Unit 1
Definition of Artificial Intelligence
Artificial Intelligence is the science and engineering of creating machines or software capable of
performing tasks that typically require human intelligence, such as perception, reasoning, learning,
decision-making, and natural language understanding. It encompasses algorithms and models that
enable computers to simulate cognitive functions autonomously.
AI is broadly categorized into rule-based systems (pre-programmed logic) and learning-based
systems (data-driven adaptation). Core attributes include autonomy (operating without constant
supervision), adaptability (improving over time), and scalability (handling increasing complexity).
Example: A self-driving car uses AI to perceive roads via cameras (perception), predict pedestrian
movements (reasoning), and adjust speed (decision-making), all in real-time without human input.
History and Evolution of AI
AI's timeline spans decades, marked by optimism, setbacks ("AI winters"), and revivals:
• 1940s-1950s Foundations: Alan Turing's 1950 paper "Computing Machinery and
Intelligence" introduced the Turing Test. Cybernetics pioneers like Norbert Wiener explored
feedback loops.
• 1956 Dartmouth Conference: John McCarthy coined "Artificial Intelligence"; birth of AI as
a field.
• 1960s-1970s Golden Age: Early successes like ELIZA (chatbot simulating therapy) and
SHRDLU (block manipulation).
• 1970s-1980s AI Winters: Overhyped promises led to funding cuts due to computational
limits.
• 1990s-2000s Revival: Machine learning advances; IBM Deep Blue beat chess champion
Garry Kasparov (1997).
• 2010s-Present Boom: Big data, GPUs, and deep learning fueled breakthroughs like AlphaGo
(2016) defeating Go world champion and GPT models revolutionizing language.
Key Drivers of Modern AI: Moore's Law (computing power doubling every 18 months), vast
datasets (e.g., internet-scale text), and open-source frameworks like TensorFlow.
Example: AlphaFold (2020) solved protein folding—a 50-year biology challenge—using deep
learning trained on 170,000 protein structures.
Strong AI vs Weak AI
Weak AI (Narrow AI): Task-specific systems lacking general intelligence or consciousness. They
excel in defined domains but fail outside them.
• Characteristics: Optimized for efficiency; no self-awareness; brittle to novel scenarios.
• Example: Siri handles voice queries but can't drive a car.
Strong AI (Artificial General Intelligence - AGI): Hypothetical systems with human-like
versatility, understanding, and creativity across any intellectual task, potentially including
consciousness.
• Characteristics: Transfer learning (skills apply broadly); common-sense reasoning;
emotional simulation.
• Challenges: Philosophical debates on consciousness (e.g., Chinese Room argument by John
Searle).
• Example: Hypothetical AGI could learn chess, compose symphonies, and negotiate treaties
from scratch.
Current AI is 99% weak; strong AI remains elusive, with estimates ranging from 2030 (optimists like
Ray Kurzweil) to never (skeptics).
Narrow AI, General AI, and Super AI
Type Definition Current Status Example/Application
Specialized for one task/domain; Image classifiers (e.g., face
Narrow AI Widely deployed
dominant today. unlock)
General AI Human-level across diverse tasks; Hypothetical: Solve puzzles,
Research stage
(AGI) seamless skill transfer. code, cook
Surpasses humans in all
Super AI Sci-fi like Skynet; existential
intellectual/economic activities; self- Theoretical/speculative
(ASI) risks
improving.
Narrow AI scales via data; AGI requires architectures mimicking brain modularity (e.g., transformer
models evolving toward it); Super AI invokes singularity (exponential self-improvement).
Example: Narrow: Recommendation engines (Netflix analyzes viewing habits for suggestions);
Super: Hypothetical system inventing fusion energy overnight.
Goals and Capabilities of AI Systems
Primary Goals:
1. Problem-Solving: Automate complex decisions (e.g., route optimization).
2. Perception: Interpret sensory data (vision, speech).
3. Learning: Improve from experience (supervised/unsupervised).
4. Reasoning: Infer knowledge (deductive/inductive).
5. Natural Interaction: Human-like communication.
Capabilities Breakdown:
• Reactive: No memory (e.g., Deep Blue).
