Unit 1 - Introduction to LLM and Prompting

1. Text Generation Models

Text generation models are artificial intelligence (AI) systems trained to generate human-like
text based on patterns learned from large datasets. Their main goal is to predict the next word,
phrase, or sentence given some input (called a prompt).

Key Concepts:

• Language Modeling: At the core, text generation models calculate probabilities of word
sequences.
Example: Given “The cat is on the …”, the model predicts “mat” as more probable than
“carpet” or “sky.”
• Training Data: They learn from massive text datasets — books, websites, articles — to
capture grammar, facts, and writing styles.
• Applications:
o Writing assistance (emails, articles, blogs)
o Chatbots and virtual assistants
o Code generation (e.g., GitHub Copilot)
o Creative writing (poetry, stories, scripts)
o Summarization, translation, Q&A

Evolution:

1. Rule-based Models → Simple predefined templates.

2. Statistical Models (e.g., N-grams, Markov Chains).
3. Neural Networks (RNNs, LSTMs).
4. Transformers (GPT, BERT, T5, LLaMA, etc.) → State-of-the-art.

2. The Magic of Large Language Models (LLMs)

LLMs like GPT-4/5, Claude, Gemini, LLaMA are text generation models at massive scale
(billions of parameters). Their magic comes from a combination of architecture, training data,
and emergent abilities.

Why LLMs Feel “Magical”:

1. Scale Creates Intelligence

o The bigger the model and data, the better it generalizes.
o Beyond grammar, it learns reasoning, logic, and even creativity.
2. Emergent Abilities
 o At smaller sizes, models can’t do advanced tasks.
o Once large enough, LLMs suddenly develop new capabilities like translation,
coding, or multi-step reasoning — abilities never explicitly programmed.
3. Contextual Understanding
o LLMs don’t just predict single words, they capture long-range dependencies,
making text coherent across paragraphs and conversations.
4. Few-shot and Zero-shot Learning
o Give a model just a few examples (few-shot) or even only instructions (zero-
shot), and it can perform new tasks without retraining.
5. Human-like Interaction
o Through fine-tuning and reinforcement learning (RLHF), LLMs respond
conversationally, adapting tone, style, and intent.
6. Versatility
o One model can handle multiple tasks: answering questions, summarizing,
generating code, writing essays, even reasoning about math.

In short:

• Text Generation Models are AI systems for predicting and generating text.
• LLMs bring the magic by scaling this up, resulting in emergent intelligence, versatility,
and human-like interactions.

History and Types of Language Models (LMs)

Language Models (LMs) are systems that learn the probability distribution of words and generate
or predict text. Over the decades, they have evolved from simple statistical models to powerful
Large Language Models (LLMs).

1. Early History: Rule-Based Systems (Pre-1990s)

• Approach: Manually designed grammar rules, dictionaries, and templates.
• Examples: ELIZA (1966), SHRDLU (1970).
• Limitations: Could only handle very narrow conversations; lacked learning capability.

2. Statistical Language Models (1990s – early 2000s)

• Based on probability and statistics.
• N-gram Models:
o Predicts the next word based on the last n-1 words.
o Example: In a trigram model, “The cat is on the” → predicts “mat.”
• Strengths: Simple, interpretable, worked well for speech recognition and machine
translation.
• Limitations:
 o Huge memory requirements for storing probabilities.
o Struggled with long-term dependencies.

3. Neural Network Language Models (2000s – 2010s)

• Introduced neural networks for better generalization.
• Feedforward Neural LMs (2003): First attempt to learn distributed word
representations.
• Recurrent Neural Networks (RNNs): Captured sequential dependencies.
• LSTMs (1997, popularized ~2014): Solved the vanishing gradient problem, handled
longer text better.
• Strengths: Learned semantic meaning, improved translation and speech recognition.
• Limitations: Training was slow, context still limited.

4. The Transformer Era (2017 – present)

• Breakthrough: The Transformer architecture (Vaswani et al., 2017).
• Replaced recurrence with self-attention, enabling models to look at all words in a
sentence simultaneously.
• Key Models:
o BERT (2018) – Great for understanding tasks (classification, Q&A).
o GPT Series (2018–2023) – Specialized in text generation.
o T5, XLNet, RoBERTa, LLaMA – Different optimizations of transformers.
• Strengths: Scalability, parallel training, handling long-range dependencies.
• Limitations: Requires massive data and compute power.

5. Large Language Models (LLMs) (2020s – future)

• Massive transformer-based models with billions/trillions of parameters.
• Trained on internet-scale data.
• Examples: GPT-4/5, Claude, Gemini, LLaMA, Mistral.
• Capabilities:
o Conversational AI
o Code generation
o Summarization & translation
o Multi-modal (text, image, audio) reasoning
• Magic: Emergent abilities (reasoning, creativity, few-shot learning).

Types of Language Models

Based on Technique:

1. Statistical Models
o N-gram models, Hidden Markov Models.
 2. Neural Models
o Feedforward Neural LMs
o RNN-based LMs (RNN, GRU, LSTM)
3. Transformer Models
o Encoder-only (BERT, RoBERTa) – for understanding tasks
o Decoder-only (GPT, LLaMA) – for generation tasks
o Encoder-Decoder (T5, BART) – for translation, summarization

Based on Learning Style:

• Autoregressive Models (e.g., GPT): Predict the next word step by step.
• Masked Language Models (e.g., BERT): Predict missing words in a sentence.
• Sequence-to-Sequence Models (e.g., T5): Translate one sequence into another.

In summary:

• History: Rule-based → Statistical → Neural (RNN/LSTM) → Transformers → LLMs.

• Types: Autoregressive, Masked, Seq-to-Seq, Encoder-only, Decoder-only, Encoder-
Decoder.

Market Size & Growth Projections

• Global Revenue:
o In 2024, the LLM market generated approximately USD 5.62 billion, projected to
soar to USD 35.44 billion by 2030, reflecting a robust CAGR of ~36.9% Grand
View Research.
o Another forecast estimates growth from USD 5.72 billion in 2024 to USD 123.09
billion by 2034, maintaining a similar CAGR of ~35.9% Precedence Research.
• Alternate Estimates:
o Some reports suggest the global LLM market was valued at USD 2.85 billion in
2023, expected to reach USD 30 billion by 2032 (CAGR ~29.9%) Market
Research Future.

