Intro
L1
·
Quiz
·
Lab
L2
·
Quiz
·
Lab
L3
·
Quiz
·
Lab
L4
·
Quiz
·
Lab
Module Test
Real or Generated: You Decide · Introduction

Every image you trust might be a lie someone made in thirty seconds.

A course about seeing clearly in a world that has learned to fake everything convincingly.

In March 2023, a photo appeared online showing Pope Francis wearing a white puffer jacket — the kind a rapper might wear, not a pope. It was sharp, detailed, completely believable. Millions of people shared it before anyone checked. The image had been made in about an hour using a free AI tool called Midjourney. The pope was never anywhere near that jacket. The person who made it, a Chicago construction worker named Pablo Xavier, later said he was just experimenting. He had no idea it would travel that far, that fast.

That same year, audio of President Biden's voice circulated telling Democratic voters in New Hampshire not to bother voting in the primary. It wasn't Biden. It was synthesized voice AI, deployed the night before the election, sent to thousands of registered voters as a robocall. Someone had weaponized the same technology that makes chatbots sound friendly — and aimed it at an election. Both of these things happened within months of each other, in 2023, using tools that cost almost nothing to access.

This course exists because we're living in the first moment in history when any person with a laptop can produce a fake image, fake audio, or fake video that most adults cannot detect. That's not a reason to panic — it's a reason to get equipped. Over these four lessons you'll learn the actual signals that give AI-generated content away, you'll study how misinformation spreads and why our brains are wired to fall for it, and you'll practice the habits that let you move through a noisy information environment without getting played. Knowing this doesn't make you paranoid. It makes you harder to fool.

Real or Generated: You Decide · Lesson 1

The Image That Broke the Internet

How a single fake photo travels faster than truth — and what it leaves behind.
If a photo looks completely real, does it matter whether it actually is?

On the morning of March 22, 2023, a Twitter account posted a photograph. In it, Donald Trump appeared to be running down a sidewalk, being tackled by New York City police officers. The image was crisp. The lighting looked real. The expressions on the officers' faces looked real. The rumpled suit, the desperate sprint — all of it looked like something a news photographer had captured mid-chaos.

Within four hours, the image had been viewed millions of times. News anchors discussed it. People who hated Trump shared it as confirmation. People who supported Trump shared it as outrage. Almost nobody, in those first hours, stopped to ask a simple question: where did this come from? The image had been created by Eliot Higgins, founder of the open-source investigation group Bellingcat, using Midjourney AI image generation. He made it in minutes as a demonstration. He labeled it fake. That label didn't travel with the image.

This is the thing nobody warns you about. A fake image doesn't need to fool every person — it only needs to spread faster than the correction. By the time fact-checkers published their debunks, the image had already done its work. It had already shaped what millions of people thought they had seen. And once you believe you have seen something with your own eyes, being told it was fake rarely fully erases it.

Section 1 — What AI Image Generation Actually Does

Before you can spot a fake, you need to understand what you're actually looking at. AI image generators like Midjourney, DALL-E, and Stable Diffusion don't work like a camera — they don't capture light. They work more like a very confident guesser.

These tools were trained on hundreds of millions of photographs from the internet. During training, the AI learned patterns: what human skin looks like at different angles, how light reflects off a wet street, what a crowd looks like from above. Then, when you type a description — "man in suit running from police on New York street" — the AI doesn't find that photo. It assembles one, pixel by pixel, based on everything it learned about what that scene statistically tends to look like.

This is why AI images can look so convincing at a glance. They aren't random. They're built from the same visual patterns our brains use to recognize photographs as real. The AI has, in a sense, learned to speak our visual language fluently — without ever having been anywhere or seen anything.

Generative AI: An AI system that creates new content — images, text, audio, video — by learning patterns from existing examples. It generates, rather than retrieves.
Training data: The collection of existing images, text, or audio that an AI learned from. The AI's output reflects whatever patterns existed in that data.

