Day 16 · Spot the Glitch

🔍 Lab · Glitch Hunt

Tap everything that's wrong

This scene was "generated" by a machine that doesn't understand the world — it just makes stuff that looks right at a glance. There are 4 planted glitches hiding in here. Tap each one you spot. Miss-taps are totally fine — that's the whole game.

0glitches found

4planted in scene

0your taps that missed

A real artist would never draw a 6-fingered hand. A generator does it constantly, because it's copying patterns — not thinking. 🖐️

🧪 Break-It · Bias in, Bias out

The generator that only knows ONE answer

Here's a pretend generator. Ask it for a "hero", a "chef", a "scientist" — any role. Watch what it draws. This one was only ever trained on one kind of example, so guess what it hands back every single time…

Tap a role above. Notice the machine gives you the same look every time, no matter what you ask. It's not being mean on purpose — it literally never saw anything else while learning. Whatever's missing from the training data goes missing from the output.

Grown-ups call this dataset bias: bias in → bias out. Fixing it isn't magic — it's giving the model a wider, fairer set of examples to learn from.

🏷️ The pro wordsA telltale flaw in generated media (extra fingers, melted text, wrong counts) is an artifact. The predictable ways a generator breaks are its failure modes. When it confidently invents something that isn't there, that's a visual hallucination. And when the output is skewed because the training data was skewed, that's bias — specifically dataset bias. Real pros inspect for all of these before anything ships.

🎯 Boss · Artifact Detective

Beat the clock

Real content-review jobs are a race: catch every glitch fast, before bad media goes out. Find all 4 glitches in a scene before time runs out. Clear 3 scenes to earn the badge!

Scenes cleared: 0 / 3

0found this scene

–seconds left

🚀 Beyond the Basics

Why a human always has to look

The scary truth: AI can't reliably catch its own mistakes. The same model that drew the 6-fingered hand thinks the hand is perfect. That's exactly why humans stay in the loop:

🧑‍💻

QA & review (a real career)

Studios hire people whose whole job is inspecting AI media for artifacts and bias before it's published. "Prompt QA" and "AI content reviewer" are real, paid roles.

🪞

Self-check is weak (the catch)

You can ask a model "are there mistakes?" and it'll often say "nope!" — while staring right at one. A model can't be the only judge of its own work.

⚖️

Bias is a safety issue (not just quality)

If a generator only ever pictures one kind of person for a job, it teaches the world a wrong, unfair story. Teams audit for this on purpose.

🏷️ Pro names to look upInspecting outputs is called QA (quality assurance) and red-teaming. Catching skew is bias auditing or fairness testing. The golden rule: human-in-the-loop — a person reviews before it ships.

🔬 Real Model Lab

Even the checker gets fooled

This is a real image-captioning model (ViT + GPT-2) running 100% on your device. We'll show it a deliberately broken scene — extra-fingered hand, floating stuff, garbled text — and ask it to describe what it sees. Watch: it answers confidently and never once mentions the glitches.

not loaded

First load downloads the model from Hugging Face, then it's cached for offline use. On a 4GB laptop give it a minute. ☕

🎉 Day 16 Complete!

You learned the golden rule of AI media: never trust the output — inspect it. Generators leave artifacts, invent things that aren't there, and pick up bias from their data. And they can't reliably catch any of it themselves. That's why your eyes matter.

Day 17: The Judge's Table →

Tomorrow: if a generator makes 4 different pictures, which one is best? You'll become the judge — and meet the machine that tries to judge for you. ⚖️