Regional Landscape
• North America:
o Holds the largest share, around 32–40% of global revenue (approx. USD 1.97
billion to USD 5.2 billion in 2024) Grand View ResearchGrowth Market
ReportsDataintelo.
o Fueled by presence of OpenAI, Google, Microsoft, IBM, NVIDIA, etc., and
strong funding and infrastructure Growth Market ReportsDataHorizzon Research.
• Asia Pacific:
o Fastest-growing region with CAGR between 35–36%, market around USD 3.1
billion in 2024, projected to expand rapidly Kings ResearchGrowth Market
ReportsConsainsights.
 o India stands out as a high-growth market Grand View ResearchPolaris.
• Europe:
o Holds ~27% share (~USD 3.7 billion in 2024), with focus on ethical AI and
regulatory frameworks Growth Market ReportsConsainsights.
• Other Regions:
o Projections for Middle East, Africa, South America also show substantial
growth—e.g., Asia Pacific from USD 7.66 billion in 2023 to USD 154 billion by
2033; North America from USD 13.4 billion to USD 271 billion in the same
timeframe Consainsights.

Market Drivers & Trends

• Key Growth Drivers:
o Rising demand for automation via chatbots, content generation, virtual assistants,
translation, code generation Precedence ResearchResearch and Markets.
o Continual advancements in hardware (GPUs, TPUs) and model architectures
enabling scalability and efficiency Research and Markets.
o Healthcare and finance sectors particularly strong users of LLMs Precedence
ResearchKings Research.
• Regulatory & Ethical Landscape:
o U.S.: Standards like NIST AI Risk Management Framework, and principles
like the AI Bill of Rights Kings Research.
o Europe: The upcoming EU AI Act categorizes foundation models as high-risk,
with mandates around transparency and bias mitigation Kings Research.

Key Market Players

• Dominant Global Leaders:
o OpenAI, Google (DeepMind – Gemini), Microsoft, Meta (LLaMA/OPT),
Amazon, IBM (Watsonx, Granite), NVIDIA, alongside Chinese giants Baidu,
Alibaba, Tencent Grand View ResearchDataHorizzon ResearchGrowth Market
Reports.
• Emerging Innovators:
o Anthropic (Claude series) – enterprise focus, rapid revenue growth Financial
TimesWikipedia.
o Mistral AI – European unicorn headed toward a €12 billion valuation,
partnerships with Microsoft and others Wikipedia.
o IBM – Enterprise-centric models via Granite and Watsonx platforms
Wikipedia+1.

Industry Highlights & Challenges

• Funding & Valuations:
o Anthropic recently valued at ~USD 183 billion, with runaway funding success
and over USD 5 billion in annualized revenue Financial TimesInvestopedia.
 o OpenAI valuation reportedly at USD 500 billion from secondary market activity
Investopedia.
o Mistral AI securing funding pushes its valuation to €12 billion (~USD 14
billion) Wikipedia.
• Geopolitical Tensions & Access Restrictions:
o Anthropic prohibits Chinese-controlled firms from using Claude, citing regulatory
risks—a move that could impact revenues Tom's Hardware.
o Chinese firms like Alibaba, Baidu, ByteDance launching open-source and low-
cost LLMs to compete Business InsiderReutersAP News.
• Legal Pressure:
o Anthropic agreed to a record USD 1.5 billion settlement with authors for
copyright infringement—highlighting legal risks in training data practices PC
Gamer.

Summary Table
Category Highlights
Global Size & USD 5–6B in 2024; projected USD 35B (2030) to USD 123B (2034);
Growth CAGR ~30–37%
Regional Leaders North America dominant; Asia-Pacific fastest growing; Europe strong
OpenAI, Google/DeepMind, Microsoft, IBM, Amazon, Meta, NVIDIA,
Top Players
Anthropic, Mistral
Growth Drivers Automation, AI apps across sectors, hardware innovation
Challenges Regulation, geopolitical access, legal exposure

Prompting Techniques and Principles

Prompting is the art of designing inputs (prompts) that guide Large Language Models (LLMs)
like GPT, Claude, or Gemini to produce useful, accurate, and creative outputs. Good prompts
turn a generic model into a domain-specific assistant, teacher, or creative partner.

1. Core Principles of Prompting

These are the foundations for effective prompting:

1. Clarity
o Use clear, unambiguous instructions.
o Example: Instead of “Tell me about trees”, say “Explain the process of
photosynthesis in trees in simple terms for high school students.”
2. Specificity
o The more details you include (tone, format, role, context), the better.
o Example: “Write a 200-word summary in bullet points” vs. “Summarize this.”
3. Contextualization
 o Give the model background so it knows what role to take.
o Example: “You are a financial analyst. Compare Tesla and Toyota in terms of
long-term investment potential.”
4. Constraint Setting
o Limit scope to get focused answers.
o Example: “List 3 pros and 3 cons of solar energy.”
5. Iterative Refinement
o Prompting is experimental — refine and re-prompt to improve answers.

2. Common Prompting Techniques

(A) Instruction-Based Prompting

• Directly tell the model what to do.

• Example: “Summarize this text in one paragraph.”

(B) Zero-Shot Prompting

• No examples given — just the task description.

• Example: “Translate ‘How are you?’ into French.” → “Comment ça va ?”

(C) Few-Shot Prompting

• Provide a few examples to guide style or format.

• Example:
• Q: What is the capital of France?
• A: Paris
• Q: What is the capital of Japan?
• A: Tokyo
• Q: What is the capital of Germany?
• A:

(D) Chain-of-Thought Prompting (CoT)

• Encourage the model to think step by step.

• Example: “Solve this math problem step by step: 245 + 378.”

(E) Role-Playing Prompting

• Assign the model a role to shape responses.

• Example: “You are a UX designer. Suggest improvements for a mobile app login page.”

(F) Delimiters for Clarity

• Use symbols (like """, ---, or <text>) to mark important text.

 • Example: “Summarize the following text between triple quotes: """[text]"""”

(G) Prompt Chaining

• Break a complex task into smaller steps, feeding outputs from one step into the next.
• Example: (1) Generate ideas → (2) Rank them → (3) Expand top choice.

(H) Self-Consistency Prompting

• Ask the model to generate multiple reasoning paths and pick the best.
• Example: “Give 3 possible solutions and then choose the most logical one.”

(I) Reflexion Prompting

• Ask the model to critique or review its own output.

• Example: “Write a short essay. Then check for grammar mistakes and rewrite.”

(J) Multi-Modal Prompting (for advanced LLMs)

• Combine text with images, audio, or code inputs to expand capabilities.

3. Advanced Prompting Principles

1. Instruction Hierarchy: Start broad, then add constraints (length, tone, audience).
2. Bias Mitigation: Phrase prompts neutrally to reduce skewed answers.
3. Evaluation and Iteration: Test multiple prompts, compare outputs, refine wording.
4. Grounding with Knowledge: Supply external context (e.g., paste reference material).
5. Safety Principles: Add disclaimers in sensitive prompts (health, legal, finance).