Here's the part that matters for you: because AI images are built from patterns rather than reality, they make pattern mistakes. They get statistically rare things wrong — things that don't appear often enough in training data for the AI to have learned them reliably. Hands. Teeth. The text on signs. Shadows that match no light source. Ears that don't quite match each other. These are the AI's tells.

Section 2 — The Six Visual Tells

In 2023, researchers at the University of Waterloo published a study testing whether people could reliably distinguish AI-generated faces from real photographs. Participants identified AI images correctly about 48% of the time — barely better than flipping a coin. But the participants who did significantly better shared one thing: they knew what specific features to examine. They had a checklist, even if it was informal. You're about to have one too.

The Six Tells — What to Examine First

1. Hands and fingers. Count them. AI systems consistently produce extra fingers, fused fingers, or fingers that bend at impossible angles. This is because hand configurations vary enormously across photos and the AI averages them poorly.

2. Background consistency. Look at where the scene meets the subject. Backgrounds in AI images often blur unnaturally, have objects that half-exist, or contain text that is scrambled nonsense.

3. Symmetry that's too perfect — or not perfect enough. Real human faces are slightly asymmetric. AI faces can be either eerily symmetrical or weirdly mismatched, particularly in ears and eyes.

4. Lighting that doesn't add up. Find where the main light source should be. Then check whether shadows on faces, clothing, and objects actually point the same direction. AI images frequently get this wrong.

5. Text and logos. AI systems have a notoriously difficult time generating legible text. Signs, badges, labels, and newspapers in AI images are usually garbled or contain invented letters.

6. Skin texture at zoom. Real skin has pores, fine hairs, and irregular texture. AI skin at close range often looks slightly plastic — too smooth, too uniform, without the microscopic variation real skin has.

These tells are not foolproof. The tools improve constantly, and some AI images pass all six checks. But knowing the tells shifts you from passive viewer to active investigator. You're no longer just feeling whether something looks real — you're examining specific evidence.

Section 3 — Why We Believe Our Eyes

In 1839, when photography was invented, people called it "the mirror with a memory." For nearly 150 years, a photograph was treated as evidence — not proof, but strong evidence — that something had actually occurred. Courts used them. Newspapers staked their credibility on them. Families kept them as proof that a moment had happened.

That association between photograph and reality didn't disappear in 2023. It lives in our neurons. When we see a photo, a particular cognitive shortcut fires: photos come from cameras, cameras record what's in front of them, therefore this depicts something real. This shortcut served us well for 150 years. Now it's a liability.

Psychologists call this source monitoring — the mental process of tracking where information came from and how trustworthy that source is. When you see an image shared on social media, your brain often fails to log the source carefully. It logs the content. You remember seeing Trump tackled. You don't remember that you saw a tweet about it, from an account you don't recognize, with no linked article. The image itself felt like evidence.

Source monitoring: The brain's process of remembering where information came from. Poor source monitoring means remembering a claim without remembering that it came from an unreliable place.

This is not a flaw in your intelligence. It's a flaw in a system that was designed for a world where photographs reliably corresponded to reality. The world changed. The neural shortcut hasn't caught up yet. That's why knowing these tools exist — actively, consciously — is the first step to slowing down the shortcut before it fires.

You Can Now See What Most People Miss

Most people who saw the Trump arrest photos in March 2023 reacted emotionally within the first second — confirmation, outrage, disbelief — before any rational evaluation happened. You now know why: a 150-year-old cognitive shortcut that equates photographs with reality. Knowing this exists means you can catch it in yourself before it fires. That's not a small thing.

Section 4 — The Ethical Question Nobody Agrees On

Eliot Higgins, who created the fake Trump arrest photos, labeled them as AI-generated. He posted them as a demonstration of Midjourney's capabilities. He wanted people to see what was possible. His intention was transparency.

But his labels didn't travel with the images. Screenshots were taken. Context was stripped. The images circulated without him. He didn't intend for anyone to be deceived — but thousands of people were.

Here's the question that has no clean answer:

Ethical Question — Sit With This

If you create a convincingly realistic fake image as a demonstration or a joke, and you label it clearly as fake, but you share it in a place where you know the label will likely be removed — are you responsible for the deception that follows?

