What is generative AI?
Most older AI tells things apart: spam or not, cat or dog. Generative AI does something new: it creates content (text, images, audio, code) that resembles what it learned from. Learn the difference, the modalities, and how it actually generates, right here on the page.
01Two jobs: telling apart vs. creating
Picture two jobs at a bakery: one worker sorts pastries into boxes labeled "croissant" or "muffin," while the other actually bakes a brand-new pastry from a recipe. Older AI mostly does the first job, sorting, and that sorting kind is called discriminative AI: it takes an input and predicts a label (is this email spam or not? is this photo a cat or a dog?). Generative AI goes the other way: from a prompt, it produces brand-new content (a paragraph, an image, a snippet of audio or code) that resembles the data it learned from. Flip between the two below.
Anchor your AI program in a charter. The AI Governance Charter: establish ownership, scope, and accountability for AI.
Your purchase helps keep our hubs free to read.
Discriminative AI sorts an input into a category. Example: given an email, it answers "spam" or "not spam."
- Discriminative answers a question about existing data: which category does this belong to?
- Generative answers a different question: what is a plausible new example that fits the pattern?
- Both learn from data; they just use what they learn for opposite purposes.
02What it can create: the modalities
"Content" covers a lot. Generative AI works across several modalities: different kinds of output, often powered by different model designs under the hood. Switch between them below.
Text: written language
Large language models (LLMs) generate sentences, summaries, answers, and translations by predicting plausible next words. This is the modality behind most chat assistants.
Image: pictures from a description
Image generators, often built on diffusion models, turn a text description into a new picture, starting from random noise and refining it toward the prompt.
Audio: voice, speech & music
Audio models synthesize spoken words from text, clone or create voices, and compose music. The output is generated sound, not a clip pulled from a library.
Code: programs & snippets
Code models write functions, fix bugs, and explain programs in many languages. Because code is just structured text, LLMs handle it as another kind of language.
03How it actually generates
Generative AI doesn't copy from a stored library. It learns the patterns in a large amount of training data, then samples from those patterns to build something new, piece by piece. Step through the idea.
04Check your understanding
You finished Generative AI
Here’s where it sits in your path, and the strongest next move.
Recommended next
Tokens, attention, training and inference, in plain language.
Prompt engineering
Techniques, patterns and pitfalls for getting better results from AI.
Open lesson → LanguageHallucinations & confabulation
Why models make things up, how to spot it, and how to reduce it.
Open lesson → LanguageTransformers
Attention, embeddings and the architecture behind modern AI.
Open lesson →Tokens & Context Windows
How LLMs chunk text and why context limits matter.
Open lesson → LanguageDiffusion & Image Generation
How AI image generation works, from noise to picture, step by step.
Open lesson → LanguageSpeech & Voice AI
Speech recognition, text-to-speech and how spoken-language AI works.
Open lesson →Two jobs: telling apart vs. creating
- Discriminative AI answers a question about existing data: which category does this belong to (e.g. spam or not, cat or dog)?
- Generative AI answers a different question: what is a plausible new example that fits the learned pattern?
- Both learn from data; they just apply what they learn to opposite purposes (classify vs. create).
What it can create: the modalities
- Text: large language models generate sentences, summaries, answers, and translations, the modality behind most chat assistants.
- Image: pictures from a description, commonly produced by diffusion models that refine from random noise toward the prompt.
- Audio: generated voice, speech, and music; synthesized sound, not a clip retrieved from a library.
- Code: programs and snippets across many languages; code is structured text, so LLMs handle it as another form of language.
How it actually generates
- It does not copy from a stored library; it learns the patterns (statistical distribution) of a large amount of training data.
- It then samples from those patterns to produce something new, built piece by piece.
- Each new piece feeds the next decision, so the output stays coherent, similar in style to the training data, but not a copy of any one example.
Limits worth knowing
- It can produce output that sounds confident yet is wrong, or that closely echoes its training data.
- Treat results as a strong first draft: review and verify before relying on them, especially for important decisions.
Every claim below links to its primary source so you can go straight to the original.
Published by Tech Jacks Solutions · Reviewed June 2026. This lesson explains established concepts and is grounded in the references below; figures shown in the interactives are illustrative and labelled as such.
Generative AI in 5 minutes
Tech Jacks Solutions · AI Knowledge Hub · educational summary
Discriminative vs generative
Discriminative AI tells things apart: it predicts a label from an input (spam-or-not, cat-or-dog). Generative AI creates new content (text, images, audio, code) that resembles its training data. Both learn from data; they use it for opposite purposes.
How it generates
It learns the patterns / distribution of a large amount of training data, then samples from those patterns to produce something new, built piece by piece, not copied from a stored library.
Modalities
Text: LLMs (sentences, summaries, answers). Image: often diffusion models (a picture from a description). Audio: voice, speech, and music synthesis. Code: functions, fixes, and explanations across languages.
Use it wisely
Powerful but fallible: output can be plausible-but-wrong or derivative. Treat results as a draft; review and verify before relying on them, especially for anything important.