business: Companies can concentrate on their primary activities instead of becoming AI
experts.
• Reduced costs: Significantly lowers operational and infrastructural expenses, allowing even small
companies to use machine learning without huge investments.
• Minimized investment risk: Lower financial exposure compared to in-house development.
• Short deployment time: Solutions can be deployed quickly.
• Accessibility: Pre-built algorithms and drag-and-drop tools make complex data management accessible
to small businesses.
• On-demand services: Provides a competitive edge through readily available services.
• Enhanced data analysis: Increases benefits from analysing data trends.
• Flexibility and transparency: Offers cost-effective, flexible, and transparent solutions.
RV Institute of Technology and
Management®
• No need for dedicated AI teams: Companies don't need to maintain and train internal AI professionals.
• Scalability: Allows companies to start with smaller AI projects and easily scale up or down services
based on demand.
8.2.3 Challenges of AIaaS:
• Reduced security: Sharing data with third-party vendors raises data security concerns during storage
and transmission.
• Reliance: Organizations become dependent on third parties for their information needs, potentially
causing delays.
• Reduced transparency: AIaaS often operates like a "black box" where organizations don't see the
internal workings or algorithms used.
• Data governance: Internal data policies might restrict storing data on third-party clouds, limiting AIaaS
adoption.
8.2.4 Vendors of AIaaS:
• Amazon Web Services (AWS)
• Microsoft Azure (e.g., Azure Cognitive Services for computer vision/text extraction, Azure Bot
Service).
• Google Cloud Platform (GCP) (e.g., Google Cloud ML for big data, AutoML for custom ML models) .
• Monkey Learn: Simplifies text analysis with intuitive, no-code tools for sentiment detection, keyword
extraction, and topic classification.
• IBM Watson: Offers a suite of AI tools, including Watson Assistant for virtual assistants and Watson
Natural Language Understand for text analysis.
• AIPaaS (AI Platform-as-a-Service) is an end-to-end cloud solution allowing businesses to use services
on a pay-per-use basis, often via third-party APIs.
8.4 Recent Trends in AI
• Collaborative Systems
• Machines Assisting Humans
• Algorithmic Game Theory and Computational Social Choice
• Multi-Agent Reinforcement Learning (MARL)
• Neuromorphic Computing
9.1 Expert System
Definition: An Expert System is an AI program designed to simulate the decision-making ability of a
human expert.
9.1.2Characteristics:
RV Institute of Technology and
Management®
• High performance
• Reliable
• Highly responsive
• Understandable
• Capable of solving complex problems
• Effective
• Delivers correct advice
9.1.3 Components:
• Knowledge Base: Stores facts and rules, representing domain-specific knowledge.
• Inference Engine: Applies logical rules to the knowledge base to derive new information and answer
queries.
• User Interface: Enables communication between the user and the system.
9.1.4 Participants in Development:
• Domain Expert: Provides the domain knowledge.
• Knowledge Engineer: Extracts, structures, and encodes the domain expert's knowledge into the system.
• End-user: The person who uses the expert system.
9.1.5 Capabilities:
• Advising
• Demonstrating
• Diagnosing
• Explaining
• Interpreting input
• Predicting results
• Justifying conclusions
• Suggesting alternative options
9.1.6 Advantages:
• Availability of solutions to complex problems.
• Reduces error rate.
• Works with incomplete/uncertain data.
• Increases consistency in decision-making.
• Reduces human risk (e.g., in dangerous environments).
• Faster decision-making.
• Handles complex information.
• No emotions or bias.
• Offers multiple solutions.
9.1.7 Limitations:
RV Institute of Technology and
Management®
• Expensive development and maintenance.
• Requires knowledge acquisition from human experts, which can be difficult and time-consuming.
• Limited to specific domains.
• Lack of common sense.
• Cannot learn on its own (requires explicit programming for new knowledge).
• Difficult to update.
• May not perform well with contradictory rules.
• Not suitable for complex problems with varying conditions.
9.1.8 Applications:
• Design and Manufacturing
• Knowledge Publishing
• Financial Management
• Diagnosis and Troubleshooting
• Planning and Scheduling
• Education
• Data Mining
9.1.9 Expert System Technology:
• Shells: Software packages that provide the basic framework for building an expert system, including an
empty knowledge base and inference engine.
• Languages: AI programming languages like Lisp and Prolog are used for developing expert systems.
