Lesson 3 of 7 · 10 min
How AI makes things
Goal:See why AI output looks finished, and learn when not to trust it.
WHAT IT IS ACTUALLY DOING
uxuiopen.comYou don't need to know the math. You need a good mental model: what the AI is doing when it answers, so you can predict where it will help and where it will mislead you.
Patterns
Learned from huge amounts of text and images
Your prompt
The only thing it knows about your project
Prediction
Picks the most likely next piece, over and over
Output
Looks finished. Might be wrong.
Under the hood, a language model picks a likely next piece of text, again and again, until the answer is done. Image tools work in a similar spirit: they build a picture that is likely, given your words. Play with it.
A poster for a jazz night should feel ____
Low randomness: the most likely word wins almost every time. Predictable, often average.
- moody30%
- smooth22%
- warm16%
- bold12%
- retro10%
- loud6%
- crunchy4%
Words to know
- Model
- The AI system itself, trained on huge amounts of text or images to recognize patterns.
- Prompt
- Everything you send it: your request, plus any files, images or earlier messages it can see.
- Temperature
- How adventurous its choices are. Low gives safe, similar answers; high gives surprising, less reliable ones.
- Hallucination
- When AI states something false with full confidence, like a fake statistic or a quote nobody said.
- Training data
- What the model learned from. It reflects the past, including its gaps and biases.
- Context
- What the model can see right now. It knows nothing about your project except what's in the context.
- Plausible isn't proven. Checking is your job.
- Leave gaps and AI fills them with averages.
- Names, numbers and quotes always need a source.
Five things an AI tool gave you. Use as a starting point, or check first?
Twenty headline ideas for a coffee shop app.
"68% of users abandon carts with more than three steps." No source given.
A summary of five interview transcripts, with three key themes.
Three color palettes for a calm meditation app.
A quote about simplicity from a famous designer.
Why vague prompts look samey
Because it predicts what's most likely, AI drifts toward the middle of everything it has seen. Ask for "a poster" and you get the average poster, again and again. Every detail you add moves it away from the average and toward your project.
Vague prompt
“Make a poster for an event”
Six versions of the average poster
Specific prompt
“Bold, playful poster for a free student design workshop, date easy to read on a phone, three different directions”
Three directions that fit the brief
AI is great at…
- lots of options, fast
- first drafts of copy and layouts
- summarizing and organizing your notes
- explaining things in simpler words
Check carefully when it…
- gives numbers, dates, names or quotes
- talks about real people or companies
- makes claims about your users
- sounds very sure about something niche
Output that looks finished is the easiest kind to wave through. Here's a card an AI drafted from the Design Night brief. Read it like an editor and click every mistake.
Click everything that's wrong. There are four mistakes.
The brief: Design Night, a free evening workshop. Friday, 7pm, in Studio 4.
0 of 4 mistakes found
AI writes "92% of students prefer evening workshops" in your poster copy. What do you do?
You asked for "a logo" and all ten results look alike. Why?
Read next