AI Basics: Definitions, History, Applications & Ethics

What is Artificial Intelligence?

Artificial Intelligence (AI) is the simulation of human intelligence in machines, enabling them to perform tasks that typically require human cognition, such as reasoning, learning, problem-solving, perception, and language understanding. AI systems analyze data, recognize patterns, and make decisions or predictions with minimal human intervention.

Key Definitions:

• AI (Artificial Intelligence): The broader field of creating systems that mimic human intelligence.
• Machine Learning (ML): A subset of AI where systems learn from data without explicit programming.
• Deep Learning: A specialized ML technique using neural networks with multiple layers (deep architectures).
• Narrow AI (Weak AI): AI systems designed for specific tasks (e.g., voice assistants, recommendation systems).
• General AI (Strong AI): Hypothetical AI with human-like cognitive abilities (not yet achieved).
• Superintelligent AI: Hypothetical AI surpassing human intelligence (theoretical).

History of AI: Milestones and Evolution

AI has evolved through key phases, shaped by technological advancements and research breakthroughs:

Timeline of AI Development:

Types of AI Systems

AI can be classified based on functionality, capability, and learning approach:

1. By Capability:

2. By Learning Approach:

Applications of AI in Nepal and Globally

AI is transforming industries worldwide, including Nepal. Below are real-world examples where AI is applied using concepts from this unit.

## In the real world

1. eSewa (Nepal):

   • AI Idea: Natural Language Processing (NLP) and Rule-Based Systems
   • How it works: eSewa uses AI to process user queries in Nepali (via chatbots) and validate transactions. For example, when you ask, "Bill payment for Ncell Rs. 500," the system parses your request, checks account balances, and executes the payment—all using predefined rules and NLP for language understanding.
   • Worked Example:
     • Input: User types "Bill payment for Ncell Rs. 500" in Nepali.
     • NLP Processing: The system extracts:
       • Action: Bill payment
       • Service Provider: Ncell
       • Amount: Rs. 500
     • Rule-Based Check: Validates if the user has sufficient balance and if Ncell is a supported service.
     • Output: Transaction successful or error message (e.g., "Insufficient balance").

2. Pathao (Nepal/Bangladesh):

   • AI Idea: Reinforcement Learning and Optimization
   • How it works: Pathao’s dynamic pricing and driver matching use AI to optimize routes and fares. For example, during Kathmandu traffic jams, the AI predicts delays and adjusts driver incentives to reduce congestion.
   • Worked Example:
     • Scenario: Traffic on the Ring Road is heavy due to a festival.
     • AI Action: The system detects increased demand and:
       • Increases driver incentives by 20% to attract more drivers.
       • Reroutes existing drivers via less congested paths (e.g., via Thapathali).
       • Adjusts fare prices dynamically (e.g., +15%) to balance supply and demand.
     • Outcome: Average wait time drops from 12 to 8 minutes.

3. NTC (Nepal Telecommunications Corporation):

   • AI Idea: Predictive Analytics and Anomaly Detection
   • How it works: NTC uses AI to predict network failures and optimize fiber-optic routes. For instance, during monsoon season, AI analyzes historical weather data and network logs to predict outages in hilly regions (e.g., Pokhara) and preemptively reroute traffic.
   • Worked Example:
     • Input Data:
       • Historical outage data (2018–2023).
       • Weather forecasts (heavy rain in Gorkha district).
       • Current network load (85% capacity in Pokhara).
     • AI Prediction: The model flags a 70% chance of a fiber cut on the Pokhara-Kathmandu link.
     • Action: NTC reroutes 30% of traffic via Chitwan and activates backup generators in Pokhara.
     • Result: Outage duration reduced from 6 hours to 30 minutes.

How AI Works: Core Components

AI systems rely on three fundamental components:

1. Data

• Role: The "fuel" for AI. Quality and quantity of data determine the system’s performance.
• Types:
  • Structured (e.g., databases, spreadsheets).
  • Unstructured (e.g., text, images, audio).
• Example: For a fraud detection system in banks like Nabil Bank, transaction records (structured) and customer reviews (unstructured) are used to train models.

2. Algorithms

• Role: Mathematical models that process data to generate insights or actions.
• Types:
  • Supervised Learning (labeled data, e.g., spam detection).
  • Unsupervised Learning (unlabeled data, e.g., customer segmentation).
  • Reinforcement Learning (learning from rewards/penalties, e.g., robotics).
• Example: Google’s search algorithm uses PageRank (a graph-based algorithm) to rank web pages.

3. Hardware

• Role: Provides the computational power for training and running AI models.
• Key Components:
  • GPUs/TPUs: Accelerate parallel computations (e.g., NVIDIA GPUs for deep learning).
  • Cloud Services: Google Cloud AI, AWS SageMaker, or Azure ML for scalable training.
• Example: Tesla’s autonomous cars use NVIDIA DRIVE GPUs to process real-time sensor data.