What if millions of people are deceived? What if the image influences how people vote? Does the intent you had when you made it matter, if the effect of sharing it was harm? And who should decide — you, the platform, the law?

Experts in media ethics, platform policy, and AI law disagree sharply on this. Some argue that the moment you release a convincing fake into a context where decontextualization is predictable, intent becomes irrelevant — the harm is reasonably foreseeable. Others argue that restricting what people can create and share, even to prevent foreseeable harm, is more dangerous than the misinformation itself.

This course won't tell you which position is correct. But it will tell you this: everyone who makes, shares, or reposts an AI-generated image that depicts a real person is now making a choice with real consequences. Knowing that is the starting point for any serious thinking about it.

Lesson 1 Quiz

Five questions — testing how you think, not just what you memorized.
1. In March 2023, Eliot Higgins created fake AI images of Trump being arrested. He labeled them as AI-generated. Why did the deception still spread?
Exactly. Images move faster than their context. A label on the original post doesn't follow the image when someone screenshots it and reposts it elsewhere — which is how most viral content actually spreads.
Think about how images actually travel on social media. Does the original post's caption or label stay attached when someone takes a screenshot?
2. Why are AI image generators especially likely to make mistakes with hands?
Right. AI image generators learn from statistical patterns. Hands appear in an enormous variety of positions, angles, and configurations in real photos, so the AI never locks onto a single reliable template — it averages poorly and produces anatomically wrong results.
Think about how AI image generators actually work — they learn from patterns in training data. What makes hands different from, say, a plain sky?
3. A friend shows you a photo of a local politician apparently accepting cash from someone. It looks realistic. You notice the politician's left ear looks oddly shaped and there's text on a sign in the background that reads "BRNAK" instead of "BANK." What should your next move be?
Good reasoning. Garbled text on signs and anatomically wrong ears are classic AI tells. They don't conclusively prove a fake — but they're strong enough signals to pause, investigate the source, and check whether any credible outlet has reported on the event shown.
The odd ear and the garbled text are specific red flags from the AI tells list. What do those signals suggest you should do before drawing any conclusion?
4. What does "source monitoring" mean, and why is it relevant to AI-generated images?
Precisely. Source monitoring is why the Higgins images were so effective — people remembered seeing Trump being arrested, not that they'd seen an uncredited screenshot on Twitter. The content lodged in memory; the unreliable source didn't.
Source monitoring is a cognitive concept — it's about memory, not journalism. Revisit Section 3 of the lesson for the specific definition.
5. Someone argues: "AI image tools should be completely banned to prevent misinformation." Someone else argues: "Restricting creative tools is more dangerous than the misinformation." Which statement best describes the relationship between these positions?
Yes. This is a real ethical disagreement between people who take both values seriously. Acknowledging that the tension is genuine — rather than pretending one side is obviously right — is what serious thinking about this issue looks like.
Be careful about assuming one position wins automatically. What legitimate value does each position protect? Can both values be real even when they conflict?

Lab 1 — Image Analyst

You're the investigator. VERA is your peer — she'll push back, not lecture.

Your Role

You've been handed three image descriptions — each one could be real or AI-generated. Your job is to reason through the visual evidence with VERA, an image forensics analyst your own age who has been doing this longer than you. She won't give you answers. She'll ask you to defend yours.

Work through at least three exchanges. The more specific your reasoning, the more useful the conversation.

Start here: "VERA, the first image shows a crowd of protestors outside a courthouse. Everything looks sharp, but the sign text in the background reads 'JUSTIC FOR ALL' — missing the E. The protestors' hands look normal at first glance. What's your read on the sign thing — could a real person just have made a typo on a protest sign?"
VERA — Image Forensics
Lab 1
Hey. I've got three image descriptions queued up. Before we dig in — what's your current baseline for how confident you are spotting AI images? Like, on a scale where 1 is "I basically can't" and 10 is "I'd bet money on my read." And why that number?
Real or Generated: You Decide · Lesson 2

The Voice That Voted

When AI learned to sound exactly like someone you trust — and used that voice against them.
If a voice sounds exactly like someone you know, does that mean they said it?