4. Example Prompt Transformations

• Weak Prompt: “Tell me about AI.”
• Strong Prompt: “Act as a computer science professor. Give me a structured 200-word
explanation of AI, covering its definition, history, and applications, in simple language.”

In summary:
Prompting combines clarity, context, and control with techniques like zero-shot, few-shot,
CoT, role-playing, and prompt chaining. The key is iterative refinement — treating prompts as
tools you fine-tune to unlock the best from LLMs.
1. Components of a Prompt
A well-designed prompt usually has four main components:

(A) Instruction

• What you want the model to do.

• Example: “Summarize this article in 100 words.”

(B) Context

• Background or extra details to guide the response.

• Example: “You are a teacher explaining to 10th-grade students.”

(C) Input Data

• The actual text, question, or material to process.

• Example: “Article: [paste text here]”

(D) Output Format

• Desired structure of the response.

• Example: “Answer in bullet points” or “Return JSON format.”

A complete prompt combines all four for precision.

2. Types of Prompts
We can classify prompts in several ways:

(A) Based on Instruction Style

1. Direct Instruction Prompt
o Straightforward task.
o Example: “Translate this sentence into French: ‘How are you?’”
2. Role-based Prompt
o Assigns the model a persona or role.
o Example: “Act as a doctor and explain the side effects of this medicine in simple
terms.”
3. Zero-shot Prompt
o No examples, just instructions.
o Example: “Write a poem about the moon.”
4. Few-shot Prompt
o Provides examples to show desired style/format.
 o Example:
o Q: Capital of Italy? A: Rome
o Q: Capital of Japan? A: Tokyo
o Q: Capital of Germany? A:

(B) Based on Cognitive Approach

1. Chain-of-Thought (CoT) Prompt
o Encourages reasoning step by step.
o Example: “Solve this math problem step by step: 245 + 378.”
2. Self-Reflection Prompt
o Ask the model to evaluate/revise its answer.
o Example: “Write a paragraph, then review it for grammar mistakes and improve
it.”
3. Multi-Step Prompt (Prompt Chaining)
o Breaks tasks into smaller stages.
o Example: (1) Generate 5 essay topics → (2) Choose the best one → (3) Write an
outline.

(C) Based on Output Format

1. Open-Ended Prompt
o Allows creative, wide-ranging answers.
o Example: “Write a story about a time traveler.”
2. Constrained Prompt
o Sets clear limits (word count, format, style).
o Example: “Summarize in 3 bullet points, max 15 words each.”
3. Structured Output Prompt
o Requests specific formats like tables, code, or JSON.
o Example:
“Return a table with columns: Country | Capital | Population.”

(D) Advanced Types

1. Delimiters Prompt
o Uses symbols to separate text for clarity.
o Example: “Summarize the text inside triple quotes: """[text]"""”
2. Multi-modal Prompt (for advanced LLMs)
o Combines text + images/audio/code.
o Example: “Describe this chart (image attached) in 3 sentences.”
3. Comparative Prompt
o Asks for side-by-side analysis.
o Example: “Compare solar and wind energy in terms of cost, efficiency, and
sustainability.”
3. Example of a Complete Prompt
• Weak Prompt: “Tell me about photosynthesis.”
Strong Prompt:

Act as a biology teacher.

Explain the process of photosynthesis to 9th-grade students.
Keep it under 200 words.
Use simple language and bullet points.

Summary:

• Components = Instruction + Context + Input Data + Output Format.

• Types = Instruction-based (direct, role, zero/few-shot), Cognitive-based (CoT, self-
reflection, chaining), Output-based (open-ended, constrained, structured), Advanced
(delimiters, multi-modal, comparative).

1. Strategic Prompting
Strategic prompting means designing prompts with deliberate structure and creativity so
LLMs deliver outputs that are not only correct but also aligned with tone, purpose, and audience.
It goes beyond simple instructions and focuses on shaping how the model “thinks” and
“speaks.”

Two powerful strategies are Personality Prompting and Mix-and-Match Prompting.

2. Personality Prompting
This technique assigns a persona, voice, or identity to the model to shape its style, tone, and
perspective.

Key Ideas:

• Role-Playing:
“You are a motivational coach. Encourage me to prepare for my exam.”
• Tone Control:
“Explain this concept in a humorous way, as if you are a stand-up comedian.”
• Audience Adaptation:
“You are a kindergarten teacher. Explain gravity to 6-year-olds.”
• Voice Simulation:
You can request styles like formal, casual, poetic, journalistic, or scientific.

Personality prompting is powerful for branding, storytelling, teaching, customer support,

and content creation.
3. Mix-and-Match Prompting
Instead of relying on a single prompt style, you combine multiple techniques to get richer and
more accurate results.

Examples:

• Role + Chain of Thought

“As a math tutor, explain step-by-step how to solve this algebra problem.”
• Few-Shot + Role-Based
• You are an HR manager.
• Here are some examples of professional email responses:
• [Example 1, Example 2].
• Now, write a polite rejection email for a job candidate.
• Delimiters + Constrained Output
“Summarize the following text between triple quotes in 3 bullet points, max 10 words
each: """[text]"""”
• Multi-Persona Collaboration
Ask the model to simulate a panel of experts with different personalities, then merge
their views.
“Imagine you are three experts (a doctor, a lawyer, and a teacher). Discuss whether AI
in education is beneficial.”

Mix-and-Match helps handle complex, multi-step, or creative tasks where one prompt style
alone isn’t enough.

4. Example
Weak Prompt:
“Write about climate change.”

Strategic Prompt (Personality + Mix-and-Match):

Act as a climate scientist speaking to high school students.

Step 1: List the main causes of climate change.
Step 2: Explain the effects in simple terms.
Step 3: Suggest 3 easy actions students can take.
Write in an engaging, hopeful tone.

Summary:

• Personality Prompting: Shape tone, role, and perspective (teacher, scientist, comedian,
coach).
• Mix-and-Match Prompting: Blend multiple strategies (role + CoT + examples +
constraints) for better control and creativity.
 • Together, they make prompts strategic, flexible, and highly adaptable for different
contexts.

Unit 2 - The Art of Text Data Generation with GenAI

1. Lists in Prompting
Using lists helps structure outputs, making them organized, scannable, and complete.

Why It Works:

• Forces the model to cover multiple points.

• Reduces the chance of missing information.
• Easier for users to parse results.

Prompting Examples:

• Weak: “Tell me about benefits of exercise.”

• Strong (List):
“List 5 key benefits of regular exercise. Present each as a short bullet point.”

Useful for: summaries, pros/cons, step-by-step instructions, recommendations.

2. Context in Prompting
Context ensures the model understands who, what, and why before answering.

Why It Works:

• Guides tone and depth.

