TL;DR - Traditional software is deterministic (same input -> same output). Generative AI is probabilistic (same input -> similar but different output). Miss this and you'll be confused and burned.
Why it matters
People expect AI to behave like a calculator and panic when it doesn't repeat exactly. Understanding why it varies tells you when to lock it down and when to let it roam.
How it works
Traditional code follows fixed rules: 2 + 2 is always 4. An LLM samples from likely next words, so there's natural variation - a feature for creativity, a risk for consistency.
Worked example
Run "summarize this in 3 bullets" twice and you'll get two valid but differently-worded summaries.
A calculator never does that. An LLM always might. Neither is broken - they're different kinds of tools.
Steal this - control the variation
Need consistency (same format every time)?
- Lower "temperature" (creativity dial), if your tool exposes it.
- Give examples of the exact output you want (few-shot).
- Use a fixed template and ask it to fill the blanks.
Need fresh ideas? Let it run hot and ask for many options.Common mistakes
- Expecting pixel-perfect repeats from a probabilistic tool.
- Pasting AI output into a process that needs exactness without verifying.
- Blaming the model for variation that's normal behavior.
Good to know
"Temperature" is a real setting in the APIs (and some apps) for ChatGPT, Claude, and Gemini: low (0-0.3) for consistent/factual, higher for creative. You'll meet it again in Level 2 (system prompts & temperature) and Level 5 (building with APIs).