Nasscom Digital Edge 101
Module -1
Module 1 – AI and Responsible AI
Topic 1: An Overview of Artificial Intelligence
What Is AI?
Artificial Intelligence (AI) refers to machines that can think, learn, and make decisions
like humans. The core objective of AI is to enable computers to analyze data, recognize
complex patterns, and solve intricate problems without requiring continuous human
intervention. AI is deeply embedded in everyday life, powering virtual assistants like
Alexa and Siri, guiding self-driving cars, and revolutionizing medical diagnostics.
The Spectrum of Artificial Intelligence
Artificial Intelligence is not a monolith; it exists on a spectrum of capabilities ranging
from simple reactive programming to theoretical self-awareness.
• Reactive AI: These systems can respond to specific situations but do not
possess the ability to learn from past experiences. A classic example is a chess-
playing AI that calculates moves based only on the current board state.
• Limited Memory AI: These systems learn from past data to make short-term
improvements and informed decisions. Self-driving cars utilize limited memory
AI to observe the speed and direction of other vehicles over time.
• Theory of Mind AI: This is a theoretical, in-development stage of AI where
machines will be able to fully understand human emotions, beliefs, and
interactions.
• Self-Aware AI: The ultimate, theoretical peak of AI development, wherein
machines would possess true human-like consciousness and self-awareness.
An Introduction to Neural Networks
Think of a neural network as a brain-inspired computational system that helps
computers recognize and categorize patterns. Much like the human brain processes
information through interconnected biological neurons, artificial neural networks
process data through layers of artificial neurons.
• Input layer: This layer receives the raw data, such as the pixels of an image or
the characters of a text.
1|Page
• Hidden layers: These intermediate layers process the data using complex
mathematical operations, extracting features and patterns.
• Output layer: This final layer produces the definitive result or prediction, such as
recognizing a cat in a provided image.
Each connection between these neurons has a designated "weight" which determines
how much influence one neuron has on the next. A "bias" is also added to help shift the
values up or down, providing the network with necessary flexibility.
AI in Manufacturing: The Fourth Industrial Revolution
AI is the driving force behind the Fourth Industrial Revolution (often called Industry 4.0).
AI-powered robots are now standard in modern factories, utilized to perform precise
manufacturing tasks, assist human workers, and automate repetitive supply chain
operations. By integrating AI, manufacturing facilities can predict product demand,
optimize inventory management, and forecast potential supply chain disruptions before
they happen.
Trends in AI in Healthcare
The healthcare sector is experiencing a massive transformation due to AI integration. AI
excels at analyzing medical images—such as MRIs and X-rays—to detect diseases with
remarkable accuracy, significantly assisting doctors in early diagnosis. From detecting
cancer anomalies in scans to streamlining patient data management, AI acts as a
critical support tool for medical professionals.
Topic 2: Introduction to Machine Learning
Machine Learning: How It Works and Why We Need It
Machine Learning (ML) is a specialized subset of Artificial Intelligence that enables
machines to learn from data and improve their performance over time without being
explicitly programmed to do so.
• Step 1: The machine is fed large volumes of data, which can be structured or
unstructured.
• Step 2: The system identifies underlying patterns within the data using complex
algorithms.
2|Page
• Step 3: The model undergoes training and testing phases to continuously
improve its accuracy.
• Step 4: Predictions are mathematically refined using techniques like
backpropagation and gradient descent.
The 4 Machine Learning Methods
Understanding ML requires looking at how systems are trained. There are four primary
methodologies:
• Supervised Learning: The system learns from labeled data, meaning every input
has a corresponding, known correct output. It learns patterns to predict
outcomes for new, unseen data. Common algorithms include Linear Regression,
Logistic Regression, Random Forest, Support Vector Machines (SVM), and
Decision Trees.