• Reduces irrelevant or overly generic answers.
• Aligns with audience expectations.

Prompting Examples:

• Weak: “Explain blockchain.”

• Strong (Context):
“You are a university professor teaching beginners. Explain blockchain to
undergraduate students in computer science with one real-world example.”

Useful for: education, professional use, industry-specific queries, role-based instructions.

3. Simplification in Prompting
Simplification directs the model to output concise, easy-to-digest responses.

Why It Works:

• Prevents overly technical or verbose answers.

• Matches the knowledge level of the audience.
• Makes complex topics accessible.

Prompting Examples:

• Weak: “Explain quantum computing.”

• Strong (Simplification):
“Explain quantum computing in simple terms, using an analogy a 12-year-old could
understand. Limit to 100 words.”

Useful for: teaching, summaries, executive briefs, non-expert audiences.

4. Combined Example (Lists + Context + Simplification)

Weak Prompt:
“Tell me about renewable energy.”

Strong Prompt:

You are an environmental science teacher.

Explain renewable energy in simple terms for high school students.
List the 4 main types (solar, wind, hydro, biomass) with one example each.
Keep the explanation under 200 words.

Result: Clear, structured, contextualized, and simplified output.

Summary Table
Practice Purpose Prompting Example

Lists Ensure structured, complete answers “List 5 advantages of AI in healthcare.”

Context Tailor response to audience & purpose “As a lawyer, explain GDPR in simple terms.”

Make output concise & easy to “Explain inflation in under 100 words, for
Simplification
understand beginners.”
1. Translation Prompts
Translation prompts ask the model to convert text from one language
to another while preserving meaning.

Types:

• Direct Translation – Word-for-word conversion.

“Translate this sentence into French: ‘The weather is nice
today.’”
• Context-Aware Translation – Maintain tone, idioms, and cultural
fit.
“Translate into Spanish for a business audience: ‘We look forward
to collaborating with you.’”
• Domain-Specific Translation – Adjust to field-specific language.
“Translate into German for a legal contract: ‘The parties agree to
the following terms…’”

Used in global business, localization, cross-cultural communication.

2. Style Transfer Prompts

Style transfer changes the tone, mood, or writing style of the same
content.

Examples:

• Formal to Informal:
Input: “We regret to inform you that your application was not
successful.”
Prompt: “Rewrite this in a friendly, casual tone.”
Output: “Sorry, your application didn’t make it this time, but keep
trying!”
• Creative Reframing:
“Rewrite this news article as a poem.”
 • Brand Consistency:
“Rewrite this product description in the Apple marketing style.”

Useful for marketing, storytelling, customer communication, and

personalization.

3. Contextual Prompts
Contextual prompts provide background information so the LLM
gives an answer that is relevant, role-specific, and aligned with
audience needs.

Examples:

• Role Context:
“You are a medical professional. Explain the symptoms of diabetes
to a patient in simple language.”
• Audience Context:
“Explain blockchain to MBA students with a business case study.”
• Task Context:
“Summarize this article for a news anchor who has 30 seconds to
read it on live TV.”

Useful for teaching, professional settings, customer support,

presentations.

4. Combined Example (Translation + Style Transfer + Context)

Prompt:

You are a professional translator working for a children’s book

publisher.
Translate the following English text into Hindi.
Make sure the style is playful and suitable for 7-year-old kids.
Text: “The little dragon flew across the sky, painting rainbows with its
tail.”
Output: Correct translation, adapted tone, and child-friendly style.

Summary Table
Technique Purpose Example Prompt

Convert across “Translate this into French:

Translation
languages ‘Good morning, everyone.’”

Change tone or writing “Rewrite this paragraph in a

Style Transfer
style humorous tone.”

Contextual Provide background for “As a history teacher, explain

Prompt tailored response WWII causes to 10th graders.”

1. Feature Extraction Prompts

Feature extraction means asking an LLM to pull out specific structured
information from raw, unstructured text (like documents, conversations,
reviews, or articles).

Why It Matters

• Makes unstructured text machine-readable.

• Useful for analytics, classification, and automation.
• Saves time in summarizing large datasets.

Prompting Examples

• Simple Extraction:
“Extract all email addresses from the following text.”
• Attribute Extraction:
“From this product review, extract: (1) Product Name, (2)
Sentiment (positive/negative), (3) Key Features mentioned.”
 • Entity Recognition:
“Identify all people, organizations, and places mentioned in this
paragraph.”

Applications: Chatbot logs, customer reviews, research papers,

resumes, contracts, medical reports.

2. Role Prompting
Role prompting assigns the LLM a specific persona or professional
role to shape the response in tone, knowledge, and perspective.

Why It Matters

• Helps align with audience expectations.

• Makes answers more relevant, contextual, and reliable.
• Enables simulation of expertise.

Prompting Examples

• Professional Roles:
“You are a financial advisor. Explain investment strategies to a
beginner.”
• Creative Roles:
“Act as a storyteller. Rewrite this news event as a bedtime story.”
• Multi-Role Prompting (Panel Simulation):
“Imagine you are a panel of three experts: a doctor, a lawyer, and
an engineer. Discuss the pros and cons of AI in healthcare.”

Applications: education, professional training, customer support,

creative writing, expert simulations.

3. Combined Example (Feature Extraction + Role Prompting)

Prompt:

You are a legal analyst.

From the following contract clause, extract:
1. Parties involved
2. Key obligations
3. Penalty conditions
Summarize in plain English for a non-lawyer.
Text: """[Insert contract text here]"""
Outcome: The LLM extracts structured features while responding in the role of a legal
analyst, making the result clear, contextual, and audience-friendly.

Summary Table
Technique Purpose Example Prompt

Feature Pull structured data from “Extract all dates and names
Extraction text from this report.”

“You are a teacher. Explain

Role Assign identity/persona
photosynthesis to 12-year-
Prompting for tailored response
olds.”

1. Why Prompt Analysis Matters

LLMs behave differently depending on how prompts are written. A
strong prompt can produce accurate, clear, and useful outputs, while a
weak one may give vague, biased, or incomplete results.
Analyzing prompts helps in refining them systematically.

2. Strengths of Good Prompts

Characteristics of Strong Prompts:

• Clarity → Simple, direct, no ambiguity.

“Summarize this article in 3 bullet points, each under 10 words.”
 • Contextualization → Includes role, audience, or scenario.
“You are a career coach. Explain how to prepare for an interview
to a college student.”
• Structured Output → Uses lists, steps, or formatting.
“List 5 challenges in cloud computing and give 1 solution each.”
• Constraints → Specifies word limits, style, or tone.
“Write a 50-word summary of this research paper in layman’s
terms.”
• Examples → Few-shot prompting shows desired style.
“Here are two examples of polite rejection emails. Write a third
one in the same style.”

