Skip to content

Generative AI

Generative AI is artificial intelligence designed to create new content such as text, images, audio, video, code, or structured information.

Generative AI is artificial intelligence designed to create new content based on patterns learned from data.

Generated outputs can include:

  • Text
  • Images
  • Audio
  • Video
  • Code
  • Structured information

Large Language Models are a common form of generative AI used for language and code.

How It Works

A generative model receives input and produces an output based on learned statistical patterns.

For a language model, the process generally involves predicting Tokens that form a response.

The generated result is not normally retrieved as a finished answer from a database.

The model constructs the output.

That distinction helps explain both the flexibility and the uncertainty of generated content.

A traditional search system primarily finds existing information.

Generative AI creates a response.

Retrieval-Augmented Generation combines those ideas:

Question
   ↓
Retrieve relevant source material
   ↓
Provide it to generative model
   ↓
Generate response

The model is still generating the answer even when the information comes from retrieved sources.

Why It Matters

Generative AI is useful where the desired output cannot be fully predefined.

Examples include:

  • Summarization
  • Drafting
  • Classification
  • Translation
  • Extraction
  • Question answering
  • Code generation
  • Information synthesis

But generated content can contain Hallucinations and should not automatically be treated as verified fact.

Common Problems

Common problems include:

  • Treating generated text as database output
  • Assuming confident language means accuracy
  • Using AI for tasks that require exact deterministic results
  • Sending unnecessary sensitive data
  • No source verification
  • Weak Prompt design
  • No AI Evaluation
  • Automating high-impact actions without oversight

How to Improve

Define what the model is allowed to generate and what must remain deterministic.

For factual use cases, consider Retrieval-Augmented Generation.

For consequential outputs, consider Human-in-the-Loop review.

Evaluate the system on realistic examples rather than a few impressive demonstrations.