Free access week – every resource, tutorial and practice application on this site is open to everyone until 10 November. No membership needed.
Home / AI Tutorials / Module 1
Module 1 of 10 · 7 lessons · about 45 min

AI Fundamentals

What AI actually is, where it came from, what it is doing in real companies today, and the limits you have to design around.

What is Artificial Intelligence?

Artificial intelligence is software that does a job we would call intelligent if a person did it – reading a document, recognising a face, deciding whether a transaction looks wrong, writing a test case. The important part for you is how it decides.

  • Ordinary software follows rules a developer wrote: if amount > 50000 then flag. Same input, same output, every time.
  • AI works out the rule itself from examples, and answers with a probability, not a certainty.
  • That single difference changes testing completely: there is often no one correct output to assert on.
  • An AI feature is still ordinary software around the edges – APIs, databases, a UI – and those parts test exactly as they always did.
Real-time example

A bank’s fraud check is AI. Your UPI payment at 2am from a new device is not blocked by a rule that says “block 2am payments” – a model scores it against millions of past payments and returns something like 0.87 risk. The team then picks the cut-off. Testing it means asking “is 0.87 reasonable?” and “what happens at the cut-off?”, not “does it equal 0.87?”.

History and evolution of AI

AI is not new. Knowing the phases tells you why some ideas that sound modern are forty years old, and why the last three years felt so sudden.

  • 1950s-60s – the idea. Turing asks whether a machine can imitate a person. The first neural network, the perceptron, is built. Expectations run far ahead of the hardware.
  • 1970s-80s – expert systems. Hand-written rules captured from specialists. They worked in narrow areas and collapsed outside them; funding dried up twice (the “AI winters”).
  • 1990s-2000s – statistics and data. Spam filters, recommendation engines, Deep Blue beating Kasparov. Learning from data beats hand-written rules.
  • 2012-2017 – deep learning. Cheap GPUs plus big labelled datasets. Image recognition passes human accuracy; speech and translation follow.
  • 2017 – the transformer. The paper “Attention Is All You Need” introduces the architecture every modern language model is built on.
  • 2022 onwards – generative AI goes mainstream. ChatGPT puts a language model in front of ordinary users; Claude, Gemini and open-weight models follow. AI becomes something testers are asked about.
Real-time example

The “self-healing locator” your automation vendor sells in 2026 is a 1990s idea – score several candidate elements, pick the most similar – running on better hardware. Knowing that stops you treating it as magic and makes you ask the right question: what is the similarity score, and what happens when two elements score the same?

Types of AI: Narrow, General and Superintelligence

Three labels get used loosely in articles and sales decks. Only one of them exists.

  • Narrow AI does one task: translate, classify, recommend, generate text. Everything in production today, including every large language model, is narrow AI.
  • General AI (AGI) would learn any task a person can and carry what it learned into a new one. It does not exist; serious researchers disagree about when or whether it will.
  • Superintelligence is the theoretical step beyond that. It is a philosophy discussion, not a release note.
  • A model that writes decent code and decent poetry is still narrow – it is doing one thing, predicting text, applied to two subjects.
Three types of AI: narrow AI exists today, general AI does not, superintelligence is theoryNarrow AIOne job, done well.Spam filters, UPI fraudchecks, ChatGPT, theself-healing locator inyour test framework.EXISTS TODAYGeneral AI (AGI)Learns any task a personcan, and moves what itlearned to a new one.Research goal, not aproduct you can buy.NOT YETSuperintelligenceBeyond human ability atmore or less everything.Debated by researchers,no working example.THEORYEvery AI feature you will be asked to test in the next few years is in the first box.
Types of AI: Narrow, General and Superintelligence
Real-time example

When a tool is sold to your company as “an AI that understands your application”, translate it: a narrow model was trained on screenshots and DOM snapshots and predicts likely elements. That tells you what to test – unusual layouts, shadow DOM, canvas-rendered UIs – instead of trusting the word “understands”.

AI vs Machine Learning vs Deep Learning vs Generative AI

These four words are used as if they were synonyms. They are nested circles, and each ring needs a different kind of testing.