Strength: They lead to reliable, focused, and high-quality responses.

3. Weaknesses of Poor Prompts

Characteristics of Weak Prompts:

• Vague / Open-Ended → Too broad, model may wander.

“Tell me about AI.” (→ could be history, uses, risks, random
facts).
• Missing Context → No guidance on audience or role.
“Explain cloud computing.” (→ Too technical for a beginner, too
shallow for an expert).
• Overloaded → Cramming multiple tasks in one.
“Explain AI, list all its types, write a poem, and code an example.”
• No Constraints → Output may be too long, off-topic, or
irrelevant.
• Ambiguous Language → Words like “interesting”, “important”,
“detailed” without clear definition confuse the model.

Weakness: They lead to unfocused, inconsistent, or unusable outputs.

4. Example Comparison
Weak Prompt:
“Write about climate change.”

Strong Prompt:

You are a climate scientist.

Explain climate change in 200 words for high school students.
Include: causes, effects, and 2 simple solutions.
Present in bullet points.

5. Summary Table
Aspect Strong Prompt Weak Prompt
Clarity Direct, specific Vague, open-ended
Context Role & audience defined Missing or generic
Structure Lists, steps, formatting Random, unorganized
Constraints Word/format/style specified No boundaries
Examples Shows model the pattern None given

1. Copywriting with Prompts

Copywriting is about persuasion and selling through words. LLMs

can generate ads, product descriptions, taglines, and website copy.

Prompting Techniques:

• Product Descriptions
“Write a 100-word product description for a smartwatch,
highlighting health tracking and style. Tone: professional,
persuasive.”
 • Tagline Creation
“Suggest 5 catchy slogans for an eco-friendly water bottle brand.”
• AIDA Framework (Attention–Interest–Desire–Action)
“Write ad copy for an online course on AI using AIDA style.”

Applications: e-commerce, marketing, landing pages, brand

campaigns.

2. Social Media Content Prompts

Social media writing focuses on brevity, engagement, and

shareability.

Prompting Techniques:

• Posts & Captions

“Write an Instagram caption for a coffee shop promoting a new
pumpkin spice latte. Use a friendly, cozy tone.”
• Hashtags
“Generate 10 trending hashtags for a tech startup’s product
launch.”
• Platform-Specific Styles
“Write a LinkedIn post summarizing an AI research breakthrough
in a professional yet approachable tone.”
“Write a funny 280-character tweet about Mondays.”

Applications: branding, campaigns, influencer marketing,

community engagement.

3. Scriptwriting Prompts

Scripts are used for videos, podcasts, training, and advertisements.

Prompting Techniques:
 • Explainer Video Script
“Write a 2-minute YouTube script explaining how blockchain
works. Audience: beginners. Tone: engaging, simple.”
• Ad Commercial Script
“Write a 30-second ad script for Nike shoes focusing on
empowerment and speed.”
• Dialogue Writing
“Write a short conversation between a teacher and a student about
choosing a career in technology. Tone: motivational.”

Applications: ads, education, entertainment, marketing, training

videos.

4. Combined Example

Task: Promote an online yoga course.

• Copywriting Prompt:
“Write persuasive website copy for an online yoga course.
Highlight benefits: stress relief, flexibility, mental clarity. Use
professional, warm tone.”
• Social Media Prompt:
“Create a 3-line Instagram caption promoting the yoga course.
Add 5 relevant hashtags.”
• Script Prompt:
“Write a 1-minute YouTube ad script introducing the online yoga
course, ending with a strong call-to-action.”

Outcome → Consistent messaging across formats, tailored to different

platforms.
5. Summary Table
Content
Purpose Example Prompt
Type
Persuade, sell, brand “Write ad copy for a smartwatch in
Copywriting
voice 100 words.”
Engage, attract, go “Write a funny tweet about AI in
Social Media
viral 280 characters.”
Inform, entertain, “Write a 30-sec ad script for an
Scripts
convert energy drink.”

Takeaway:

• Copywriting → persuasive storytelling for sales & branding.

• Social Media → short, catchy, and platform-optimized.
• Scripts → structured narratives for ads, videos, and training.
Together, they form a content generation ecosystem powered by
LLMs.

1. What is Personalized Messaging?

Personalized messaging means generating content that adapts to the

individual’s identity, preferences, or context.
Instead of a generic “one-size-fits-all” response, prompts guide the LLM
to tailor tone, style, and content to the recipient.

Example:

• Generic: “Thanks for subscribing!”

• Personalized: “Hi Sarah, thanks for joining our design community!
Since you’re into UI/UX, we’ll send you weekly design tips and
exclusive toolkits.”
2. Principles of Personalized Engagement

Key Elements:

• Audience Awareness → who you’re speaking to.

• Tone Adaptation → casual, formal, playful, supportive.
• Dynamic Content → adjust recommendations, offers, or
examples.
• Interactivity → ask questions, invite replies, create dialogue.

3. Prompting Techniques for Personalized Messaging

a) Name & Context Inclusion

“Write a welcome email to John, a college student interested in AI,

highlighting how our beginner-friendly tutorials will help him.”

b) Segmentation-Based Messaging

“Write three versions of a product promotion: one for teenagers, one for
working professionals, and one for retired users.”

c) Behavior-Driven Messaging

“Generate a personalized discount message for a customer who added

shoes to their cart but didn’t complete the purchase.”

d) Tone & Emotion Adaptation

“Send an encouraging SMS to a student who failed an exam, motivating

them to try again.”

4. Personalized Engagement in Different Domains

• Marketing → Personalized emails, product recommendations,

loyalty rewards.
 • Education → Tailored feedback for different learning styles.
• Customer Support → Empathetic responses based on frustration
level.
• Healthcare → Patient-friendly explanations of test results.

5. Combined Example

Scenario: Fitness App

• Generic Prompt:
“Write a message encouraging users to exercise today.”
• Personalized Prompt:
• You are a fitness coach.
• Write a motivational morning message for Sarah, a 25-year-old
beginner runner.
• Mention that she ran 3 km yesterday and suggest a light stretching
routine today.
• Keep the tone supportive and energetic.

Output: Feels human, relevant, and engaging.

6. Summary Table
Technique Description Example Prompt
Name & Use user details “Write a thank-you note for Alex,
Context for relevance who attended a coding bootcamp.”
“Create 3 versions of an email for
Segmentation Tailor by group students, professionals, and
retirees.”
 Technique Description Example Prompt
Behavior- “Generate a reminder for a
Respond to actions
Based customer who left items in cart.”
Tone & Match user’s state “Send a cheerful note to a nervous
Emotion of mind first-time traveler.”

