Introduction to
Generative AI
Revolutionizing Creativity and Knowledge Work
Dr. Prakash R Patil
KLE Technological University, Hubballi
Agenda
Introduction
2 Conceptualization
3 Limitations of Current Generative AI
Engaging the audience
Visual aids
Final tips & takeaways
Introduction to Generative AI 2
Quote
“Innovation is taking two things that exist and
putting them together in a new way.” — Tom
Freston
Presentation title 3
Human Creativity – Then and Now
Historically: Creative tasks seen as uniquely
human
Poetry, software, music, fashion
Paradigm shift: AI now rivals human creativity
Presentation title 4
What is Generative AI?
Definition:
AI techniques that generate new,
meaningful content from training data
Content Types:
Text, images, audio, video
Examples:
• Text: GPT-4
• Images: DALL·E 2
• Code: GitHub Copilot
• Used for artistic and productive
tasks
• Mimicking writers, illustrators
Impact on
• Practical applications:
Daily Life • IT helpdesks
& Work • Recipe suggestions
• Medical advice
Introduction to Generative AI 6
Goldman Sachs (2023):
• 7% potential increase in
Economic global GDP
• 300 million knowledge worker
& Social jobs may be affected
Impact • Major implications for business,
technology, and society
Introduction to Generative AI 7
• Opportunities:
• Innovation acceleration
• Improved knowledge work
Opportunit
• Risks:
ies & Risks • Job displacement
• Ethical and sustainability
concern
Introduction to Generative AI 8
• Conceptualizing GenAI as part of
complex systems
Generative • Focus on interaction between:
AI in Socio- • Humans
Technical • AI systems
• Organizational structures
Systems
Introduction to Generative AI 9
• Conceptualizing GenAI as part of
complex systems
Generative • Focus on interaction between:
AI in Socio- • Humans
Technical • AI systems
• Organizational structures
Systems
Introduction to Generative AI 10
Conceptualization of
Generative AI
Enhancing your presentation
Understanding Generative AI
Generative AI Key Components of Generative
AI
• Built upon generative
• Model:
modeling (vs. discriminative Machine learning architecture
modeling) (e.g., deep neural networks)
• Aims to model the • System:
underlying data Infrastructure including model,
distribution to create new data processing, and UI
data • Application:
Practical real-world use (e.g.,
content creation, code
generation)
Introduction to Generative AI 12
Mathematical Principles of
Generative AI
Feature Discriminative AI Generative AI
Purpose Classifies or predicts labels for input data Generates new data that resembles training
data
Focus Models decision boundary between X))
classes (P(Y
Output Type Class labels or probabilities New content (text, images, audio, etc.)
Data Use Learns from existing labeled data Learns patterns to create synthetic data
Example Task Spam email detection Writing an entire email in the style of a
human
Example Algorithm Logistic Regression, SVM, Random Forest GANs, VAEs, GPT, DALL·E
Use Case Example Classifying an image as cat or dog Generating a realistic image of a cat
Introduction to Generative AI 13
A Model-, System-, and
Application-Level View of
Generative AI
Model-Level View 🔹 Foundation Models
• Large, versatile generative models trained on vast datasets.
🔹 Definition & Core Role Capabilities:
• A generative AI model: machine learning architecture that • Emergence – exhibit unexpected abilities.
creates new data instances. • Homogenization – power many systems via one model (e.g., Copilot).
• Learns patterns from training data to generate novel, Modalities :
synthetic outputs. • Unimodal: same input/output type (e.g., text → text).
• Multimodal: different input/output types (e.g., text → image).
🔹 Deep Learning Architectures
Popular Models
•Deep Neural Networks (DNNs): suited for different data • GPT-4 – text and image input to text output.
types (e.g., text, images). • Midjourney – text/image to image.
•Examples: • Codex / AlphaCode – text-to-code.
• Stable Diffusion / MusicLM – text-to-image / music.
• Transformers & LLMs – GPT for text generation.
🔹 Training Techniques
• Diffusion Models – used for text-to-image (e.g., • GANs – dual-network competition to generate realistic samples.
DALL·E 2). • RLHF – used in ChatGPT; improves quality using human feedback and
reinforcement learning.
Introduction to Generative AI 14
Fig. 1 A model-, system-, and application-level view on
generative AI
Presentation title 15
Fig. 2 Examples of different training procedures for generative AI
models.
A) Generative adversarial network (GAN).
B) Reinforcement learning from human feedback (RLHF) as used in
conversational generative AI models
Presentation title 16
Application-Level View of Generative AI
🔹 Definition & Purpose 🔹 Shaping Future Work
Applications = Generative AI systems embedded in real-world New technology-enabled work modes emerging.
settings. Tasks evolve from basic (e.g., email writing) to complex
Goal: Solve business problems and address stakeholder needs.