  • AI is the umbrella: the goal of making software behave intelligently.
  • Machine learning is the main method: learn the rule from data. Fraud scoring, churn prediction, defect triage.
  • Deep learning is ML using neural networks with many layers – needed for images, audio and language.
  • Generative AI is deep learning that produces new content: text, code, images, audio. ChatGPT, Claude and Gemini are generative models.
  • Rule of thumb: all generative AI is deep learning, all deep learning is machine learning, all of it is AI – and the reverse is never true.
Artificial intelligence contains machine learning, which contains deep learning, and generative AI sits inside deep learningArtificial Intelligenceany system that does something we call intelligentMachine Learninglearns the rules from data instead of being told themDeep Learningneural networks with many layersGenerative AIwrites text, code and imagesA tester reads it like thisAI = the goalML = the methodDL = the engineGenAI = what youtype into every dayYour test strategychanges at each ring:fixed rules -> probabilities
AI vs Machine Learning vs Deep Learning vs Generative AI
Real-time example

One insurance claim, four rings. A rule rejects claims over the policy limit (plain code). A model scores the claim for fraud (ML). A network reads the uploaded hospital bill image (DL). An LLM writes the rejection letter (GenAI). Four different test strategies in one user journey – and only the first one has a single correct answer.

Real-world AI applications by domain

Where the money actually is. These are the systems you are most likely to be handed as a tester in India right now.

  • Finance: transaction fraud scoring, credit decisioning, document reading for KYC, customer-support assistants over policy documents.
  • Healthcare: triage and appointment routing, scan analysis as a second opinion, discharge-summary drafting, drug-interaction checks.
  • Retail and e-commerce: search ranking and “customers also bought”, demand forecasting, pricing, review summarisation, returns-fraud detection.
  • Insurance: premium scoring, claim fraud detection, document extraction from bills and reports, first-notice-of-loss chatbots.
  • Logistics and supply chain: route and slot optimisation, delivery-time prediction, warehouse demand planning, damage detection from photographs.
Real-time example

Open our own practice applications and you can see where each one would sit. In ShopKart, search ranking is the AI surface. In MediCare, it is appointment triage. In InsureHub, it is the premium and the claim decision. In CargoLink, it is the delivery estimate. Practising the non-AI journey first is how you learn to test the AI one.

AI ethics, bias, privacy and responsible AI

This is not a lecture section. Bias and leakage are defects, and increasingly they are the defects that get a feature pulled.

  • Bias comes from the training data. If past loan approvals favoured one group, a model trained on them repeats that pattern – and does it at scale, consistently.
  • Privacy. Anything typed into a hosted model leaves your network. Customer names, card numbers, patient data and internal code must not go into a prompt without an approved agreement.
  • Explainability. “The model said so” is not an answer a regulator accepts. Finance and insurance need a reason for a decision.
  • Accountability. A human has to own the outcome. Test the override path, not only the happy path.
  • As a tester: run the same case with the name, gender, age or city changed and compare outcomes. A difference you cannot explain from the business rules is a finding worth raising.
Real-time example

A hiring screener was tested with the same CV submitted twice – one with a male name, one female. Different scores. That is a two-minute test any tester can run, and it is the single most valuable AI test case there is. Keep the pair of inputs in your regression pack.

AI tools and where the jobs are

A short, honest map of what to learn and what it is worth.

  • Assistants: ChatGPT, Claude, Gemini and Copilot. Everyone will use these; on their own they are not a skill an employer pays extra for.
  • Building blocks: Python, an LLM API, embeddings, a vector store, LangChain or LangGraph, FastAPI. This combination is what “AI engineer” job posts ask for.
  • Testing side: AI-assisted test generation, self-healing automation, Ragas and DeepEval for evaluating LLM output, plus the usual Selenium, Playwright and Rest Assured underneath.
  • Roles: AI engineer, LLM/RAG developer, prompt engineer (increasingly folded into the other two), AI quality engineer, data scientist. The AI quality route is the shortest path for an experienced tester.
  • The honest bit: what moves a salary is not a prompt library. It is being able to say “I tested a RAG chatbot, here is how I measured groundedness, here is the defect I found.”
Real-time example

The fastest credible story for a tester today: take one of the practice applications on this site, point an LLM at its requirements, generate test cases, find where the generated cases are wrong, and write up that gap. That is a portfolio piece you can defend in an interview – unlike a certificate.

Practical: explore a chatbot and build a tiny AI-assisted workflow

  1. Open any assistant and give it a real requirement from your project. Ask for test scenarios. Count how many are usable without editing – that number is your baseline.
  2. Now give it the same requirement plus the field validations and the error messages. Count again. The jump you just measured is what prompt engineering means in practice.
  3. Ask it for the same answer as JSON with fields id, title, steps, expected. Notice whether every row really has every field – this is the first thing you will have to validate in code.
  4. Save both prompts. Module 6 turns them into something repeatable.
Use a feature from our practice applications if you cannot use your employer’s requirements – nothing you type there is confidential.
← Course overview2. Python for AI →