• Limited Memory: Learns from past data (e.g., self-driving cars).
• Theory of Mind: Understands emotions (emerging).
• Self-Aware: Consciousness (future).
Example: GPS navigation AI goals: Minimize time/fuel (optimization capability), adapt to traffic
(learning), reroute verbally (interaction).
AI Subfields Overview
Subfield Definition Key Techniques Example
Fraud detection
Algorithms learning patterns from Regression, clustering,
Machine Learning (ML) via anomaly
data without explicit rules. SVMs
spotting
ML subset using multi-layered neural CNNs (images), RNNs ChatGPT
Deep Learning (DL)
networks for hierarchical features. (sequences) generating essays
Processing human language for Tokenization, Google Translate
NLP
understanding/generation. transformers real-time speech
Interpreting visual data Object detection Self-driving car
Computer Vision
(images/videos). (YOLO), segmentation lane detection
MYCIN
Rule-based mimics of human experts Forward/backward
Expert Systems diagnosing
in narrow domains. chaining
infections
AI for physical agents integrating SLAM (mapping), path Boston Dynamics
Robotics
perception/action. planning robots dancing
Knowledge
Structuring info for inference Semantic nets, Legal reasoning
Representation &
(ontologies, logic). PROLOG systems
Reasoning (KR&R)
Example: Autonomous drone uses CV (obstacle detection), robotics (flight control), ML (path
learning).
Multidisciplinary Applications
AI in Engineering
Manufacturing: Predictive maintenance analyzes IoT sensors to predict failures, reducing downtime
50%.
Automation: Cobots (collaborative robots) learn human workflows via imitation learning.
Design Optimization: Generative design explores millions of iterations for optimal structures.
Example: Siemens uses AI to simulate factory lines, cutting design time from months to days.
AI in Healthcare
Diagnosis: ML models outperform doctors in specific tasks (e.g., diabetic retinopathy detection).
Medical Imaging: CNNs segment tumors with 95% accuracy.
Example: PathAI assists pathologists, reducing errors by 30% in cancer biopsies.
AI in Finance
Fraud Detection: Real-time anomaly detection flags 99% of threats.
Trading Systems: High-frequency trading executes microsecond decisions.
Example: JPMorgan's LOXM optimizes trades, saving billions in costs.
AI in Education
Adaptive platforms tailor content to learner pace, boosting retention 30%.
Example: Carnegie Learning math tutor adjusts difficulty dynamically.
AI in Management and Business Analytics
Forecasts demand via time-series models; optimizes supply chains.
Example: Walmart uses AI for inventory, preventing 1B+ in stockouts.
AI in Engineering and Management
Role in Decision-Making
AI augments via prescriptive analytics (recommends actions) over predictive.
Example: Reinforcement learning in energy grids balances supply-demand optimally.
AI-Driven Optimization Problems
Solves NP-hard issues like traveling salesman via genetic algorithms/evolutionary computing.
Example: Airlines use AI for crew scheduling, saving $100M+ yearly.
Intelligent Systems in Enterprise Management
ERP systems with AI (e.g., SAP Leonardo) enable real-time anomaly alerts.
Example: Digital twins mirror factories for virtual testing.
Ethical Considerations
Ethical considerations in AI address the moral, societal, and practical implications of deploying
intelligent systems that impact human lives profoundly. These issues arise because AI can amplify
biases, invade privacy, and make opaque decisions at scale, demanding frameworks for responsible
development.
Privacy and Data Security
Definition: Privacy concerns the right to control personal information, while data security protects it
from unauthorized access. AI exacerbates risks by requiring massive datasets often containing
sensitive details like health records or biometrics.
AI systems train on vast personal data, creating vulnerabilities to breaches, surveillance, or misuse.
Techniques like data minimization (collect only essentials) and encryption mitigate risks, but
challenges persist with cloud storage and third-party sharing.