AI vs. Human Intelligence: Key Differences

While AI mimics human intelligence, there are critical distinctions:

Ethical Considerations in AI

AI raises ethical dilemmas that must be addressed for responsible deployment:

Key Ethical Challenges:

1. Bias and Fairness:

   • AI systems can inherit biases from training data (e.g., facial recognition tools performing poorly on darker-skinned individuals).
   • Example: In Nepal, if a loan approval AI is trained mostly on data from Kathmandu, it may unfairly reject applications from rural areas like Doti.

2. Privacy:

   • AI relies on vast amounts of personal data, raising concerns about surveillance and misuse.
   • Example: WhatsApp’s AI chatbots (like "WhatsApp Business") collect user data for personalized ads, sparking debates on privacy.

3. Accountability:

   • Who is responsible when an AI system makes a harmful decision (e.g., a self-driving car accident)?
   • Example: If a Pathao driver’s AI misjudges a pedestrian, is the company, the driver, or the AI developer liable?

4. Job Displacement:

   • Automation may eliminate jobs in sectors like manufacturing or customer service.
   • Example: Daraz’s AI-powered warehouses use robots for sorting, reducing the need for human pickers.

5. Transparency:

   • "Black box" models (e.g., deep neural networks) make it hard to explain AI decisions.
   • Example: A bank’s AI denying a loan without clear reasons violates transparency principles.

Ethical Frameworks:

• Asilomar AI Principles (2017): Guidelines for AI development (e.g., prioritize safety, fairness, and transparency).
• EU’s General Data Protection Regulation (GDPR): Regulates AI and data usage in Europe.
• Nepal’s Digital Identity Act (2018): Aims to protect digital privacy but lacks specific AI regulations.

AI in Everyday Life: Examples from Nepal

AI is embedded in everyday tools and services in Nepal, often without users realizing it:

1. Khalti:

   • AI Idea: Biometric Authentication and Fraud Detection
   • How it works: Khalti uses AI to verify transactions via fingerprint or face recognition. If an unusual login (e.g., from a new device in a different city) is detected, the system flags it for manual review.
   • Visual:

     flowchart LR
       A["User Initiates Payment"] --> B["Biometric Scan"]
       B --> C["AI Verification Module"]
       C --> D["Check: Device Location, Time, Transaction History"]
       D -->|"Normal"| E["Approve Payment"]
       D -->|"Suspicious"| F["Send OTP for Verification"]

2. Nepal Police’s Crime Prediction:

   • AI Idea: Predictive Policing
   • How it works: The Nepal Police use AI to analyze crime patterns (e.g., theft hotspots in Thamel) and deploy patrols proactively.
   • Worked Example:
     • Data Input: Crime reports from 2020–2023, weather data, and tourist footfall in Thamel.
     • AI Prediction: High theft risk on Fridays (20–25% increase) due to crowded markets.
     • Action: Police increase patrols on Fridays and deploy undercover agents near jewelry shops.

3. Farmers’ AI Assistants (e.g., "Kisan Suvidha" apps):

   • AI Idea: Expert Systems and NLP
   • How it works: Apps like "Kisan Suvidha" use AI to answer farmer queries in Nepali (e.g., "Best time to plant maize in Dhankuta?") and provide weather alerts.
   • Example Trace:
     • Query: "My rice plants are turning yellow in Sindhupalchok."
     • NLP Processing: Extracts symptoms (yellow leaves), location (Sindhupalchok), and crop (rice).
     • Response: "Possible nitrogen deficiency. Apply 20-20-20 fertilizer and check soil pH."

Limitations of AI

Despite its potential, AI has inherent limitations:

1. Data Dependency:

   • AI models are only as good as the data they’re trained on. Garbage in, garbage out (GIGO).
   • Example: If a medical AI in a Nepalese hospital is trained only on data from Kathmandu hospitals, it may misdiagnose diseases common in rural areas.

2. Lack of Generalization:

   • AI excels at specific tasks but struggles with tasks outside its training domain.
   • Example: An AI trained to recognize cows in Pokhara may fail to identify yaks in Solukhumbu.

3. High Computational Costs:

   • Training large models (e.g., LLMs like GPT) requires massive computational resources.
   • Example: OpenAI’s GPT-3 training cost ~$4.6 million for 300 billion parameters.

4. Interpretability:

   • Complex models (e.g., deep neural networks) are often "black boxes."
   • Example: A bank’s AI rejecting a loan application cannot explain why, violating transparency.

5. Ethical Risks:

   • AI can be weaponized (e.g., deepfake scams, autonomous weapons).
   • Example: In Nepal, deepfake voice calls have been used to scam businesses by impersonating managers.

AI and the Future: Trends to Watch

Emerging trends in AI will shape its future applications:

1. Explainable AI (XAI):

   • Developing AI models whose decisions can be easily understood (e.g., using decision trees instead of black-box neural networks).
   • Example: A hospital AI explaining "Patient X has a 78% risk of diabetes due to high HbA1c levels and family history."

2. Edge AI:

   • Running AI models on local devices (e.g., smartphones) instead of the cloud to reduce latency.
   • Example: Pathao’s real-time traffic rerouting happens on the driver’s phone, not a remote server.