• Unsupervised Learning: The system is fed unlabeled data and must
independently find hidden patterns, similarities, or clusters without predefined
outputs. Common techniques include Clustering (grouping similar data) and
Dimensionality Reduction (removing irrelevant info). Algorithms include K-Means
Clustering and Principal Component Analysis (PCA).
• Semi-Supervised Learning: A hybrid approach combining a small amount of
labeled data with a large amount of unlabeled data to vastly improve learning
efficiency. The system is trained on the labeled data, then uses that knowledge
to label new data (pseudo-labeling).
• Reinforcement Learning: An action-based method where the system interacts
with an environment, makes decisions, and learns through trial and error. The
system receives a positive reward for correct actions and a penalty for incorrect
ones, continuously adjusting its approach to maximize positive rewards.
Algorithms include Q-Learning and Monte Carlo Tree Search (MCTS).
Artificial Intelligence vs. Machine Learning vs. Deep Learning
While often used interchangeably, these terms represent a hierarchy:
• Artificial Intelligence: The overarching concept of machines mimicking human
intelligence.
• Machine Learning: A subset of AI focused on systems learning from data to
improve over time.
3|Page
• Deep Learning: A further subset of ML that utilizes multi-layered (deep) neural
networks to process vast amounts of data, mimicking the human brain's
architecture.
Applications of Machine Learning
ML helps automate tasks, analyze large datasets, and improve business decision-
making. Real-world applications include:
• Fraud Detection: Identifying suspicious patterns in banking transactions.
• Recommendation Engines: Suggesting products or content based on past user
behavior (e.g., Netflix, Amazon).
• Spam Filtering: Identifying and blocking spam emails using labeled examples.
• Traffic Prediction and Self-Driving Cars: Optimizing routes and helping
autonomous vehicles navigate surroundings.
Top 10 Games that Use Machine Learning
Machine learning is heavily used in gaming, specifically for Dynamic Difficulty
Adjustment, ensuring the game adapts to the player's skill level. The top 10 games
utilizing this include:
1. Resident Evil 4 Remake
2. Left 4 Dead 2
3. Call of Duty: Warzone
4. Celeste
5. Middle-earth: Shadow of Mordor
6. Sekiro: Shadows Die Twice
7. Forza Horizon 5
8. Dead Cells
9. Hades
10. Gears 5
4|Page
Topic 3: Introduction to Deep Learning
An Introduction to Deep Learning
Deep learning enables machines to process vast amounts of data, recognize complex
patterns, and make predictions with exceptionally high accuracy. Unlike traditional ML,
which sometimes struggles with messy data, deep learning efficiently handles both
structured and unstructured data, continuously extracting complex patterns and
improving as the dataset grows.
A Quick Guide to Deep Learning
Deep Learning relies on Neural Networks (NN). Different types of NNs serve different
purposes:
• Feedforward Neural Networks (FNN): A basic model where data moves strictly
in one direction.
• Recurrent Neural Networks (RNN): Designed to process sequential data,
making them ideal for speech recognition.
• Convolutional Neural Networks (CNN): Highly specialized for image
recognition and analysis.
• Generative Adversarial Networks (GANs): Used when the AI needs to generate
entirely new data, such as deepfake images.
Artificial Neural Networks Explained
Artificial Neural Networks (ANNs) replicate biological networks to perform tasks like
pattern recognition and decision-making. Traditional computers excel at numerical
processing but fail at natural tasks like vision; ANNs bridge this gap. When calculating
the output of a single neuron before activation, the network computes a weighted sum.
For example, with inputs $x1$ to $x4$ and weights $w1$ to $w4$:
$z=x1w1+x2w2+x3w3+x4w4+b$ The result then passes through an activation function
to determine if the neuron fires: $a=\sigma(z)$ To learn from its mistakes, the network
utilizes a cost function to calculate the error between the predicted output $\hat{Y}$
and actual output $Y$. A standard cost function is: $J=(Y-\hat{Y})^2$
5|Page
Activation Functions
Activation functions decide which features are important and strip away negative or
useless values. Common mathematical functions used in Deep Learning include:
• Sigmoid: $\sigma(x)=\frac{1}{1+e^{-x}}$
• tanh: $\tanh(x)$
• ReLU (Rectified Linear Unit): $\max(0,x)$
• Leaky ReLU: $\max(0.1x,x)$
An Overview of Convolutional Neural Networks (CNNs)
A CNN is designed to understand images by learning important features like shapes and
textures.