Takeaway:
Personalized messaging isn’t just about addressing someone by name —
it’s about adapting tone, context, and content to make interactions feel
authentic and engaging.
LLMs + good prompts = scalable personalization for millions of users.
Unit 3 - Learning to Craft Image Data with GenAI

1. What are Diffusion Models?

Diffusion models are a class of generative models that create new data (like images, audio, or
text) by learning to reverse a process of noise addition.

Think of it like this:

• Start with a clear image.

• Add random noise step by step until it becomes pure static.
• Train a model to reverse the noise process → gradually “denoise” static back into a
meaningful image.

During generation, you start from random noise, and the model learns how to “denoise” it into a
completely new, realistic output.

2. Core Idea
• Forward Process (Diffusion): Add noise step by step until data is destroyed.
• Reverse Process (Denoising): Train the model to remove noise step by step and
reconstruct data.

This is why they’re called diffusion (spreading noise) and denoising diffusion probabilistic
models (DDPMs).

3. Why Diffusion Models Matter

• High Quality → They generate images with stunning detail (used in Stable Diffusion,
DALL·E 3, MidJourney).
• Diversity → They can create a wide variety of outputs from the same prompt.
• Controllability → By conditioning on text, sketches, or other signals, outputs can be
steered.
• Stable Training → Compared to GANs (Generative Adversarial Networks), diffusion
models are less prone to instability and mode collapse.

4. Applications of Diffusion Models

• Image Generation → Stable Diffusion, DALL·E, MidJourney.
• Text-to-Image → “A cat wearing sunglasses on the beach.”
• Image Editing (Inpainting/Outpainting) → Fill missing parts or expand pictures.
• Super-Resolution → Turn blurry/low-res images into HD.
 • Audio & Speech → DiffWave for speech synthesis.
• Video Generation → Emerging area (text-to-video, motion interpolation).

5. Simple Example (Analogy)

Imagine you have a jigsaw puzzle picture.

• Forward process = scramble the puzzle pieces step by step until it’s just chaos.
• Reverse process = teach the model how to put the puzzle back together step by step.
Once trained, the model can take random scrambled pieces and build completely new
puzzles (images, sounds, etc.).

6. Summary
Aspect Explanation

Definition Generative models that learn to reverse noise into data

Process Forward (noise → static), Reverse (denoise → data)

Strengths High quality, diversity, stable training

Key Applications Text-to-image, inpainting, super-resolution, audio/video generation

1. Why Prompt Design Matters in Image Generation

Unlike text-only LLMs, image generation models rely heavily on descriptive prompts.
A good prompt = more control over style, detail, and composition.
A weak prompt = random or mismatched results.

2. Core Components of an Image Prompt

When designing an AI art prompt, think of it as giving instructions to a digital artist.

a) Subject (What’s in the image?)

• “A futuristic city skyline”

• “A golden retriever puppy wearing sunglasses”

b) Style (How should it look?)

• Photorealistic, oil painting, watercolor, cyberpunk, Pixar-style, anime, surrealism, etc.

c) Details & Attributes

• Lighting → dramatic, soft, neon glow, natural sunlight

• Mood → dreamy, eerie, vibrant, calm
• Camera Angles → close-up, wide shot, aerial, portrait

d) Quality & Resolution Keywords

• “Ultra-detailed, 8K, cinematic, high resolution, hyperrealistic”

3. Prompting Framework: The Formula

[Subject] + [Medium/Style] + [Details/Attributes] + [Quality/Lighting/Camera]

Example:
"A majestic lion, digital painting, glowing eyes, fantasy jungle background, dramatic lighting,
ultra-detailed, 8K resolution"

4. Prompting Techniques

a) Descriptive Prompting

“A cozy cabin in the snowy mountains, warm lights glowing inside, photorealistic.”

b) Style Transfer Prompting

“A portrait of Einstein in Van Gogh’s Starry Night style.”

c) Contextual/Role Prompting

“Imagine you are a fashion photographer. Create a high-end photoshoot of a model in futuristic
clothing, neon city background.”

d) Negative Prompting (for Stable Diffusion)

Specify what you don’t want.

“A fantasy dragon flying in the sky — no blurry details, no extra limbs, no watermark.”

5. Examples of Strong vs Weak Prompts

Weak: “Draw a cat.”
Strong:
"A fluffy white Persian cat sitting on a velvet armchair, sunlight streaming through the window,
in the style of a 19th-century oil painting, ultra-detailed, warm tones."
6. Practical Applications
• Marketing → Generate ad visuals, product mockups.
• Entertainment → Concept art, storyboarding, game design.
• Education → Visual explanations, historical recreations.
• Personal Use → Profile pics, wallpapers, art experiments.

7. Summary Table
Prompt Element Purpose Example

Subject What’s in the image “A futuristic car”

Style Artistic look & feel “Cyberpunk neon style”

Details Attributes & environment “Rain-soaked streets, glowing signs”

Quality Resolution, realism “Ultra-detailed, 8K, photorealistic”

Negative What to avoid “No blur, no distortion”

Takeaway:
Well-designed prompts = more control, quality, and creativity in AI image generation.
Think of your prompt like directing a photographer or artist — the clearer and more vivid
your instructions, the better the result.

1. What Are Image Generation Models?

They are AI systems that convert text prompts into images. Most are built on diffusion
models or transformer-based architectures.
You type a description (“a cat in a spacesuit walking on the moon”) → the model generates an
image.

2. Major Image Generation Models

a) DALL·E (by OpenAI)

• Technology: Based on diffusion + transformer architectures.
• Strengths:
o Great at literal text-to-image accuracy.
o Strong at inpainting/outpainting (editing parts of an image).
o Integrates into ChatGPT, so you can generate and edit images in one place.
• Limitations:
 o Sometimes less stylistically diverse than MidJourney.
o Stricter safety filters (avoids sensitive content).
• Best For: Quick, accurate visualizations, design mockups, creative edits.

b) MidJourney
• Technology: Runs on Discord; uses diffusion + proprietary training data.
• Strengths:
o Famous for artistic, cinematic, stylized outputs.
o Excellent with lighting, atmosphere, surrealism.
o Community-driven → users can view and remix others’ prompts.
• Limitations:
o Less precise literal interpretation (sometimes “hallucinates” details).
o Not open-source, requires subscription.
• Best For: Concept art, digital art, fantasy/sci-fi styles, creative industries.

c) Stable Diffusion (by Stability AI)

• Technology: Open-source diffusion model.
• Strengths:
o Customizable — anyone can fine-tune with their own dataset.
o Works offline; models can be trained for specific styles.
o Supports negative prompting (tell it what not to generate).
• Limitations:
o Requires technical knowledge to set up.
o Quality depends on available fine-tuned checkpoints.
• Best For: Developers, researchers, hobbyists who want full control and custom styles.

d) Others Worth Noting

• Adobe Firefly → Integrated into Photoshop/Illustrator, focused on commercial use &
copyright-safe training data.
• Runway Gen-2 → Expands into AI video generation.
• Imagen (Google) → Very high-quality, but not fully public yet.