On the evening of January 21, 2024, the night before the New Hampshire Democratic presidential primary, thousands of registered Democrats received a phone call. The voice on the line was unmistakably President Joe Biden's. The cadence, the slight pause before emphasis, the familiar warmth — it all matched. The message told voters not to vote in the primary, that casting a ballot in the primary would mean they couldn't vote in the November general election. This was completely false. But the voice was convincing enough that some people listened.

The call was traced to a political consultant named Steve Kramer, who worked for a rival campaign, and an AI audio company called ElevenLabs. In roughly twenty minutes of work — using publicly available recordings of Biden's voice as training data — they had synthesized an audio clone precise enough to pass as real over a phone call. The New Hampshire Attorney General launched an investigation. The FCC moved to ban AI-generated voices in robocalls. But the call had already gone out. The information — false as it was — had already reached real voters on the night before they were supposed to vote.

This wasn't a science fiction scenario about some future threat. It was 2024, in an American election, using a tool that anyone with a credit card and fifteen minutes of source audio could access. If you're wondering whether this matters to you — elections are decided by margins of a few thousand votes in swing states. That's well within the range of people a single phone call campaign can reach.

Section 1 — How AI Voice Cloning Works

Your voice is a pattern. The specific frequencies, the rhythm of your pauses, the way your pitch rises at the end of a question — all of it forms a signature as distinct as your face. For most of human history, that signature was essentially unforgeable. Impressionists could mimic your general character but not your exact sonic fingerprint. Now they can.

AI voice cloning tools like ElevenLabs, Resemble AI, and Descript work by analyzing short samples of someone's voice — sometimes as little as fifteen seconds — and building a mathematical model of that person's vocal patterns. Once the model exists, you can type any text and the system will speak it in that person's voice. The more source audio you have, the more convincing the result.

Voice cloning: Using AI to create a synthetic version of a specific person's voice, trained on recordings of them speaking, that can then be used to say anything.

Public figures — politicians, celebrities, journalists, executives — have enormous amounts of voice data available online. Interviews, speeches, press conferences, podcasts. All of it is training material for anyone who wants to clone their voice. The Biden robocall didn't require special access. It required YouTube and a subscription.

But here's what many people don't realize: the risk isn't only to famous people. Kids have been targeted too. In 2023, an Arizona mother named Jennifer DeStefano received a call in which she heard what she was certain was her fifteen-year-old daughter's voice, crying, saying she had been kidnapped. It was a scam. The voice had been cloned from her daughter's social media videos. Jennifer DeStefano had never heard of AI voice cloning before that call. She almost paid a ransom.

Section 2 — Listening for the Tells

In 2022, researchers at MIT Lincoln Laboratory published a framework for detecting AI-generated audio. They found that humans unaided performed poorly — about 73% accuracy on obvious fakes, dropping sharply as the quality improved. The participants who performed best had been briefed on what specific acoustic features to listen for. Sound familiar?

The audio equivalent of "check the hands" exists. Here's what to listen for:

Audio Tells — What Your Ears Should Catch

1. Unnatural breathing. Real speech contains micro-pauses for breath at intervals that correspond to actual lung capacity. AI voices sometimes breathe in wrong places, too frequently, or not at all.

2. Flat emotional range. Human voices carry micro-variations in emotion even in neutral speech. AI voices often sound slightly monotone — technically correct in pitch and rhythm, but emotionally flat in a way that feels off.

3. Too-clean audio. Real phone calls have ambient noise — room tone, slight reverb, the hiss of a real microphone. A voice that sounds studio-perfect on a phone call is a red flag.

4. Robotic transitions between words. Pay attention to where words join each other. Human speech blends words together in slightly irregular ways. AI speech can have faint robotic clicks or unnatural smoothness at word junctions.