• Convolution: Applying small grids of numbers (filters/kernels) to an image to
detect edges.
• Stride & Padding: Stride is the step size the filter takes across the image.
Padding adds extra space around the image edges to prevent shrinking during
convolution.
• Pooling: Reduces the size of the feature map while retaining critical details (e.g.,
Max Pooling picks the highest value in a small region).
• Fully Connected Layer: Takes all extracted features, flattens them, and makes
the final prediction (e.g., classifying the image as a dog).
Deep Learning Use Cases
• Banking & Finance: Used for algorithmic trading, credit scoring, risk
assessment, and fraud detection.
• Healthcare: Medical imaging analysis, drug discovery, and disease prediction.
• Automotive: Powering autonomous driving systems and predictive vehicle
maintenance.
• Retail & Manufacturing: Inventory management, visual search, quality control,
and supply chain optimization.
6|Page
Infographic: 10 Truths about Deep Learning
1. Efficiency: It automates complex tasks much faster than human capabilities.
2. Adaptability: It inherently learns and improves its accuracy over time.
3. Accuracy: Drastically reduces human errors in decision-making pipelines.
4. Scalability: Capable of processing massively large datasets almost instantly.
5. Data Dependency: It requires massive, high-quality datasets for training; it fails
with poor data.
6. Lack of Transparency: Often referred to as a "black box" because tracing how
the network arrived at a decision is incredibly difficult.
7. Ethical Concerns: Deep learning in autonomous weapons and surveillance
raises severe moral questions.
8. Job Displacement: Unchecked automation threatens to replace human jobs.
9. Computational Cost: Training deep networks requires immense computing
power and energy.
10. Susceptibility to Bias: If trained on biased human data, the network will
replicate and amplify that bias.
Topic 4: ChatGPT
What Is ChatGPT?
Introduced by OpenAI in November 2022, ChatGPT is a highly advanced generative AI
chatbot based on a generative pre-trained transformer architecture. It was trained on
massive amounts of textual data from the internet. ChatGPT utilizes two core
methodologies: Natural Language Processing (NLP) and Large Language Models (LLMs).
The model uses algorithms and statistical analysis to "understand" language, allowing it
to produce text that mimics human speech in conversational settings. It was trained
using Proximal Policy Optimization (PPO), a reinforcement learning algorithm.
Seven Steps of NLP in ChatGPT
The encoder region of the NLP model processes language through seven distinct steps:
1. Segmentation: Breaking text into manageable segments.
2. Tokenization: Converting text into pieces of words called tokens.
7|Page
3. Removal of Stopwords: Stripping out common words that lack meaningful
value.
4. Stemming: A rule-based process removing suffixes to reduce words to their root.
5. Lemmatization: An advanced process reducing words to their base dictionary
form (lemma).
6. Speech Tagging: Assigning tags to tokens indicating their grammatical role in the
sentence.
7. Named Entity Recognition: Classifying proper nouns referring to specific
people or organizations.
Applications of ChatGPT
ChatGPT dramatically impacts business and digital marketing.
• Customer Engagement: Providing 24/7 real-time responses to queries.
• Content Generation: Marketers use it to generate blog posts, social media
posts, video scripts, and email marketing campaigns.
• Data Organization: It assists marketers in managing data, acting as a shortcut
for complex spreadsheet formulas.
• SEO Optimization: Generating relevant keywords, meta descriptions, and link-
building strategies.
• Market Research: Conducting surveys and analyzing customer feedback to
gauge market trends.