Viewed as human-task-technology systems or information
(e.g., legal advice).
Raises concerns around accuracy, ethics, and
systems.
🔹 Real-World Use Cases accountability.
SEO content generation – optimizing digital visibility. 🔹 Trust & Adoption
Synthetic movie creation – AI-generated scripts & visuals. User trust will affect use vs. disuse of applications.
AI music composition – automatic music generation. Transition toward consequential decisions involving:
Natural language-based coding – tools like GitHub Copilot. o Moral judgment
o Legal/medical recommendations
Introduction to Generative AI 17
A Socio-Technical View on
Generative AI
🔹 Redefining AI Capabilities 🔹 From Delegation to Co-Creation
• Analytic AI: Human delegates task, AI
• Traditional AI: Analytical tasks (e.g., executes.
decision-making). • Generative AI: Promotes co-creation –
• Generative AI: Adds creative, content shared, iterative interaction.
generation capabilities. • AI interprets, suggests, and guides user
• Results can be novel, artistic, and decisions.
expressive.
🔹 Emergence of Hybrid Intelligence
🔹 From Passive to Agentic IT • Combines:
• Traditional IT: Human-directed, passive • Human creativity & empathy
tools. • AI speed, accuracy, scale
• Now: Agentic AI systems with decision- • Enables enhanced problem-solving and
making roles. innovation.
• Requires rethinking human-AI relationships • Calls for new interaction patterns and
and roles.to Generative AI
Introduction models. 18
A Socio-Technical View on
Generative AI
🔹 Theory of Mind in AI Context 🔹 Socio-Technical Implications
•In humans: Ability to infer others' • Design must account for trust,
thoughts and emotions.
perception, and human
• In AI: As interfaces become more behavior.
natural, humans begin to:
• Attribute "intent" or "emotion" to
• Interaction frameworks should
AI. support:
• Demand explainability and • Empathy in design
predictability.
• Shared understanding
• A "Theory of Artificial Mind" is needed
to guide human-AI c • Responsible AI agency
Introduction to Generative AI 19
Limitations of Current
Generative AI
🔹 1. Incorrect Outputs (Hallucinations) 🔹 2. Bias and Fairness Issues
• AI generates probable, not always factual • AI models can inherit & amplify societal biases
from training data.
content.
• Risks: Toxic language, stereotypes, and
• Risk of plausible-sounding discrimination.
misinformation.
• Sources of bias:
• Hallucinations: Semantically plausible but • Training datasets (e.g., CLIP, LAION).
factually wrong outputs.
• RLHF (Reinforcement Learning from
• Verification is difficult due to black-box Human Feedback) processes.
model nature. • Mitigation strategies:
• Solution attempts: • Prompt engineering (e.g., DALL·E 2’s
diversity prompts).
• Embedded correctness checks.
• Bias-aware design and quality checks.
• Probabilistic explanations/references
• Fair AI remains a work-in-progress research area.
to aid user judgment.
Introduction to Generative AI 20
Limitations of Current
Generative AI
🔹 3. Copyright and Legal Risks 🔹 4. Environmental Impact
• Risk of reproducing or imitating
• Large models (e.g., GPT-3) require
protected content (e.g., logos, art).
high energy consumption.
• Two types of violations:
• Reproduction rights (copying • Training emits hundreds of tons
content). of CO₂.
• Transformation rights (derivative • Sustainability concerns are rising.
works).
• Ongoing solutions:
• Legal gray areas:
• Ownership of AI-generated • Efficient training algorithms.
intellectual property. • Model compression and
• Training data provenance is critical. hardware optimization.
Introduction to Generative AI 21
Key Limitations of Current
Generative AI
🧠 1. Incorrect Outputs (Hallucination) ©️3. Copyright & Legal Concerns
• Models generate the most probable response, not • Can reproduce or imitate copyrighted works
always the correct one
• Prone to hallucinations—plausible but false content • May violate reproduction or transformation
rights
• Limited verifiability, often lacks grounding in real
facts • Raises questions of IP ownership for AI-
• Rooted in probabilistic inference and training data generated content
quality • Legal frameworks still evolving to catch up
⚖️2. Bias & Fairness Issues
• Training data often reflects societal biases (gender,
race, politics, etc.)
• Biases can be amplified in generated content
• Alignment processes like RLHF may introduce new
biases
• Fairness is still an ongoing research area
Introduction to Generative AI 22
Thank you
Dr. P R Patil
[email protected]
www.kletech.ac.in
