Chapter 7 — Introduction to Generative AI

Look at two portraits side by side. One is a photograph of a real person; the other was produced by a computer, and that person has never existed. In the CBSE activity Guess the Real Image vs. the AI-Generated Image, students try to tell which is which and give reasons. Many find it surprisingly hard. Clues that sometimes give an AI image away include odd hands or fingers, earrings that do not match, text in the background that is gibberish, hair that melts into the background and lighting that comes from two directions at once. But newer tools make fewer such mistakes, which is exactly why this topic matters.

Definition. Generative Artificial Intelligence (Generative AI) refers to AI systems that create new content, such as text, images, audio, music, video, computer code or simulations, that resembles content made by humans. The system is trained on a large amount of existing content, learns its patterns, and then produces something new that follows those patterns.

Content it can generate.

How is it different from the AI you have already met? A face-unlock system answers a question about existing data: 'Is this the owner's face?' A spam filter decides 'spam or not spam'. These systems recognise and sort. Generative AI goes one step further: it produces something that did not exist before.

A short history. Computer programs that generated text by simple rules existed decades ago, but they were limited. The big jumps came with deep learning. In 2014, Generative Adversarial Networks (GANs) began producing realistic images. In 2017, a new kind of neural network called the transformer made it possible to train much larger language models. From 2022 onwards, chatbots and image generators built on these ideas became available to the public and were used by millions of people within months.

Deepfakes. A deepfake is a fake video, image or audio clip, made with generative AI, that shows a real person saying or doing something they never said or did. Deepfakes show both the power and the danger of this technology, and you will study the ethics in the last topic.

Where you meet it. Chat assistants that answer questions and draft emails, apps that turn a photo into a cartoon, tools that remove a background or fill a missing part of a picture, voice assistants that speak naturally, and autocomplete that writes whole sentences are all generative AI.

Imagine teaching a child about animals in two different ways. In the first, you show cards and say 'this is a cat', 'this is a dog', until the child can name new pictures. In the second, the child spends weeks looking at thousands of animal pictures with no names attached, and learns so much about what animals look like that she can draw a new, believable animal herself. These two ways of learning match the two main kinds of AI modelling.

Discriminative modelling (supervised learning). A model is trained on labelled data: each example comes with the correct answer, given by people. The model learns to separate (discriminate between) the categories, for example 'cat' and 'dog'. Given a new picture, it answers 'cat' or 'dog'. It does not learn enough to draw either.

Generative modelling (unsupervised or self-supervised learning). A model is trained on huge amounts of unlabelled data. Without human labels, it figures out the patterns and structure in the data by itself: what usually comes next in a sentence, which shapes, colours and textures go together in a picture. Having learnt this, it can produce new examples that fit the same patterns.

How a text generator learns. A large language model reads an enormous amount of text and plays a guessing game billions of times: hide the next word, guess it, compare with the real word, and adjust itself slightly to guess better next time. The text itself supplies the answers, which is why this is called self-supervised. After training, it writes by predicting one likely word after another, guided by your prompt. It does not look up a stored copy of an answer; it generates new text from the patterns it learnt.

How an image generator learns. Many modern image generators are trained to remove noise. During training, random noise is added to millions of images step by step, and the model learns to reverse each step, guided by the image's caption. To make a new picture, it starts from pure noise and removes it gradually, steered by your text prompt, until an image appears.

Training and prompting. Two stages are involved. Training happens once, uses huge datasets and powerful computers, and is done by the company or lab that builds the model. Using the model is quick: you give a prompt (an instruction or description) and the model generates an output. A clear, specific prompt that states the topic, the audience, the style and the length usually gives a better result than a vague one.

What this means for you. Because a generative model predicts what is likely rather than what is true, it can produce fluent, confident text that contains mistakes. This is often called a hallucination. Always check facts from a generative tool against a reliable source.

A bank uses AI to flag suspicious transactions. A designer uses AI to produce fifty logo ideas in a minute. Both are AI, but they do very different jobs. The CBSE course compares generative AI with conventional AI (also called traditional or discriminative AI) on four points.

Same input, different jobs. Give both kinds of AI a photo of a street. A conventional system answers: 'three cars, two people, one traffic light'. A generative system, asked to, can produce a new version of the street at night, in the rain, or painted in the style of a famous artist.

