AI Skill Course
Module 06 of 08
Lesson 06 · 36 minutes · Beginner

AI is amazing.
It's also wrong
a lot.

After five lessons celebrating what AI can do, here's the lesson that protects you. By the end, you'll spot AI mistakes that fool most people — and you'll know exactly when not to trust the answer.

You'll play
A Real-or-AI game
You'll see
4 failure modes
You'll learn
5 rules to live by
See the cracks
Hallucinated
"The internet is the great equalizer of our modern age — it gives every person an equal voice."
— Mark Twain (allegedly)
AI confidently attributed this quote to Mark Twain. Twain died in 1910. The internet wasn't invented until the 1980s.
This is what an AI hallucination looks like.
Part 01 · The cracks

There are four ways
modern AI goes wrong.

Each one is a different failure of a different system. Each one is a real risk you've probably already encountered without noticing. Each one is fixable — if you know what to look for.

01

Hallucinations

The AI confidently states things that are completely false. Made-up quotes. Fabricated citations. Invented studies. Often delivered with the same tone as a verified fact.

Real case · 2023 A US lawyer submitted a court brief citing 6 legal cases that didn't exist. ChatGPT had invented every one — complete with realistic case numbers, judges, and quoted opinions. He was sanctioned by the court.
02

Bias

AI inherits the prejudices baked into its training data — sometimes amplifying them. Decisions that look "objective" hide patterns of historical discrimination.

Real case · 2018 Amazon built an AI to screen resumes. It learned from 10 years of past hiring (mostly male) and started penalizing resumes that contained "women's" anywhere. Scrapped before launch — but only because they caught it.
03

Deepfakes

AI-generated images, voices, and videos so convincing that you can't reliably tell them from real ones. Used in fraud, political manipulation, and harassment.

Real case · 2024 A finance worker in Hong Kong transferred $25 million after a video call with the company "CFO" and several "colleagues." Every person on the call was an AI-generated deepfake.
04

Adversarial inputs

Tiny changes that fool AI completely. A few pixels added to an image. A subtle audio overlay. The AI confidently misreads it; humans don't even notice the change.

Real case · 2017 MIT researchers added imperceptible noise to a turtle photo. Google's vision AI classified it as a rifle. They later printed a real 3D turtle that AI consistently misidentified — from any angle.
Part 02 · Hands on

8 mystery items.
Real or made up by AI?

Each one sounds plausible. Each one could go either way. Trust your gut — but check it at the end. Most people score 4–5 out of 8. The world's worst score: most students who started this module.

Round 1 of 8
Score 0
Quote
Loading…
Answer
0/8

You 0/8
Average learner 4.5/8
Random guessing 4.0/8
Part 03 · The unfair mirror

Bias in AI is not a software bug.
It's a history bug.

AI doesn't invent its bias. It inherits the patterns we already had — and then it scales them up to millions of decisions. Watch what happens.

A simulated hiring AI

100 fictional applicants — 50 women (pink), 50 men (blue). Equal qualifications. Watch how a model trained on 10 years of past hiring decides.

100 applicants
50 / 50 split
Trained AI ranks them
"Top 20" picked
Top 20 finalists
Why this happened The AI learned from 10 years of past hires — when only 18% of successful candidates were women. It picked up patterns like "graduated from school X" or "used the word executed in resume" — patterns that were proxies for male candidates. The bias was never in the code. It was in the training data, and the AI didn't know it was being unfair. It just maximized "looks like a successful past hire."

Famous real-world cases

2016 · USA
COMPAS recidivism algorithm

Used in US courts to predict re-offending risk. ProPublica showed it falsely flagged Black defendants as "high risk" at nearly 2x the rate of white defendants.

2018 · MIT
Facial recognition gender gap

Joy Buolamwini found leading face-recognition systems were 99% accurate on white men — and as low as 65% on Black women. Same model, different reliability.

2019 · Healthcare
The hidden race proxy

A widely-used US healthcare algorithm directed less care to Black patients. It used past healthcare spending as a proxy for "sickness" — but Black patients historically receive less care, so the proxy hid the bias.

2024 · Image gen
Generative image stereotyping

Studies showed asking image AIs for "a CEO" almost always returned a white man; "a janitor" returned a person of color. The stereotypes weren't programmed — they were absorbed from billions of internet images.

Part 04 · Train your eye

5 ways to spot an
AI-generated image.

Generative AI is getting better fast. These tells used to be obvious in 2022. In 2026, they're more subtle — but most still hold. The faster you can spot them, the safer your judgment.

01
6th finger
Look at the hands

Six fingers. Three fingers. Fingers melting together. Hands are still the #1 tell — AIs were trained on photos where hands are usually relaxed and partially hidden.

02
CAFFE esprresoo misspelled · gibberish
Read the text in the image

Words on signs, t-shirts, books — AI almost always botches them. Letters that aren't quite letters. Words that look right at a glance and gibberish on close inspection.

03
asymmetric pair
Check pairs of things

Earrings that don't match. Eyes slightly different sizes. Eyebrows shaped differently. Two glasses lenses that warp differently. The AI generates each side independently — and they rarely agree.

04
reflections don't match
Reflections & shadows

Mirrors that don't reflect what's in front of them. Shadows falling the wrong direction. Light sources that physics wouldn't allow. AI doesn't simulate physics — it pattern-matches what photos look like.

05
background melts
Scan the background

Foreground subjects look sharp. But edges in the background often warp, melt, or smear. Lines that should be straight curve subtly. Patterns that should repeat shift mid-image.

The honest truth in 2026:

Top-tier AI images can already pass casual inspection. The tells are getting subtler. Verifying source matters more than spotting AI. If a photo claims something important (a politician said X, a celebrity did Y) — find the original source, the news organization, the verified account. If you can't, treat it as fiction.

Part 05 · The wisdom

5 rules for working with AI
that will save you.

Print these. Tape them to your monitor. The people who get the most out of AI follow them obsessively.

01

Treat AI output as a draft, never a finished product.

Every AI response is a starting point. Read it critically. Fact-check claims that matter. The mistake isn't using AI — it's accepting its output without review.

02

Verify any specific claim before using it.

Names, dates, statistics, citations, legal facts, medical facts — anything specific must be confirmed against a real source. The AI is wrong with confidence; you need to be right with verification.

03

Don't outsource decisions only humans should make.

Medical diagnoses, legal advice, financial moves, hiring decisions — the AI may help you think, but humans (and licensed experts) must decide. Liability and judgment don't transfer to a machine.

04

If it sounds too plausible, get suspicious.

Hallucinations sound great. They mimic the texture of truth — confident, specific, well-structured. That smoothness is a warning sign, not a confirmation. Real research has rough edges.

05

Source verification beats AI detection.

Don't ask "is this AI?" Ask "where did this come from?" A verified source from a known publication beats any AI-detection tool. Trace the origin — most fakes don't have one.

Part 06 · Knowledge check

Five questions.
One step from module 7.

Aim for 4/5. Wrong answers explain themselves.

Question 01 of 05

0/5

Continue
Module 06 complete

You're now harder
to fool than most people.

You know the four failure modes. You played the Real-or-AI game. You can spot deepfake tells. You have 5 rules taped to your forehead. That's a real, practical, professional-grade skill — and you'll use it tomorrow.

Up next · Module 07

The ethics episode

Jobs, privacy, copyright, alignment — the questions that will define the next decade of AI. With a branching dilemma simulator that puts you in the chair where these decisions actually get made.

Continue to Module 07