What is generative AI?
Generative AI is a kind of artificial intelligence that creates new content similar to what it learnt from. Older AI is mostly discriminative: it sorts or predicts (spam or not, cat or dog). Generative AI makes a new email, a new picture of a cat or a new tune.
| Type | Creates | Example use |
|---|---|---|
| Text (large language models) | answers, essays, summaries, code | chatbots, writing helpers |
| Images | pictures from a text prompt | posters, concept art |
| Audio | speech, music, sound effects | voice-overs, read-aloud |
| Video | short clips | animation drafts |
How does it work?
1. Training on huge data
A neural network with millions or billions of adjustable numbers (parameters or weights) is shown enormous amounts of data. Each time its guess is wrong, the numbers are nudged. In the end the model holds patterns, not a library of copies.
2. Text: predicting the next token
A large language model (LLM) splits text into tokens (words or parts of words). Given a prompt, it gives each possible next token a probability, picks one, adds it, and repeats. Temperature controls the choice: low temperature picks the most likely word (safe, repetitive); high temperature lets less likely words in (creative, more errors).
3. Images: diffusion and GANs
A diffusion model learns to remove noise from pictures. To create, it starts from pure noise and removes noise step by step, steered by the prompt. A GAN (generative adversarial network) has two parts: a generator makes fakes and a discriminator tries to catch them; both improve by competing.
Prompts
A good prompt is clear and specific: give the task, the audience, the format and any limits. "Explain gravity to a 10-year-old in 3 short sentences" works better than "gravity".
Uses of generative AI
- Learning: explaining ideas simply, making practice questions, translating.
- Work: drafting emails and reports, summarising long documents, writing and fixing code.
- Creative: idea sketches, music drafts, storyboards.
- Access: reading text aloud, describing images for people with low vision, voice help in local languages.
- Science and health: designing new molecules, drafting medical notes for a doctor to check.
Generative AI is a helper, not an expert you trust blindly. A person must stay in charge of important decisions.
Risks and ethics, and responsible use
Risks
- Hallucination: the model can state made-up facts, quotes or references confidently.
- Bias: if the training data is unfair, outputs can be unfair (for example showing only men as engineers).
- Deepfakes and misinformation: fake images, voices and videos of real people.
- Privacy: personal data typed into a tool may be stored or leaked.
- Copyright and credit: questions about using creators' work for training and about who owns outputs.
- Cheating and over-reliance: copying AI answers stops you learning.
- Energy: training and running big models uses a lot of electricity and water.
Responsible use (C-C-P-K)
- Check facts with trusted sources.
- Credit: say when you used AI; follow your school's rules; respect creators.
- Protect privacy: never share passwords, addresses, phone numbers or others' photos.
- Kind: never use AI to bully, fake or deceive.
Try it: the 3D and at home
In free play, choose prompt 1, set temperature to 0.2 and press Generate five times. Then set 2.0 and do it again. How many different words did you get each time?
At home: play "human language model" with your family. One person says a sentence start ("Every morning I…"); each person adds one word in turn. Notice how you choose likely words, just like an LLM. Then try adding only surprising words: that is high temperature!
Key formulas and definitions
- Generative AI: learns patterns from data → creates new content
- LLM loop: prompt → probabilities for next token → pick one → add → repeat
- Low temperature → most likely word; high temperature → more variety
- Diffusion: noise → remove a little noise × many steps (guided by prompt) → image
- GAN: generator (makes) vs discriminator (checks)
- Responsible use: Check · Credit · Protect · Kind
Worked examples
1. Is a spam filter generative AI? Is a chatbot that writes a poem generative AI?
The spam filter sorts emails into two groups, so it is discriminative. The poem-writing chatbot creates new text, so it is generative AI.
2. A model gives next-word probabilities: tea 0.45, coffee 0.40, milk 0.15. At very low temperature, which word is chosen? At high temperature?
Low temperature: almost always "tea" (the highest). High temperature: any of the three can be chosen, and "milk" becomes more likely than before.
3. Improve this prompt: "Write about water."
For example: "Write 5 bullet points for Class 8 students on how to save water at home, with one example each." It states task, audience, format and length.
4. A chatbot gives a book reference that you cannot find anywhere. What happened and what should you do?
It is likely a hallucination: the model produced a believable but made-up reference. Check with a library catalogue or trusted website and do not use it unless verified.
Common mistakes
- Believing everything the AI says. It predicts likely words, not checked facts.
- Thinking the model stores and copies whole documents. It learns patterns, though it can sometimes repeat training text, which is a copyright concern.
- Typing personal details (address, phone, passwords) into a chatbot.
- Submitting AI-written work as your own without permission or credit.