Predictable versus varied output. If you ask a conventional spam filter about the same email twice, you get the same answer both times, which is what you want from a checking system. If you ask a generative chatbot for a story twice, you will usually get two different stories. This variety is useful for creativity but means the output must be checked every time.

They often work together. In practice, the two kinds are combined. A shopping app may use conventional AI to predict which products you will like and generative AI to write a short description of each. A medical system may use conventional AI to detect a possible disease in a scan, and generative AI to draft a plain-language explanation for the patient, which a doctor then checks.

Which should you use? Ask what the task needs:

• If you need a reliable yes/no answer, a category, a number or a ranking from existing data, conventional AI is usually the right tool.
• If you need new material, such as a draft, a design idea, a summary in a new style, or an illustration, generative AI is the right tool.
• If mistakes are costly, such as in medicine, law or exams, generative output must always be checked by a knowledgeable person.

A note on words. Both kinds use machine learning, and many generative systems are built from the same neural-network ideas as conventional ones. The difference is in what the model is trained to do: decide about data, or create data.

Just as there are different kinds of engines for cars, bikes and aircraft, there are different kinds of generative models. The CBSE course names four families. You do not need their mathematics, but you should know what each does and where it is used.

1. GANs (Generative Adversarial Networks). A GAN is made of two neural networks that compete and, in doing so, improve each other:

• the generator produces new data, for example a fake face;
• the discriminator examines data and tries to tell real examples from the generator's fakes, and its verdict is fed back to the generator.

Think of a forger and a detective. Each time the detective catches a fake, the forger learns what gave it away and improves. This feedback loop continues until the generator's output is so realistic that the discriminator can no longer tell it from real data. GANs are used to create portraits of people who do not exist, turn daytime photos into night scenes, generate images from descriptions and make realistic video.

2. VAEs (Variational Autoencoders). A VAE learns the overall distribution of the data, a kind of map of what typical examples look like and how they vary. To produce something new, it picks a point from that map and turns it into an example. VAEs are used to generate new images similar to the training set, rebuild damaged images, create rough drafts for writers, and produce new sounds and music.

3. RNNs (Recurrent Neural Networks). RNNs are designed for sequential data, where order matters, such as text, speech or music. They read a sequence one item at a time and carry a memory of what came before, using it to predict what comes next. They can generate new text in the style of a particular author or genre and predict the next character or word. Today, most large language models use a newer design called the transformer, which handles long passages better, but the idea of predicting the next item from what came before is the same.

4. Autoencoders. An autoencoder learns a compressed representation of data. It has two halves: an encoder that squeezes the input into a small code, and a decoder that rebuilds the original from that code. Because it must keep only the essentials, it learns the most important features. Autoencoders are used for removing noise from images, compressing pictures, artistic image creation and even drug discovery. A VAE is a special kind of autoencoder that can also generate new examples.

Hands-on: GAN Paint. In the GAN Paint demonstration, you start from a picture of a building and 'paint' with brushes labelled tree, door, brick, dome or grass. Each brush switches on or off a small group of neurons inside a GAN, and the network redraws the scene so that the new object fits naturally, with the right lighting and perspective. It shows that the network has learnt real-world concepts such as what a door looks like and where it belongs.

Generative AI is already used in art, music, writing, design and everyday study. Here are well-known examples from the CBSE course and the main kinds of tools you can try, always with a teacher's or parent's guidance and within each tool's age rules.

Art. In The Next Rembrandt project (2016), a team studied hundreds of paintings by the 17th-century Dutch painter Rembrandt, analysed features such as faces, lighting and brush strokes, and used data analysis and 3D printing to produce a new portrait in his style. It was not painted by Rembrandt, but many viewers found it convincing.

Music. AIVA is an AI composer that creates original pieces in genres such as classical, cinematic and pop. Film-makers and game designers use such tools for background music.

Language. Chatbots hold conversations, answer questions, summarise long documents, translate and draft letters or stories. Natural language generation systems also write routine reports, such as sports scores or weather summaries.

Common tools.

Tools change quickly; the names above are examples, not recommendations, and new ones appear every few months. What matters is knowing what kind of tool suits the task.