9.1.10 Development of Expert Systems: The development process typically involves:
1. Problem identification: Defining the problem the expert system will solve.
2. Knowledge acquisition: Gathering knowledge from domain experts.
3. Knowledge representation: Encoding the acquired knowledge into the knowledge base.
4. Inference engine design: Developing the reasoning mechanism.
5. Testing and refinement: Evaluating the system and improving its performance.
9.2 Internet of Things (IOT)
Definition: The Internet of Things (IOT) refers to a network of physical objects embedded with sensors,
software, and other technologies for the purpose of connecting and exchanging data with other devices
and systems over the internet.
9.2.1 Examples of Applications:
• Smart homes (smart locks, switches, refrigerators, washing machines).
• Smart cities (optimizing traffic, waste management).
RV Institute of Technology and
Management®
• Healthcare (remote patient monitoring, smart wearables).
• Industrial automation (predictive maintenance, supply chain management).
• Agriculture (monitoring crops, soil conditions).
9.2.2 IOT Products: smart devices like smart locks, switches, and appliances are mentioned as examples
within smart homes.
9.2.3 Challenges: Challenges related to data security and privacy are general concerns for interconnected
systems.
9.2.4 Sensors: Sensors are crucial components of IOT devices, enabling them to collect data from their
environment. The source lists "Sensors" as a subsection for IOT but doesn't elaborate further on the types
or functions of sensors within this specific section.
9.3 Artificial Intelligence of Things (AIOT)
Definition: Artificial Intelligence of Things (AIOT) is a synergy that combines AI technologies with
IOT infrastructure. It involves embedding AI capabilities into IOT devices and integrating AI
algorithms with IOT data.
9.3.1 How Does AIOT Work? : AIOT works by connecting AI algorithms to IOT sensors and devices.
IOT devices collect vast amounts of data, and AI algorithms then process and analyze this data to enable
intelligent decision-making and automation. This allows devices to learn from the data they collect and
react intelligently without constant human intervention.
9.3.2 Where Does AI Unlock IOT?: AI unlocks the full potential of IOT by:
• Enabling intelligent insights: Processing massive datasets from IOT devices to extract actionable
information.
• Automating decision-making: Allowing devices to make autonomous decisions based on data analysis.
• Improving efficiency: Optimizing operations and resource utilization.
• Enhancing user experience: Providing more personalized and responsive services.
9.3.3 Applications and Examples of AIOT:
• Smart Homes: AIOT devices learn user behavior patterns to automate functions, save time, and
optimize resource usage (e.g., smart thermostats that predict when to turn on/off based on weather).
• Healthcare: AIOT can enable remote patient monitoring, predictive diagnostics, and personalized
treatment plans using data from wearables and medical devices.
• Manufacturing: AIOT devices monitor equipment and production lines to predict maintenance needs
and detect defects in real-time.
• Retail: AIOT can optimize inventory management and enhance customer experience through
personalized recommendations.
RV Institute of Technology and
Management®
• Autonomous Vehicles: Self-driving cars rely on AIOT to process sensor data (LIDAR, cameras) for
navigation, object detection, and decision-making.
• Predictive Maintenance: Devices with AI capabilities can predict equipment failures before they occur,
reducing downtime.
9.3.4 Benefits and Challenges of AIOT:
Benefits:
• Enhanced efficiency and productivity: Automation of tasks and intelligent decision-making.
• Cost savings: Optimized resource use and predictive maintenance.
• Improved decision-making: Data-driven insights.
• Scalability: Ability to handle vast amounts of data and devices.
• New business opportunities: Creation of innovative services and products.
Challenges:
• Security and privacy: Protecting vast amounts of sensitive data collected by IOT devices from cyber
threats.
• Interoperability: Ensuring different devices and platforms can communicate seamlessly.
• Complexity: Managing and integrating numerous devices, data streams, and AI algorithms.
• Data quality: Ensuring the accuracy and reliability of data collected from sensors.
• Ethical concerns: Issues related to surveillance, bias, and accountability.
ent infinite loops.
• Working: Nodes at the depth limit are treated as leaf nodes.
• Advantages: Memory efficient, faster execution, terminates in finite time.
• Disadvantages: Incomplete (if solution is beyond limit), not optimal (may not find best solution even if
ℓ > d).
3.6.3 Breadth First Search (BFS)
• Strategy: Traverses the graph layer-wise (breadth-wise)....