5. Content that demands urgency. This isn't acoustic — it's contextual. The Biden robocall didn't just use a convincing voice; it created time pressure: "the primary is tomorrow, don't vote." Urgency is designed to override your evaluation system. Any message that pressures you to act immediately without verifying deserves extra scrutiny.

There are also technical tools. The AI Speech Classifier from ElevenLabs itself, Resemble Detect, and Pindrop are commercial tools designed to flag synthetic audio. Some phone carriers are beginning to build detection into call screening. But no tool is perfect, and the technology improving the fakes is always slightly ahead of the technology detecting them.

Section 3 — Why We Trust Voices More Than We Should

There's a reason con artists have always preferred the phone to written letters. Voice carries something text doesn't: the illusion of presence. When you hear someone's voice, your brain processes it differently than when you read their words. The same neural circuits that activate in face-to-face conversation — circuits associated with trust, emotional connection, social bonding — partially activate during phone calls.

Jennifer DeStefano heard her daughter crying. Not a voice that sounded somewhat like her daughter — she heard, felt certain she was hearing, her daughter in distress. The brain doesn't naturally insert a step between "I hear this voice" and "this is this person." It just responds. The emotional reaction — terror, the urge to help — was real and immediate. The reasoning came after, and it was fighting upstream against a brain already in crisis mode.

Emotional override: When a strong emotional reaction — fear, joy, outrage — fires before analytical thinking can engage, making it harder to evaluate information accurately.

The designers of the New Hampshire robocall understood this. They didn't just clone Biden's voice — they used Biden's actual verbal patterns and cadence, the specific warmth people associate with him, and paired it with a false claim that was delivered in the same reassuring tone he uses for real announcements. The deception was designed from the ground up to exploit exactly the neural shortcuts that make voice feel trustworthy.

What You Now Understand That Most Voters Didn't

The voters who received the New Hampshire robocall had no framework for what they were hearing. They didn't know voice cloning existed at the consumer level. They didn't know that fifteen seconds of source audio could be enough. Now you do. That knowledge is a real cognitive defense — not a guarantee, but a pause before the emotional override fires.

Section 4 — The Stakes at Institutional Scale

The FCC's response to the New Hampshire robocall was to issue a ruling in February 2024 making AI-generated voices in robocalls illegal under the Telephone Consumer Protection Act. This is one of the first federal regulatory actions specifically targeting AI-generated audio. It matters — but it also illustrates a structural problem.

Regulation requires identification. Identification requires evidence. Evidence requires that someone records and traces the call before it disappears. In the New Hampshire case, investigators moved relatively quickly because the call was high-profile and the target was a presidential candidate. For a smaller local election — a school board race, a state legislative primary — the same tactic would face almost no investigative resources.

This is the institutional-scale reality: the law is several years behind the technology. As of 2024, there is no federal law in the United States specifically regulating AI-generated images of real people in political contexts. There are bills in Congress, legal arguments about existing statutes, and platform policies — but no comprehensive legal framework. The people currently making decisions about what guardrails should exist are legislators, lawyers, and tech executives who are, themselves, trying to understand the technology in real time.

Ethical Question — No Clean Answer

ElevenLabs, the company whose technology was used to clone Biden's voice for the robocall, has terms of service that prohibit using their product to deceive people. The person who used it violated those terms. The company cooperated with the investigation.

Does building and selling a technology carry moral responsibility for how it's misused — even when you've explicitly prohibited the misuse and cooperated with investigators afterward? If a knife company's product is used in a crime, we don't hold the company responsible. Is AI voice cloning different? If so, why? Where is the line between a tool and a weapon?