Future Applications and Limitations
As ChatGPT evolves, it is expected to integrate seamlessly into voice assistants like
Amazon Alexa and Google Home, providing inclusive, conversational customer service.
However, ChatGPT has distinct limitations. It cannot perform physical tasks, it is not
100% accurate (as it ingests biases and misinformation from the internet), and it cannot
replace human instinct and decision-making. Users must rigorously fact-check,
proofread, and avoid entering sensitive data into the platform.
8|Page
Topic 5: GPT-4 & Evolving Models
Comparing GPT-4 and GPT-3
Released on March 14, 2023, GPT-4 is a massive leap forward from its predecessor,
GPT-3.
• Modality: GPT-3 is strictly unimodal, accepting only text. GPT-4 is multimodal,
capable of accepting and producing both text and image inputs/outputs.
• Complexity: GPT-4 utilizes more data, parameters, and computing power,
allowing it to handle complex tasks requiring profound understanding—areas
where GPT-3 struggled.
• Relevance: GPT-4 has a high degree of steerability, meaning it is far less likely to
generate results irrelevant to the user's input.
GPT-5 Features Explained (Evolving Models)
As the AI landscape rapidly progresses, newer iterations like GPT-5 (and interim models
like GPT-4.5 or GPT-OSS) are pushing the boundaries of generative AI. Current
developments focus on moving beyond manual prompting toward autonomous
"agentic" behavior.
• Advanced Reasoning and Reduced Hallucinations: Newer models feature
adaptive reasoning, spending more compute time on complex mathematical or
logical tasks to drastically reduce factual errors (hallucinations).
• Expanded Context Windows: Models are evolving to handle hundreds of
thousands of tokens simultaneously, allowing users to upload entire codebases
or hundreds of PDF pages in a single prompt.
• Native Multimodality: Future models aim to process text, image, and high-
fidelity audio natively and simultaneously in real-time without latency.
OpenAI Text Classifiers and AI Detection
With the rise of generative text, OpenAI introduced the AI Text Classifier, an algorithm
designed to distinguish AI-generated text from human-written text to combat plagiarism
and misinformation.
9|Page
• Limitations: The classifier is not foolproof. It can mislabel texts, easily be
evaded with minor edits, struggles with non-English text, and requires a
minimum of 1,000 characters to function.
Document classification via Machine Learning is widely used across industries.
• Gmail Spam Classifier: Google trains ML to identify patterns and block 100
million spam messages daily.
• Facebook Hate Speech Detection: Analyzing cultural context and subtle
nuances via NLP to flag rule-violating content without penalizing innocent
speech.
• Great Wolf Lodge Sentiment Classifier (GAIL): Using AI to "read" monthly
customer surveys to gauge positive, negative, or neutral sentiment.
Point-E and DALL-E
• Point-E: An open-source 3D model generator released by OpenAI. It uses a text-
to-image model paired with an image-to-3D model to generate point clouds
(groupings of data points in space) representing 3D shapes. It operates much
faster than traditional 3D rendering and has massive potential for 3D printing and
gaming.
• DALL-E 2: A transformer language model that creates photorealistic images from
text prompts. It uses a text encoder to map a prompt, a "prior" to map semantic
information, and a decoder to stochastically generate the image. It can even
"inpaint," seamlessly editing or replacing specific areas of an existing
photograph.
Topic 6: AI Ethics
An Introduction to AI Ethics
AI Ethics is a multidisciplinary field of study dedicated to understanding how artificial
intelligence should be designed, used, and governed to maximize benefits to society
while minimizing risks and adverse outcomes. Because AI systems simulate human
intelligence and make decisions that affect human lives, they must be bound by strong
moral frameworks. Foundational guidelines, such as those derived from the Belmont
Report (Respect for persons, Beneficence, and Justice), are heavily applied to modern
AI development to ensure technologies "do no harm."