• Working: Starts from root, explores all neighbour nodes at the present depth before moving to the next
level.
• Implementation: Uses a FIFO queue.
• Advantages: Finds a solution if it exists, finds the optimal solution if step cost is uniform8182.
• Disadvantages: High time and space complexity.
• Applications: Crawlers in Search Engines, GPS Navigation Systems, Finding Shortest Path for
Unweighted Graphs
3.6.4 Uniform Cost Search (UCS)
RV Institute of Technology and
Management®
• Strategy: Finds an optimal solution when step costs are not the same
• Working: Computes the cumulative cost to expand each node from root to goal.
• Implementation: Uses a priority queue.
• Advantages: Finds optimal solution (by least cost), complete if states are finite and no zero-weight
loops, optimal if no negative cost.
• Disadvantages: May get stuck in infinite loops (considers only cost, not number of steps)84.
• Equivalence to BFS: UCS is equivalent to BFS if all path costs are the same85.
3.6.5 Iterative Deepening Depth-First Search (IDDFS)
• Combination: Combines DFS (memory efficiency) and BFS (fast search)8687.
• Working: Gradually increases the depth limit for DLS until the goal is found8788.
• Advantages: Combines benefits of BFS and DFS, complete if branching factor is finite, optimal if path
cost is non-decreasing86.
• Applications: Widely used when search space is large and goal node depth is unknown86.
3.6.6 Bidirectional Search
• Strategy: Searches simultaneously from both initial and goal states until they meet8789.
3.7 Informed Search Algorithms (Heuristic Search)
• Contain information about how far from the goal, path cost, etc.90.
• This knowledge helps agents explore less search space and reach the goal more efficiently90.
• Uses a heuristic function that estimates how close the agent is from the goal90.
• May not always give the best solution but guarantees a good solution in reasonable time90.
3.7.2 Best-First Search Algorithm (Greedy Search)
• Combination: Combines DFS and BFS91.
• Strategy: Always selects the path that appears best at that moment91.
• Working: Uses the heuristic function to choose the most promising node (closest to the goal)91.
• Implementation: Uses a priority queue91.
--------------------------------------------------------------------------------
Chapter 4: Knowledge Representation
4.1 Introduction
• To build AI systems with conscience, knowledge needs to be inculcated in them92.
• Knowledge and Intelligence: Knowledge of the real world is crucial for intelligence and creating
intelligent AI agents92. Without knowledge, decision-makers cannot sense the environment accurately or
make appropriate decisions93.
4.2 Knowledge Representation (KR)
• Focus: How AI agents think and how their thinking contributes to intelligent behaviour94.
• Purpose: To represent real-world information in a computer-understandable way so AI can solve
complex problems (e.g., medical diagnosis, natural language communication)94.
• KR involves modelling intelligent behaviour for an agent by representing beliefs, intentions, and
judgments95.
RV Institute of Technology and
Management®
• It goes beyond merely storing data; it facilitates machines to learn from knowledge and experiences to
behave intelligently95.
4.2.1 What Knowledge Needs to be Represented? To make an AI system truly intelligent, it needs to
incorporate human-like intuition, intentions, prejudices, beliefs, judgments, common sense, and facts96:
• Object: Information and facts about relevant objects (e.g., vehicles, roads in a self-driving car)96.
• Events: Information and facts about actions occurring in the real world (e.g., applying brakes)97.
• Performance: How actions are performed (behaviour)97.
• Meta-knowledge: Knowledge about knowledge (what we know)97.
• Facts: Truths about the real world97.
• Knowledge base: Stores a group of technical sentences97.
4.2.2 What is Knowledge? Knowledge is the basic element for logical understanding, gained by
experience98. Five types of knowledge:
• Meta knowledge: Knowledge about knowledge98.
• Heuristic knowledge: Knowledge about a specific topic, often from experts, based on experience (rule
of thumb)98.
• Procedural knowledge (imperative knowledge): Information on "how to" achieve something,
including rules, strategies, procedures99.
• Declarative knowledge: Information about an object, describing concepts, facts, and attributes
(descriptive knowledge)99.
• Structural knowledge: Basic knowledge for complex problems, describing relationships between
concepts/objects (e.g., "kind of," "part of")100.
4.2.3 What is Logic?
• Logic: Main component of knowledge, facilitating drawing conclusions by filtering information100.