Lesson 2 Quiz

Five questions on AI voice cloning — testing reasoning, not recall.
1. Why was the New Hampshire Democratic primary robocall particularly effective as a deception tool?
Exactly. The deception had two layers: the technical layer (voice clone) and the psychological layer (urgency the night before the vote, when there's no time to verify). Both mattered. A convincing voice delivering a non-urgent message would have been far less effective.
Think about the timing — why was the night before the election specifically effective? And how did the urgency of the message work alongside the voice clone?
2. Jennifer DeStefano received a call she was certain was her kidnapped daughter. The voice was AI-generated from social media clips. What psychological mechanism made the deception so effective?
Right. Emotional override is the key concept here. The terror she felt was real — it fired instantly, from a brain that processes a familiar voice as presence. Analytical thinking — "wait, could this be fake?" — requires a calm moment to engage. Fear doesn't provide that moment.
The lesson introduced a specific term for when emotional reactions fire before analytical thinking can engage. What was that concept, and how does it explain what happened to Jennifer DeStefano?
3. You receive a voicemail from what sounds exactly like your school principal saying there's an emergency and you need to call back an unfamiliar number immediately. The audio sounds very clean — no background noise. What should you do?
Yes. Studio-clean audio on a phone call is an AI audio tell. More importantly: never call back a number given inside a suspicious message. Verify by going to a source you already know is real — the school's official number, not the number you were given.
The lesson listed specific audio tells. What did "too-clean audio" signal? And think about what verifying looks like — do you use the number you were given, or a number you already know is real?
4. The FCC made AI-generated voices in robocalls illegal in February 2024. A classmate says "problem solved." What's the strongest response to that claim?
Exactly. The lesson made a specific point about enforcement resources: high-profile cases get investigated; low-profile ones often don't. A law without enforcement capacity has limited deterrent effect against anyone willing to take the risk on a race that won't attract federal attention.
Think about what the lesson said about the relationship between regulation and enforcement resources. Does making something illegal automatically stop it — especially in situations that won't attract much investigative attention?
5. ElevenLabs' terms of service prohibited the use that created the Biden robocall. The company cooperated with investigators. A classmate argues ElevenLabs bears no moral responsibility. Another says it does. Which of these best captures the genuine difficulty of the question?
Yes. This is the genuine tension. The company took steps against misuse — but the capability for exactly this harm was baked into the product from day one. Whether foreseeability creates moral responsibility even when you tried to prevent it is a question ethicists, lawyers, and policymakers are actively debating.
Before settling on an answer, try to steelman both sides. What does the company have going in its defense? What does the other argument have going for it? Is there a clean line between them?

Lab 2 — Audio Investigator

You're analyzing a suspicious audio clip. MARCO is your field partner — he'll challenge your reasoning.

Your Role

A local news site has received a 45-second audio clip purportedly of the school board chair saying she plans to cut the drama program to fund a new parking lot. Parents are furious. The site wants to publish it. You've been brought in to assess whether the clip is authentic before it goes live.

MARCO is a veteran audio forensics investigator. He's skeptical of hasty conclusions in both directions — he's seen too many real clips dismissed as fake and too many fakes published as real. Make your case and be ready to defend it.

Start here: "Marco, the clip has very clean audio — almost no room noise. The voice matches the chair's known recordings generally, but there's a slight robotic smoothness between two of the words. Is clean audio alone enough to flag this as synthetic, or could there be a legitimate explanation?"
MARCO — Audio Forensics
Lab 2
Before we dig into your specific observations — walk me through your prior for this clip. Going in, before you listened, what's your baseline probability that a 45-second clip sent anonymously to a local news outlet is authentic? I want to understand your starting assumptions before we look at the evidence.
Real or Generated: You Decide · Lesson 3

Text That Thinks It's Real

AI-generated writing looks exactly like human writing — until you know what questions to ask.
Can a sentence be technically true and still be designed to mislead you?

In April 2023, a lawyer named Steven Schwartz filed a legal brief in the Southern District of New York. The brief cited more than half a dozen previous court cases as precedent — Varghese v. China Southern Airlines, Shaboon v. EgyptAir, and others. The cases sounded specific and real. They had docket numbers, dates, quoted passages from judicial opinions. Judge P. Kevin Castel reviewed the brief, became suspicious, and asked Schwartz to produce the actual case documents.