Deep Dive: Differential privacy adds statistical noise to datasets, preventing individual identification
while preserving utility. Federated learning trains models on user devices without centralizing raw
data.
Example: The 2018 Cambridge Analytica scandal used AI to profile 87 million Facebook users for
targeted political ads, eroding trust and sparking GDPR regulations mandating consent and breach
notifications within 72 hours.
Real-World Impact: Facial recognition apps like Clearview AI scraped billions of public photos
without permission, leading to bans in cities like San Francisco.
Transparency and Explainability
Definition: Transparency reveals how AI systems operate internally; explainability clarifies specific
decisions (e.g., "why this loan denial?"). Many deep learning models are "black boxes" due to billions
of parameters, hindering oversight.
Lack of transparency erodes user trust and complicates debugging or regulatory compliance.
Explainable AI (XAI) techniques include feature importance scores (e.g., SHAP values showing
variable contributions) and surrogate models (simpler approximations of complex ones).
Deep Dive: Levels of explainability range from intrinsic (inherently interpretable like decision trees)
to post-hoc (LIME generates local explanations). The EU's "right to explanation" under GDPR pushes
for this in high-stakes decisions.
Example: In healthcare, IBM Watson's oncology recommendations lacked explainability, leading
doctors to override 87% of suggestions due to unverifiable reasoning paths.
Real-World Impact: Aviation uses transparent rule-based systems for safety; opaque neural nets
struggle here, as seen in Boeing 737 MAX controversies indirectly tied to automated decision opacity.
Accountability in AI Systems
Definition: Accountability assigns responsibility for AI outcomes—developers for flaws, deployers
for misuse, or users for inputs? It ensures redress for harms like wrongful arrests from biased facial
recognition.
Challenges include diffused responsibility in supply chains (e.g., open-source models) and
autonomous actions (e.g., self-driving car accidents). Legal frameworks propose "accountability by
design," auditing trails, and liability insurance.
Deep Dive: Chain-of-responsibility models trace errors: training data providers liable for biases,
engineers for bugs, executives for deployment. Algorithmic impact assessments (like environmental
ones) evaluate risks pre-launch.
Example: Uber's 2018 self-driving car fatality highlighted gaps—neither the company nor safety
driver faced clear charges, prompting NHTSA probes into accountability standards.
Real-World Impact: The EU AI Act (2024) classifies AI by risk: prohibited (social scoring), high-risk
(hiring tools) requires human oversight and audits, with fines up to €35M for violations.
Social Impact of AI Adoption
Definition: Social impact encompasses job displacement, inequality amplification, and cultural shifts
from widespread AI integration. Automation favors capital over labor, widening wealth gaps.
AI displaces routine jobs (e.g., manufacturing) but creates high-skill ones (e.g., prompt engineering).
Inequality arises from uneven access—tech hubs thrive while rural areas lag. Broader effects include
misinformation spread via deepfakes and eroded human skills from over-reliance.
Deep Dive: Oxford studies estimate 47% of US jobs at risk, but McKinsey predicts net job growth by
2030 via augmentation. Universal basic income (UBI) pilots address displacement; diversity in
training data counters biases.
Example: Amazon scrapped a hiring AI in 2018 after it discriminated against women, trained on
male-dominated resumes—social bias perpetuated digitally.
Real-World Impact: AI in lending (e.g., Apple's credit card glitch charging wives less) revealed
gender biases, sparking FTC investigations; in warfare, autonomous drones raise "killer robot" ethics
bans pushed by 30+ countries.
Ethical Issue Key Risks Mitigation Strategies Example Case
Data breaches, Anonymization, federated Cambridge
Privacy surveillance learning Analytica
Black-box decisions, bias XAI tools (SHAP, LIME), Watson Health
Transparency hiding audits failures
Diffused liability, no Audits, EU AI Act,
Accountability redress insurance Uber AV fatality
Job loss, inequality, Amazon hiring
Social Impact deepfakes UBI, reskilling, bias audits bias