What you will learn
  • Distinguish generation from classification.
  • Explain token-by-token text generation.
  • Describe diffusion without treating it as collage.
  • Recognize what prompts can and cannot control.
  • Evaluate a generated result against requirements.

Before you begin

Understand the relationship between AI, machine learning, and deep learning from AI02.

Producing content differs from selecting a label

A photo classifier may return bicycle. A generative system might create a paragraph describing a repair or an image of a bicycle in a garden. Generative AI is defined by producing content, not by one architecture. Systems can use different mechanisms and accept different input types.

Training exposes a model to patterns in data and adjusts its parameters. During generation, it uses those relationships together with supplied input. It does not normally retrieve one complete stored answer and paste it unchanged. However, memorization and reproduction can occur, so new-looking output is not proof of originality.

A plausible explanation, an accurate explanation, and a formally verified explanation are different achievements. Fluency is immediately visible; correctness requires a separate check.

Text grows one token at a time

Many text generators process tokens: units representing words, word pieces, punctuation, or other symbols. Given the input and previously generated tokens, the model calculates scores for possible next tokens. A decoding procedure chooses a token, adds it to the sequence, and repeats until a stopping condition is reached.

For an imaginary exercise beginning The garden robot found a, several continuations might fit: stone, leaf, or broken sensor. Previous words constrain the possibilities but do not determine a single ending. Selection settings can change variety; their names and availability depend on the tool.

Repeated prediction can create structured answers because training has shaped relationships across language, facts, and reasoning examples. Token prediction describes a mechanism; it does not mean behavior is limited to finishing short phrases. It also does not guarantee that a generated chain of statements is true.

One route to images: removing noise

Diffusion models offer another mechanism. During training, a model learns relationships between clean examples and versions altered by noise. During sampling, a procedure starts with noise and repeatedly refines a representation toward a structured result. A text condition can guide the image toward a requested scene.

The original denoising diffusion work describes a probabilistic model and learned reverse process. Generation emerges through numerical transformations, not through a person selecting picture fragments. Denoising Diffusion Probabilistic Models.

Not all image generators use the same architecture or sampling process. Some operate in a compressed representation before converting it to pixels. Judge whether the result meets the task: correct object count, consistent geometry, readable lettering, or another explicit requirement.

A worked prompt and its checks

A weak prompt is Write about robots. It leaves reader, purpose, length, and boundaries open. An improved prompt is: Write a 150-word explanation for a 13-year-old about how an obstacle-avoiding robot uses a distance sensor. Define sensor and motor. Explain one limitation. Do not name products or invent test results.

This specifies audience, scope, output size, and evidence boundary. These constraints make evaluation easier but provide no guarantee. A useful follow-up is: Identify any sentence implying that a distance sensor recognizes object identity, and revise it. This targets a likely conceptual error.

Verify the result yourself. Count words, check definitions, and compare a named device's claims with its documentation. Ask whether the limitation is meaningful. A model describing its answer as checked is not independent evidence that those checks happened.

Choose checks for the intended use

Generation helps when alternatives are valuable: explanations, test cases, names, or practice problems. Retain responsibility for selecting and correcting results. Keep source material separate from invented examples so readers know which is which.

A classroom poster may benefit from an imaginary robot illustration even if the mechanism would not work. A wiring guide requires a diagram that agrees with actual connections. The same generation capability needs different evaluation depending on what a reader will do with the output.

Important terms

Generative AI
Systems producing content from learned patterns and supplied conditions.
Token
A unit processed by a language model.
Decoding
Choosing output tokens from model scores.
Diffusion model
A generative model associated with learning to reverse noise corruption.
Conditioning
Using supplied information to influence generation.
Hallucination
An unsupported or incorrect generated claim.

Mini project: Compare two explanations

  1. Run the weak and improved robot prompts in an available AI tool.
  2. Check audience fit, definitions, length, limitation, and factual support.
  3. Record one helpful change and one remaining problem.
  4. Revise the best answer and identify supporting evidence. Finish with a checked paragraph.

Common mistakes and debugging

  • Treating polished language as proof: verify claims separately.
  • Assuming every image generator uses diffusion: consult its documentation.
  • Equating generated with automatically original: examine provenance and reuse conditions.
  • Increasing prompt length without purpose: add constraints that help define or evaluate output.

Independent challenge

Adapt the robot prompt for an eight-year-old while preserving the distinction between distance measurement and object recognition.

Check your understanding: 10 questions

  1. What distinguishes generation from classification?

  2. What happens after choosing a token?

  3. Does image output prove diffusion is used?

  4. Why request a limitation?

  5. Why is self-reported verification insufficient?

  6. In your own words, what does “Generative AI” mean?

  7. In your own words, what does “Token” mean?

  8. In your own words, what does “Decoding” mean?

  9. In your own words, what does “Diffusion model” mean?

  10. In your own words, what does “Conditioning” mean?

Quiz answers

Reveal all 10 answers after your attempt
  1. Generation produces content; classification assigns categories, although one system can combine both.
  2. It is added to the sequence and influences subsequent predictions.
  3. No. Different architectures can generate images.
  4. It encourages the explanation to state boundaries readers need to understand.
  5. It is itself generated text; verification requires observable checks and evidence.
  6. Systems producing content from learned patterns and supplied conditions.
  7. A unit processed by a language model.
  8. Choosing output tokens from model scores.
  9. A generative model associated with learning to reverse noise corruption.
  10. Using supplied information to influence generation.

Summary

Generative models create content through learned transformations. Prompts guide output; checks establish whether the result is useful and correct.

Continue learning

AI04 examines the language models behind text-generating assistants.

Choose a connected learning path

Sources and further reading

Prepared 2026-09-18. Draft — generation mechanisms source checked; tool-independent exercise