3. AI for Social Good:

   • Using AI to solve global challenges like climate change, poverty, and healthcare.
   • Example: In Nepal, AI is used to predict landslide risks in hilly regions using satellite data.

4. Hybrid AI:

   • Combining AI with human expertise for better decision-making.
   • Example: Doctors using AI tools (e.g., IBM Watson Health) as a second opinion for diagnoses.

5. AI Governance:

   • Developing policies and regulations to ensure ethical AI deployment.
   • Example: Nepal’s upcoming AI Strategy (2024–2030) aims to integrate AI into national development plans.

Exam Tip

For the Tribhuvan University (TU) exam on this unit, focus on the following high-yield areas:

What to Prioritize:

1. Definitions:

   • Be able to distinguish between:
     • Narrow AI vs. General AI.
     • Machine Learning vs. Deep Learning.
     • Supervised vs. Unsupervised Learning.
   • Example Question: "Differentiate between reactive machines and limited memory AI systems with examples."

2. Applications:

   • Link AI concepts to real-world Nepalese examples (e.g., Khalti’s NLP, NTC’s predictive analytics).
   • Example Question: "How does Pathao use reinforcement learning to optimize driver routes during traffic congestion in Kathmandu?"

3. Ethical Dilemmas:

   • Discuss bias, privacy, and accountability with case studies.
   • Example Question: "A facial recognition system in Nepal performs poorly for people with darker skin tones. Analyze the ethical and technical reasons behind this bias."

4. Limitations:

   • Explain why AI fails in certain scenarios (e.g., lack of generalization, data dependency).
   • Example Question: "Why might an AI trained to diagnose diseases in Kathmandu hospitals fail in rural hospitals like those in Achham?"

5. Future Trends:

   • Mention XAI, Edge AI, and AI governance as emerging topics.
   • Example Question: "What is Explainable AI (XAI), and why is it important for healthcare applications in Nepal?"

Common Pitfalls to Avoid:

• Vague Answers: Always use specific examples (e.g., eSewa, Pathao, NTC) instead of generic explanations.
• Ignoring Ethics: TU exams often test ethical implications—always include a paragraph on bias, privacy, or accountability.
• Overcomplicating: Stick to concrete concepts (e.g., how NLP works in Khalti) rather than abstract theories.
• Memorization: Avoid rote learning of definitions. Explain with diagrams, traces, or real-world ties.

Sample Exam Question and Answer Structure:

Question: "Explain the role of AI in digital payment systems like eSewa, with a focus on Natural Language Processing (NLP) and rule-based systems. Discuss one ethical challenge associated with such systems."

Model Answer:

1. Introduction (1 mark): "Digital payment systems like eSewa leverage AI to enhance user experience and security. Two key AI techniques used are Natural Language Processing (NLP) for understanding user queries and rule-based systems for transaction validation."

2. NLP in eSewa (3 marks):

   • How it works:
     • Users input queries in Nepali (e.g., "Bill payment for Ncell Rs. 500").
     • NLP processes the text to extract:
       • Intent: Bill payment
       • Entity: Ncell
       • Amount: Rs. 500
   • Example Trace:

     flowchart LR
       A["User Input: 'Bill payment for Ncell Rs. 500'"] --> B["NLP Tokenization"]
       B --> C["Intent Classification: Payment"]
       C --> D["Entity Recognition: Ncell, Rs. 500"]
       D --> E["Rule Engine Validation"]

   • Visual:

3. Rule-Based Systems (3 marks):

   • How it works:
     • Validates the extracted data against predefined rules:
       • Is the user’s account active?
       • Is Ncell a supported service?
       • Is the balance sufficient?
   • Example:
     • If balance < Rs. 500 → "Insufficient funds. Top up now?"
     • If Ncell is unsupported → "Service unavailable. Try another provider."

4. Ethical Challenge (2 marks):

   • Bias in NLP Models:
     • If the NLP model is trained mostly on Kathmandu-centric data, it may misinterpret rural dialects (e.g., "Bill" vs. "Billaharu" in Western Nepal).
     • Impact: Users from rural areas may face errors or require manual intervention.
   • Privacy Concerns:
     • Storing transaction histories raises risks of data breaches (e.g., hackers accessing bank details).

5. Conclusion (1 mark): "AI transforms digital payments by enabling seamless interactions and secure transactions. However, challenges like linguistic bias and privacy must be addressed for inclusive and ethical AI deployment."

Summary of Key Points

Visual Summary: AI System Pipeline

flowchart LR
  A["Data Collection"] --> B["Data Preprocessing"]
  B --> C["Model Selection\n(Supervised/Unsupervised/RL)"]
  C --> D["Training\n(Algorithms + Hardware)"]
  D --> E["Evaluation\n(Accuracy, Bias Checks)"]
  E -->|"Deploy"| F["AI System\n(e.g., eSewa Chatbot)"]
  F --> G["User Interaction\n(e.g., Payment Query)"]
  G --> H["Feedback Loop\n(Improve Model)"]