• In AI, knowledge is represented using logic, which has three main elements100:
◦ Syntax: Rules specifying how legal sentences in a language are constructed100.
◦ Semantics: Defines the meaning of syntactically correct sentences, relating to the real world101.
◦ Logical Inference: Deducing conclusions from facts/problems using inference algorithms101.
4.2.4 Cycle of Knowledge Representation in AI An AI system's intelligent behaviour involves several
components102:
• Perception: Retrieves data from the environment using sensors, identifies noise sources, checks damage,
defines responses102.
• Learning: Learns from data captured by perception, focusing on self-improvement through knowledge
acquisition, inference, and heuristics103.
• Knowledge Representation and Reasoning (KRR): The core component for human-like intelligence.
Defines what an agent needs to know and how automated reasoning procedures make this knowledge
available103.
• Planning and Execution: Analyses KRR. Planning selects an initial state, enumerates
preconditions/effects, and sequences actions to achieve goals. Execution performs these actions104.
4.2.5 Knowledge Representation Requirements A good KR system must have104105:
• Representational accuracy: Represents all required knowledge104.
RV Institute of Technology and
Management®
• Inferential adequacy: Manipulates structures to produce new knowledge105.
• Inferential efficiency: Directs inference mechanism to generate appropriate results105.
• Acquisitional efficiency: Easily acquires new knowledge automatically105.
4.3 Knowledge-Based Agent (KBA)
• Agents that mimic human knowledge, using knowledge and reasoning to act efficiently and make
appropriate decisions105106.
• Functions: Maintain internal state, deduce reasoning, update knowledge from observations, and take
actions106.
• Store knowledge about surroundings as "sentences" (technical facts)106....
4.3.1 The Architecture of Knowledge-Based Agent
• KBAs contain a knowledge-base (KB) and an inference engine (IE)109.
◦ Knowledge base (KB): Stores real-world facts using sentences in a knowledge representation
language109. The learning element regularly updates it107.
◦ Inference engine (IE): Infers new knowledge from old sentences and adds them to the KB, applying
logical rules (forward or backward chaining)108109.
• Process: KBA perceives environment, input goes to IE, IE interacts with KB for decisions107.
4.3.2 Operations Performed By KBA
• TELL operation: Informs the KB what knowledge it has or needs, and what action was
selected/performed108110.
• ASK operation: Queries the KB for what action to perform110.
• PERFORM operation: Executes the selected action110.
4.3.3 A Generic Knowledge-Based Agent
• Accepts environmental percepts as input, returns an action as output111.
• Maintains a knowledge base (KB) with background knowledge and a time counter111.
• Steps:
◦ TELLS KB what it perceives (using MAKE-PERCEPT-SENTENCE)111.
◦ ASKS KB what action to take (using MAKE-ACTION-QUERY)112.
◦ TELLS KB about the chosen action (using MAKE-ACTION-SENTENCE)112.
4.3.4 Various Levels of Knowledge-Based Agent
• Knowledge Level: Specifies what the agent knows and its goals, used to fix agent behaviour (e.g.,
optimum path from A to B)112113.
• Logical Level: Understands knowledge representation by encoding sentences into different logics (e.g.,
deducing logic for the path)113.
• Implementation Level: Physical representation of logic and knowledge; agent performs actions based
on knowledge from higher levels (e.g., actually moving from A to B)113.
4.3.5 Approaches to Designing a Knowledge-Based Agent
• Declarative Approach: Initializes with an empty KB and gradually adds sentences (facts) until it's
knowledgeable114115.
• Procedural Approach: Directly encodes desired behaviour into the agent as program code (e.g., in
LISP, Prolog)115.
RV Institute of Technology and
Management®
• Hybrid Approach: Combines both, compiling declarative knowledge into efficient procedural code115.
4.4 Types of Knowledge Knowledge can be expressed in a KR system in different ways116:
• Simple Relational Knowledge:
◦ Facts about objects stored systematically using relations (tables) in databases116.
◦ Shortcoming: Little opportunity for inference116.
• Inheritable Knowledge:
◦ Stores data using a hierarchy of classes (generalised to specialised)116.
◦ Shows relation between instance and class (IS-A relation)116.
• Inferential Knowledge:
◦ Represents knowledge in formal logic, used to derive more facts accurately117.
◦ Example: "Diya is a student. All students are bright." -> "Student(Diya)" and "∀
x = Student (x) Bright (x)"117