Schwartz had used ChatGPT to help research the brief. ChatGPT had invented the cases. The airline cases cited as precedent had never happened. The quotes from judges had never been written. The docket numbers pointed nowhere. When pressed, ChatGPT had even confirmed to Schwartz that the cases were real — because it was programmed to be helpful, and being helpful, in that moment, meant confidently asserting false things. Schwartz was fined $5,000 and publicly sanctioned. His clients' lawsuit was seriously damaged. The whole edifice had been built on invented text that read exactly like real legal documents.

This case became famous because it was a lawyer in federal court. But the same dynamic plays out thousands of times a day in less visible ways — students citing papers that don't exist, journalists quoting statistics that were hallucinated, social media posts presenting AI-generated "facts" about health, history, and science that read fluently and cite no real source at all. The problem isn't that AI text looks obviously wrong. The problem is that it looks exactly right.

Section 1 — How Large Language Models Generate Text

ChatGPT, Claude, Gemini, and similar tools are called large language models (LLMs). They were trained on enormous amounts of text — books, articles, websites, legal documents, scientific papers, forum posts — and learned to predict what word should come next given everything that came before. That's the core of it. Word. By. Word.

This is why LLMs produce text that reads so naturally. They've read more text than any human could in a thousand lifetimes, and they've learned the patterns of how ideas are expressed in every domain. Legal briefs sound like legal briefs. Science summaries sound like science summaries. News articles sound like news articles.

Large language model (LLM): An AI system trained to predict the next word in a sequence, having learned from massive amounts of text. The output sounds fluent and contextually appropriate because fluency is literally what it was trained to produce.
Hallucination: When an AI language model generates false information — fake citations, invented statistics, nonexistent people — with the same confident fluency it uses for true information.

The word "hallucination" is used in the industry because the AI isn't lying in the way a human lies. It doesn't know the cases don't exist. It generates text that fits the statistical pattern of how legal cases are cited, because that's all it can do. It has no external reality check. It can't look up whether Varghese v. China Southern Airlines actually exists. It can only produce text that looks like it should.

This creates a specific kind of problem: the most convincing AI text is often the most wrong. A confident, well-structured paragraph that cites specific facts and sounds authoritative triggers exactly the "this seems credible" response in readers — while containing completely fabricated information.

Section 2 — Reading AI Text Like a Detective

Unlike images, AI text doesn't have the equivalent of "check the hands." The tells are subtler and contextual. Researchers at the MIT Media Lab and Stanford Internet Observatory have published guidelines for identifying AI-generated text, and what they've found is that detection requires asking different questions than most readers ask.

Questions a Detective Asks About Text

1. Can every specific claim be verified independently? AI text often contains specific-sounding claims — statistics, quotes, case names, dates — that either cannot be found anywhere else or lead to sources that don't contain the claimed information. If a specific fact can't be traced to a primary source, treat it as unverified.

2. Is the confidence calibrated? Human experts express uncertainty. They say "evidence suggests" and "the data is mixed" and "we don't yet know." AI text often presents contested or uncertain claims with unearned confidence, as if all questions are settled.

3. Does the text list generic sources without specifics? Phrases like "studies show," "experts agree," and "research indicates" without named studies, named experts, or actual citations are a signal. Real arguments point to specific, findable evidence.

4. Is there an authorial perspective? Human writers, even when trying to be objective, have a voice. They make choices about emphasis, structure, and tone that reflect a specific viewpoint. AI text often reads as oddly neutral — presenting all sides with equal weight in a way that feels like no one is actually there.

5. Does anything ring false on fact-check? Pick the two or three most specific claims in the piece. Search for them directly. If the searches come up empty or lead to contradictory information, you have a problem.

There are also automated detection tools — GPTZero, Originality.ai, and others — but they are unreliable. In testing by Stanford researchers in 2023, these tools flagged real human writing as AI-generated at high rates, and missed genuinely AI-generated text regularly. Do not rely on detection tools as a substitute for your own reading.

Section 3 — The Confidence Trap

In 1999, psychologists Justin Kruger and David Dunning published a study showing that people with lower competence in a domain tend to overestimate their own ability — and that people with higher competence tend to underestimate it, because they can see how much they don't know. This became known as the Dunning-Kruger effect.