Hands-on: compare two chatbots. The CBSE activity asks you to give the same prompts to two or three chat assistants, such as 'Write a summary of the history of the internet' or 'Design an app that solves a real-world problem', and record your findings on parameters such as: how human-like the response is, the training data and technology behind it (as described by its maker), how accurate and authentic the answer is, whether it can access the internet, how easy the interface is, how well it summarises and writes paragraphs, and whether it is free or paid.

Writing a good prompt. A prompt works best when it states:

• the task: explain, summarise, list, compare, draft;
• the topic and context: 'for a Class 9 student', 'for a school notice';
• the format and length: 'in 5 bullet points', 'in under 100 words', 'as a table';
• the tone: formal, friendly, simple.

Compare 'Tell me about water' with 'Explain the water cycle to a Class 6 student in 5 short bullet points, with one everyday example'. The second gives a far more useful answer.

Using tools in real fields. Architects generate early design ideas; programmers get help writing and explaining code; musicians create backing tracks; content creators draft scripts and thumbnails; teachers make practice questions. In every case, the human chooses, checks and improves what the AI produces.

A small bakery owner who cannot afford a designer uses generative AI to create a logo, a menu and social-media posts in an afternoon. A student with dyslexia uses it to turn long paragraphs into short, clear points. These are real benefits. But the same tools can produce a wrong date in a history answer or a picture that repeats a stereotype. Using generative AI well means knowing both sides.

Benefits.

Limitations.

Other limitations to keep in mind.

• No real understanding: the model predicts likely words or pixels; it does not know whether something is true or appropriate.
• Out-of-date knowledge: a model knows only what was in its training data, unless it is connected to a live source.
• Copyright questions: models are trained on existing works, and who owns AI-made content, and whether creators should be paid, is still being debated in many countries.
• Over-dependence: if students let AI do all their writing or problem solving, they miss the practice that builds their own skills.

Weighing benefits against limits. A useful rule: the more serious the consequences of a mistake, the more carefully a human must check the AI's output. A fun birthday poem needs little checking; a medical explanation, a legal document, a news report or an exam answer needs a lot. Generative AI is best treated as a fast, capable assistant whose work you review, not as an authority whose word is final.

Fact-checking habit. When a generative tool gives you a fact, date, number or quotation, verify it in a textbook or a trusted website before you use it. If a tool gives references, check that they actually exist.

A video of a well-known leader announcing a fake emergency spreads on social media. It looks and sounds real, but it was made with generative AI. Before it is exposed as a deepfake, thousands of people have shared it and some have panicked. This is why the CBSE course asks you to study the ethics of generative AI: what is right and wrong in how it is built and used.

Three key ethical concerns.

• Bias: generative AI can copy and even amplify biases in its training data. If such content is used in serious decisions, such as hiring, loans or justice, it can treat groups of people unfairly.
• Misinformation: generative AI can produce fake news and deepfakes that spread falsehoods and manipulate public opinion. This can damage trust in elections, institutions and the media.
• Privacy: models can reveal or invent sensitive personal information, and images or voices of real people can be copied without consent and used to harm, cheat or bully them.

Wider effects on society. Some jobs that involved routine writing, illustration or customer support may shrink, even as new jobs appear in building, checking and guiding AI. Questions of ownership arise: if an AI produces a picture in the style of a living artist, should that artist have a say? And schools must decide how to keep assessment fair when AI can write essays.

Responsible use of generative AI. Developers, companies and users all share responsibility:

• use training data that is diverse and representative, so outputs are fair;
• check outputs for bias and misinformation before using or sharing them;
• put user privacy and consent first: never upload other people's photos, voices or personal details without permission;
• follow clear rules on ownership and attribution, and say when content was made with AI;
• take part in public discussion about the social and ethical effects of the technology.

Guidelines for students. Sakshi has an essay on 'The Impact of Climate Change on Coral Reefs', and a friend suggests using a text generator. Responsible use would look like this:

Spotting misinformation. Before sharing a shocking image, video or message, pause. Check whether trusted news sources report it, look for signs of editing, and do a reverse image search if you can. If in doubt, do not forward it.

The balance. Generative AI can make creativity and learning available to millions of people. Used without care, it can spread falsehoods, invade privacy and treat people unfairly. Laws, company policies and personal habits all have a role in getting the benefits while limiting the harm, and as a user you control the last of these.