3. Feature Comparison
Model Strengths Limitations Best Use

Accurate, strong editing, integrates Quick, precise

DALL·E Limited style diversity
with ChatGPT illustrations

Artistic, cinematic, strong Less literal accuracy, Creative art, concept

MidJourney
community closed-source design
 Model Strengths Limitations Best Use

Stable Open-source, customizable, Setup complexity, variable Research, custom AI

Diffusion negative prompts quality art

Commercial design
Adobe Firefly Copyright-safe, pro tools integration Limited creative freedom
workflows

Runway Gen-
Text-to-video, creative media Still experimental Filmmaking, animation
2

4. Real-World Applications
• Marketing & Branding → ad visuals, product mockups.
• Entertainment → movie concept art, game design.
• Education → historical recreations, visual storytelling.
• Design → logo exploration, fashion design.
• Personal Use → profile pictures, fantasy art, storytelling illustrations.

5. Key Takeaways
• DALL·E → Best for practical, accurate, safe outputs.
• MidJourney → Best for artistic, stylized, and cinematic creations.
• Stable Diffusion → Best for open-source customization & experimentation.

1. What is Prompt Comparison?

It means testing different prompts for the same idea across models or variations — to see how
style, detail, and wording change the results.

Example Prompt: “A futuristic city at night with neon lights”

• DALL·E → Clear skyline, modern skyscrapers, accurate depiction.

• MidJourney → More cinematic, atmospheric, glowing cyberpunk vibe.
• Stable Diffusion → Flexible — depends on fine-tuned checkpoint (anime-style, realistic,
surreal, etc.).

Purpose: To learn how prompt structure affects outcome and to pick the best phrasing for
your goal.

2. What is Reverse Engineering?

Reverse engineering in AI image prompts = looking at an existing image and figuring out:
 • What keywords were likely used?
• Which style, lighting, and camera details influenced it?
• Was it photorealistic, painterly, anime, cinematic?

Example: You see an image of “A dragon flying in a storm, hyperrealistic, glowing scales,
cinematic lighting”.
From the look:

• Subject = dragon
• Style = hyperrealistic, cinematic
• Details = glowing scales, storm background
• Likely Prompt = “A hyperrealistic dragon flying through a thunderstorm, glowing
scales, cinematic dramatic lighting, ultra-detailed 8K.”

Purpose: To deconstruct successful prompts and learn which ingredients make them work.

3. Prompt Comparison Techniques

Keyword Swapping → Change one word at a time (“happy cat” → “angry cat”) and compare.
Stylistic Variations → Apply same subject with different styles (anime, oil painting,
photorealistic).
Cross-Model Testing → Run the same prompt in DALL·E, MidJourney, and Stable Diffusion
→ compare results.
Long vs Short Prompts → See if details improve with specificity (“a cat” vs. “a fluffy Persian
cat with blue eyes on a velvet chair”).

4. Reverse Engineering Techniques

Look for Visual Clues:

• Photorealism? → Likely had “photorealistic” or “ultra-detailed” keywords.

• Painterly brush strokes? → Likely “oil painting” or “watercolor style.”
• Cinematic mood/lighting? → Words like “dramatic lighting, cinematic, volumetric light.”

Check Metadata (when possible):

• Some Stable Diffusion images include embedded prompt metadata.

• MidJourney galleries often show the full prompt used.

Community Learning:

• MidJourney/Reddit/Discord groups share prompt + image pairs → great way to practice

reverse engineering.
5. Practical Example: Prompt Comparison
Prompt: “A portrait of Albert Einstein”

• Simple Prompt → “Albert Einstein portrait” → basic photo.

• Stylized Prompt → “Albert Einstein in Van Gogh’s style, oil painting, swirling colors”
→ impressionist look.
• Cinematic Prompt → “Albert Einstein, ultra-detailed, cinematic lighting, 8K resolution,
dramatic shadows” → movie-poster style.

Changing only style + details transforms the same subject drastically.

6. Benefits of Comparison & Reverse Engineering

• Learn what words matter most in shaping results.
• Improve your creative control.
• Borrow and remix successful prompt patterns.
• Avoid “prompt engineering guesswork” → become strategic.

Takeaway:

• Prompt Comparison = Experimenting with variations to see what works best.

• Reverse Engineering = Learning from existing AI art by breaking it down into subject,
style, details, and quality keywords.

1. Negative Prompts
Negative prompts tell the model what NOT to include in the generated image.

Why It Matters

• Reduces unwanted artifacts or errors.

• Helps avoid hallucinations or inappropriate content.
• Improves focus on desired elements.

Example:

• Prompt: “A fantasy castle on a mountain, sunset, detailed, photorealistic”

• Negative Prompt: “No blurry edges, no extra towers, no fog”

Combined Effect: The model generates a crisp, clean castle without unwanted features.
2. Prompt Rewriting
Prompt rewriting is the process of iteratively improving prompts for better results. It’s about
clarity, detail, and structure.

Key Techniques

1. Add Details – Include style, lighting, or mood.

o Weak: “A dragon”
o Strong: “A fierce dragon flying over a volcano, glowing scales, cinematic
lighting, ultra-detailed”
2. Specify Perspective or Composition
o “Close-up, 3/4 view, eye-level perspective”
3. Combine Styles or References
o “In the style of Studio Ghibli and digital painting, soft colors, whimsical”
4. Iterative Refinement
o Generate → Analyze → Rewrite prompt → Generate again → Repeat

3. Prompt Tuning
Prompt tuning is more technical. It involves training the model to respond better to certain
prompts.

Types

• Soft Prompting – Learn embeddings that guide the model without rewriting text.
• Fine-Tuning – Train on specific datasets for desired output styles or domains.
• Embedding Injection – Insert learned “style vectors” for consistent art styles.

Practical Use

• Companies train models for brand consistency in visuals.

• Artists create signature styles across multiple prompts.
• Researchers adjust prompts for medical or scientific accuracy in generated images.

4. Combined Example
Scenario: Generate a fantasy forest illustration.

• Initial Prompt: “A forest” → Output: Generic, lacks style.

