Introduction to AI

The most critical concept for an AI tester is the structural difference in how traditional software and AI software are built. In a conventional system, developers code the exact rules and logic, and the system applies those rules to input data to produce an answer. In an AI-based system, developers instead supply data and the desired outcomes, and the system applies algorithms to that data to learn the rules for itself.

AI is classified into three evolutionary stages based on its capabilities and autonomy - from performing one task well, to reasoning like a human across any domain, to (hypothetically) surpassing human intelligence altogether.

"AI" is an umbrella term encompassing several specialized fields. Testers must identify which technology is actually being used in order to design an appropriate test strategy for it.

• Machine Learning (ML): Systems that use statistical models to learn from data without being explicitly programmed.

  Example: An airline pricing algorithm that adjusts ticket prices based on historical demand trends.

  Example: An e-commerce system grouping customers into marketing segments based on purchasing behavior.

• Deep Learning (DL): A highly complex subfield of ML that uses multi-layered artificial neural networks inspired by the human brain, requiring massive amounts of data.

  Example: Real-time language translation that analyzes live audio streams.

  Example: Autonomous vehicles using sensor data to navigate city streets.

• Natural Language Processing (NLP): Technology that allows machines to understand, interpret, and generate human language.

  Example: A customer service chatbot identifying the sentiment (angry vs. happy) of a user's text message.

• Computer Vision (CV): Technology that enables machines to extract information from digital images and video.

  Example: A medical AI scanning an MRI image to detect early signs of tumors.

• Expert Systems: An older, rule-based form of AI that does not learn from data - it uses a massive, hard-coded database of "if-then" rules created by human experts.

  Example: A legacy medical diagnosis tool where doctors programmed 5,000 specific symptom combinations, each mapped to a specific disease.

A major focus of the v2.0 syllabus is the split between systems that categorize existing data and systems that create new data.

• Large Language Models (LLMs): A specific type of Generative AI trained on vast amounts of text. They work by predicting the next most logical word in a sequence, rather than by "understanding" facts.
• The testing challenge: hallucinations: Generative AI suffers from "hallucinations" - instances where the model confidently states fabricated, illogical, or false information as absolute fact. Because the system is only predicting the next word, it lacks actual factual comprehension, so testing for hallucinations requires entirely different approaches than standard functional testing.

Training AI requires massive computational power. Standard Central Processing Units (CPUs) process tasks sequentially - one at a time - which is too slow for the millions of calculations AI training requires.

Where the AI "lives" determines how a tester will verify latency, privacy, and integration - the two main options sit at opposite ends of a trade-off between raw compute power and speed/privacy.

• Cloud Hosting / MLaaS (Machine Learning as a Service): The heavy ML model runs on powerful remote servers (AWS, Google Cloud). The local app sends data via an API, and the cloud returns the prediction.

  Example: Using the OpenAI API inside a web app. Test focus: network latency, API security, and data transmission.

• Edge Computing (on-device): The AI model is shrunk down and deployed directly onto the local device (smartphone, IoT device, car) - it offers ultra-low latency (no waiting on the internet), works offline, and protects privacy because data never leaves the device.

  Example: A smartphone's fingerprint scanner. Test focus: device battery drain and offline capability, in addition to the recognition logic itself.

Developers do not write the complex matrix calculus behind AI from scratch - they use frameworks, pre-built code libraries that provide the building blocks for creating AI models.

• Key frameworks: TensorFlow (Google), PyTorch (Meta), and scikit-learn (for more standard, non-deep-learning ML) are the most common examples.
• Testing impact: Testers often need to write automation scripts that interact natively with these specific frameworks, requiring some understanding of how data structures (like tensors) are passed into the system.

Because AI models are probabilistic, can exhibit severe biases (e.g. learned from biased historical training data), and act as "black boxes," global governments are increasingly regulating them.

• Regulations (the law): The EU AI Act classifies AI systems by risk. A spam filter is low risk; AI used in hiring, loan approvals, or criminal justice is high risk and faces intense legal scrutiny.
• Standards (best practices): Frameworks like ISO/IEC 42001 and the NIST AI Risk Management Framework guide organizations on how to build safe, transparent, and unbiased AI - they are non-mandatory guidance, unlike a binding regulation.
• Testing impact: A tester's job is no longer just finding bugs - it includes verifying legal compliance, auditing datasets for demographic bias, and ensuring the model handles user data according to privacy laws like GDPR.