synthetic media = images, audio or video made or changed by AI
AI can now make pictures, voices and video clips that look and sound like real recordings. Media made or heavily edited this way is called synthetic media. A deepfake is synthetic media that shows a real person doing or saying something they never did. A voice clone is an AI-made copy of how a real person sounds. Plenty of synthetic media is harmless: art, jokes, game assets, movie effects. The problem starts when it is passed off as a real record of something that happened.
Why "just look closely" stops working
In 1.2.2 you saw that a generator learns patterns from lots of examples and then makes new things that fit those patterns. Every flaw people learn to spot, like odd hands or garbled text, is a pattern the next version of the tool can learn to fix. Careful looking still helps, but a missing "tell" is not proof that something is real.
What about AI-image detectors?
A detector is a classifier, like the ones in 1.1.2. It says "AI-made" or "real," and it can be wrong in two ways: it can miss a fake, or it can flag a real image. Below is an imaginary detector. Nothing here describes a real product. Set it up the way the sliders start, then try making fakes common.
Imaginary AI-image detector · 1,000 images
AI-made images50
Real images950
AI-made and flagged45
AI-made but missed5
Real but wrongly flagged95
Total flagged140
Of the images the detector flags, only 32% are really AI-made (45 of 140).
This is the precision idea from the confusion matrix in 1.1.2: when it says "AI-made," how often is it right? Try making AI images common and rare. The same detector gives very different results.
With the starting settings, 50 of the 1,000 images are AI-made. The detector flags 45 of them, but it also wrongly flags 95 real ones, so of 140 flagged images only 32% are really AI-made. If fakes are common (40%), the same detector is right about 86% of the time it says "AI-made." Same tool, different usefulness. That is why a detector's verdict is a clue, never a verdict.
🔑Real or fake isn't something you can reliably decide by staring at the pixels. Ask where the media came from and who else confirms it.
Principles of AI · 1.2.3SkillConcept
Scams: The Pressure Pattern
A scam is built to get you to act before you think: send money, read out a code, or hand over a login. Most scams, with or without AI, are built from the same few pressure tools. Learning the pattern works better than memorizing individual scams.
Seven pressure signals
Urgency.Act now, within the hour, or something bad happens
Secrecy.Don't tell anyone, keep this between us
Unusual payment.Gift cards, crypto, wire transfers, or sending money back
Claims to be someone.A relative, a bank, a boss, or an official you can't confirm
Asks for a login, code or personal details.Passwords, one-time codes, account or ID numbers
Too good to be true.A prize, a giveaway, or an easy high-paying offer
Pushes a link or attachment.Click here or open this file to fix, claim or confirm
Practice on invented messages
Signal spotter · message 1 of 8
Hi, it's Dad. I lost my phone, so this is a new number. I'm in a jam and need $400 in gift cards right now. Please don't tell Mom, I'll explain later.
Tick every pressure signal you see. (These messages are invented for practice.)
Notice that real messages can contain a signal too. A school closing early is urgent. One signal is a reason to slow down and check, not proof of a scam. The scams stack several signals and push you to skip the check.
Where AI changes things
Older scam messages were often easy to spot: generic, full of mistakes, aimed at everybody. As you saw in Unit 1, a language model writes fluent text, and when it is given context (a name, a school, a coach, a recent event) it can use that context. Scammers can use that to write messages that sound specific to you, in good grammar, in large numbers, and in a friendly tone. Cloned voices can add a phone call that sounds like someone you know.
What gets harder: spotting a scam from bad spelling, a generic greeting, or a voice that sounds off.
What doesn't change: the pressure pattern. Urgency, secrecy, unusual payment and requests for codes are still the giveaways.
Where the personal details come from: mostly what is posted publicly. Less you share openly means less to build a believable message from.
⚠️A real company or school will not ask you to read a one-time code back to them, pay in gift cards, or keep a request secret from your family.
Principles of AI · 1.2.3Skill⚠ Watch Out
What You Can Check (and Can't)
A five-step check that works on any claim
Check ladder
Stop.Notice how it makes you feel. Fear, excitement and outrage are what fakes and scams aim for.
Who is behind it?.Is it an account or site you know? Does it have a history, or did it just appear?
What do others say?.Search for the same story from other reliable outlets or the official source.
Where did it start?.Try to trace the image, clip or quote back to the earliest version you can find.
Use a second channel.For a message about money or a crisis, contact the person or company using a number or site you already have.
ExampleA viral image
An invented example: a dramatic photo of a flooded street is spreading online. The caption says it was taken in your town this morning.
Step 1 — Stop
The photo makes you feel alarmed, and you feel an urge to share it. That is the moment to pause, not to repost.
What you can't check by yourself
Things that feel like checks but aren't
Looking for "tells".Extra fingers, odd text or strange shadows are less common as tools improve, and real photos can have odd flaws too.
A detector's verdict.No detector is known to be fully reliable, and the detector lab shows why even a decent one misleads when fakes are rare.
"It looks too real to be fake".Realistic is exactly what these tools are built to produce.
Missing provenance info.Some media carries origin information (such as content credentials), but support is uneven and the info can be removed, so its absence proves nothing.
Sometimes you'll end up not knowing, and that is an honest answer. "I can't tell, so I won't share it or act on it" is a good result of a check.
Habits that protect you
Pause before money, codes or logins. Real emergencies survive a two-minute check.
Use a second channel. If a call or message asks for money, hang up and contact the person or company on a number you already have.
Agree on a family code word that a caller must know. Don't post it anywhere.
Never share one-time codes. They exist to prove it's you.
Report it to the platform, a trusted adult, or your school, and don't forward it as a joke.
💡Your side of it: don't make fake images, voices or videos of real people, even as a joke. Even if nobody believes it, it can hurt the person, and it is against the rules of many platforms and schools. Making something clearly fictional and labeled is different. Don't upload other people's photos or voices to AI tools without their permission.