• Rewritten Prompt: “A mystical forest at dawn, glowing flowers, misty atmosphere,
fantasy painting, ultra-detailed”
• Negative Prompt: “No humans, no animals, no dark shadows”
 • Tuned Prompt (optional) → Fine-tuned model for fantasy art ensures every forest
image has consistent color palette and style.

Result: A highly polished, stylistically consistent fantasy forest.

5. Summary Table
Technique Purpose Example

Negative
Exclude unwanted elements “No blurry edges, no watermarks”
Prompt

Prompt
Improve clarity, detail, style Add “ultra-detailed, cinematic lighting, fantasy style”
Rewriting

Train model for consistent Fine-tune on fantasy artwork to match style across
Prompt Tuning
output prompts

Takeaway:

• Negative prompts → control unwanted content.

• Prompt rewriting → improve quality via iterative refinement.
•

1. Content Outline Generation

Before writing, create a structured outline for clarity and flow.

Steps:

1. Define Topic & Purpose

o Example: “Guide on productivity apps for remote workers.”
2. Generate Headings & Subheadings
o Prompt: “Create a detailed blog outline on productivity apps for remote workers
with 5 main sections and 2–3 subpoints each.”
3. Include Key Points
o Each subheading should include bullet points for facts, tips, or examples.

Sample Outline:

Title: Top Productivity Apps for Remote Workers

• Introduction → Importance of productivity tools

• Section 1: Task Management Apps
 o Subpoint: Features (to-do lists, Kanban boards)
o Subpoint: Popular apps (Trello, Asana)
• Section 2: Communication Tools
o Subpoint: Video conferencing apps
o Subpoint: Team collaboration platforms
• Section 3: Time Tracking Apps
• Section 4: Note-Taking & Knowledge Management
• Conclusion & Recommendations

2. Writing Style for Blogs & Apps

Key Considerations:

• Audience: Beginner, professional, or tech-savvy users

• Tone: Informative, engaging, friendly, or formal
• Length & Depth: Short blog vs. long-form guide
• SEO-Friendly: Include keywords naturally for search engines

Example Prompts for Style:

• “Write a 500-word blog section about Trello for beginners. Tone: friendly, informative,
with examples.”
• “Rewrite this app description in a concise, app-store friendly tone.”

3. Integrating Images
Images increase engagement and help explain concepts. Use AI-generated visuals or stock
images.

Techniques:

1. Feature Illustration
o Prompt: “Generate an image of a person using a Kanban board app on a laptop,
modern office setting, photorealistic.”
2. Visual Infographics
o Prompt: “Create an infographic showing top 5 productivity apps for remote
workers, colorful, clean, simple icons.”
3. Screenshots / Mockups
o Prompt: “Generate a mobile app interface showing a task management
dashboard, modern UI, minimalistic style.”

4. Combined Workflow Example

Step 1: Outline
 • Prompt: “Create a blog outline on ‘Top 5 Productivity Apps for Remote Workers’ with 5
sections and key points.”

Step 2: Blog Content

• Prompt: “Write the introduction section, 150 words, friendly and informative tone.”

Step 3: Image Generation

• Prompt: “Generate an illustration of a remote worker using a productivity app on a

laptop, bright and modern office setting.”

Step 4: Style Refinement

• Prompt: “Rewrite the blog in a conversational style suitable for tech-savvy professionals,
keep it under 800 words.”

5. Summary Table
Step Purpose Prompt Example

Outline Structure content “Create a blog outline on top productivity apps.”

Style Writing Ensure tone & readability “Write 300 words in friendly, informative tone.”

Image Integration Visual appeal & explanation “Illustration of remote worker using a task app.”

Refinement SEO, style, consistency “Rewrite in concise, app-store friendly style.”

Takeaway:

• Start with outline → writing → visuals → style refinement.

• LLMs + AI image models = end-to-end content generation.
• Iterative prompting ensures clarity, engagement, and visual appeal.

1. Key Ethical Considerations

a) Accuracy & Misinformation

• AI can hallucinate facts or misinterpret prompts.

• Text Example: A generated health article may include false medical advice.
• Image Example: AI may create non-existent historical scenes presented as real.

Best Practice:
 • Always verify facts from trusted sources.
• Use AI as an assistant, not sole author.

b) Copyright & Intellectual Property

• AI-generated images or text may replicate existing styles or content.

• Risk: Infringing artist or writer copyrights.

Best Practice:

• Use copyright-free training data or commercial-licensed AI models.

• Avoid passing AI output as someone else’s work.

c) Bias & Fairness

• AI reflects biases in training data (gender, race, culture, politics).

• Text Example: Job descriptions favoring certain demographics.
• Image Example: AI generating stereotypical depictions.

Best Practice:

• Review outputs for bias.

• Use prompts that explicitly promote inclusivity.

d) Privacy & Data Security

• Avoid inputting personal, sensitive, or confidential data.

• AI could generate outputs that expose private information.

Best Practice:

• Use anonymized data.

• Follow GDPR and local privacy regulations.

e) Transparency & Disclosure

• Users should know when AI generated content is used.

• Builds trust and prevents misleading audiences.

Best Practice:

• Label AI-generated text or images clearly (e.g., “Created with AI assistance”).

2. Guidelines for Ethical Use in Text Applications
• Verify & fact-check: Especially in news, medical, or educational content.
• Avoid plagiarism: AI outputs should be original or properly cited.
• Promote inclusivity: Use neutral language, avoid stereotypes.
• Disclose AI involvement: e.g., blog, social media post, app content.

3. Guidelines for Ethical Use in Image Applications

• Respect copyrights: Avoid replicating artists’ unique styles without permission.
• Avoid deepfake misuse: Don’t generate misleading or harmful images of real people.
• Diversity & representation: Ensure visual content is inclusive.
• Watermark AI-generated images: Helps distinguish AI outputs from real photography.

4. Practical Examples
Scenario Ethical Risk Responsible Use

AI Blog Generation False health advice Fact-check, cite sources

AI Art for Commercial Use Style infringement Use open-source models, credit AI-generated art

Social Media Campaigns Biased portrayals Inclusive prompts, diverse visuals

App UI/UX Content Private data leakage Avoid sensitive user data in prompts

AI Deepfakes Misleading depictions Strict prohibition, watermarking

5. Key Takeaways
1. Verify AI-generated information; don’t blindly trust it.
2. Respect IP and avoid plagiarism in text and images.
3. Mitigate bias by reviewing outputs and crafting inclusive prompts.
4. Protect privacy and sensitive data.
5. Disclose AI use to maintain transparency.

Conclusion:
Generative AI is powerful, but ethical use is essential to prevent misinformation, bias,
copyright infringement, and privacy violations. Following these guidelines ensures
responsible, safe, and trustworthy AI deployment in text and image applications.