10 | P a g e
AI Ethics and Principles
Establishing an ethical code for AI requires adherence to several core principles:
• Fairness and Bias Prevention: AI systems must treat all individuals equitably.
Because AI learns from historical data, it can easily absorb and amplify human
prejudices. Developers must actively monitor training data to ensure the AI does
not discriminate based on race, gender, or socioeconomic status.
• Transparency and Explainability: AI decisions should not happen in a "black
box." Developers, users, and regulators must be able to understand how an AI
model arrived at a specific conclusion, especially in high-stakes fields like
healthcare or criminal justice.
• Accountability and Governance: When an AI system makes an error, human
operators and parent organizations must take responsibility. Clear governance
structures are required so that blame is not shifted onto the technology itself.
• Privacy and Data Protection: AI systems demand massive data ingestion.
Ethical AI respects user consent, employs anonymization techniques, and
utilizes robust cybersecurity to prevent unauthorized surveillance or data
breaches.
• Human Wellbeing: AI should be designed to augment and support human
intelligence and prosperity, not explicitly to replace humans or compromise
human dignity.
Ethical and Social Implications of AI
The widespread adoption of AI introduces severe social challenges.
• Job Displacement: As AI systems handle complex cognitive and physical tasks
faster than humans, significant segments of the workforce face unemployment.
• Bias in Action: Real-world failures include facial recognition software struggling
to accurately identify people of color, or HR screening algorithms unfairly
favoring male candidates over female ones.
• Liability Constraints: In the event of an autonomous vehicle crash, complex
legal battles ensue over who is liable—the human passenger, the software
developer, or the car manufacturer.
11 | P a g e
Topic 7: Ethical Considerations of Generative AI
Generative AI: Key Risks and Ethical Considerations
Generative AI (like ChatGPT and DALL-E) accelerates these ethical dilemmas due to its
sheer accessibility and rapid adoption. Key risks include:
• Distortion of Reality: GenAI can generate hyper-realistic deepfakes, augmented
reality filters, and fake news. This capability threatens to deceive the public,
distort societal realities, and facilitate severe identity theft.
• Hallucinations: AI models often output completely false information but
present it with absolute confidence, leading to dangerous misinformation if
users do not verify facts.
• Filter Bubbles: Over-reliance on AI can limit decision-making by serving users
highly personalized, specific options based on past behavior, cutting them off
from diverse perspectives.
• Environmental Impact: Training and maintaining massive generative models
requires immense electricity and water for server cooling, exacerbating
environmental concerns.
Key Questions About Generative AI and Copyright
Generative AI threatens traditional intellectual property laws and academic integrity.
• Authorship and Academic Integrity: Tools that bypass traditional plagiarism
detectors make it difficult for academic institutions to verify if a student or an AI
wrote an assignment. This devalues honest academic effort.
• Intellectual Property Violations: Generative models are trained on billions of
images and texts pulled from the internet, often without the original creators'
consent or compensation. When an AI generates an image mimicking a specific
artist's style, it raises fierce legal questions about copyright infringement and
ownership.
Generative AI Regulations and Business Risk
Because of these risks, global regulators (such as the European Union via the AI Act) are
establishing strict laws to govern AI development, outright banning unethical
applications like mass biometric surveillance or social scoring.
12 | P a g e
For businesses, unregulated AI usage poses massive reputational and legal risks. To
adopt generative AI safely, enterprises must:
1. Set Quality Standards: Define criteria for AI outputs regarding accuracy,
fairness, and brand voice.
2. Ensure Transparency: Openly communicate with consumers when content or
customer service is generated by an AI rather than a human.
3. Implement Human Oversight: Ensure human experts always review and fact-
check AI outputs before public release to prevent algorithmic errors from
damaging the brand.
4. Reskill the Workforce: Rather than viewing AI strictly as a tool for layoffs, ethical
companies must commit to retraining and upskilling their employees to work
securely alongside AI technologies.
13 | P a g e