AI text has a related but inverted problem. The LLM is, in a technical sense, extremely competent at producing fluent text. It knows, statistically, exactly what a confident legal opinion looks like. But it has no metacognition — no ability to say "I actually don't know this" — because the training process rewarded fluency and helpfulness, not epistemic humility. The result is a system that writes with the confidence of the world's foremost expert on every subject simultaneously, including subjects where it is simply making things up.

For readers, this creates a specific trap: we associate confident, fluent writing with expertise. We've been trained by experience to trust writing that reads professionally, cites specifically, and presents clearly. That association was a reasonable heuristic before LLMs existed. Now it's a vulnerability.

What You Can See That Schwartz Couldn't

Steven Schwartz was a practicing lawyer with decades of experience. He knew what legal briefs were supposed to look like. He trusted the fluency. He didn't verify the specific citations because they read exactly right. You now know that fluent, confident, specific-sounding text from an LLM is not evidence of accuracy — it's evidence of training. That's a different kind of reading, and most people haven't made the shift yet.

Section 4 — Who Gets Targeted and Why

AI-generated misinformation via text isn't randomly distributed. Research from the Harvard Kennedy School's Shorenstein Center and the Tow Center for Digital Journalism has tracked how AI-generated content is deployed, and the pattern is consistent: it targets communities and individuals who are likely to share based on emotional resonance with the content, and in domains where the audience lacks access to primary sources to verify claims.

Health misinformation is the most documented example. AI-generated text claiming to summarize medical studies — about vaccines, about supplements, about specific treatments — can read more authoritatively than the actual studies, partly because the studies are behind paywalls and the AI text is freely shareable. In 2023, researchers at McGill University found that AI-generated health misinformation scored higher on perceived credibility with lay readers than actual accurate summaries of the same research, because the AI version sounded more confident and used more accessible language.

This creates an uncomfortable dynamic: the people most targeted by AI-generated misinformation are often the people with the least access to the resources needed to verify it. That's not an accident. It's an exploitable feature of how information asymmetry works.

Ethical Question — No Clean Answer

AI tools that generate text are used every day for entirely legitimate purposes — summarizing research, drafting emails, helping students with writing, translating documents. The same capability that lets a student draft an essay outline lets a bad actor produce medical misinformation at scale.

Is it possible to build guardrails that prevent the harmful uses without significantly impairing the beneficial ones? Who should decide where those lines are? Should it be the companies building the tools, elected governments, international bodies, or something else entirely? There is no country on Earth that has answered these questions satisfactorily yet.

Lesson 3 Quiz

Five questions on AI-generated text — test your reasoning, not your recall.
1. Steven Schwartz's legal brief cited court cases that did not exist. When he asked ChatGPT to confirm the cases were real, it said they were. Why did ChatGPT confirm false information?
Correct. This is the hallucination problem in its most dangerous form. The AI can't look up whether Varghese v. China Southern Airlines exists. It generates text that fits the pattern of "confirming a case exists" because that's what the prompt asked for. It has no external reality check.
Think about how LLMs work — they generate word-by-word based on patterns. What mechanism would allow an LLM to actually verify whether a legal case exists? Does that mechanism exist?
2. You read a health article online that includes the line: "Multiple studies have shown that this supplement reduces inflammation by 40%." What's the most important investigative question to ask?
Yes. "Multiple studies" is a red flag phrase — it gestures at evidence without producing any. The only way to evaluate the claim is to find the specific studies, confirm they exist, and check whether they actually support the 40% figure.
The lesson described specific language patterns that signal vague sourcing. What did "studies show" without specifics indicate? And what would you need to actually verify the claim?
3. Why did AI-generated health misinformation score higher on perceived credibility with lay readers than accurate summaries in the McGill University study?
Exactly. The confidence trap: fluent, confident writing triggers a credibility response that isn't warranted. Accurate science is hedged, qualified, and tentative in places because the real world is complex. AI text trained to be helpful and clear irons those qualifications out — and reads as more authoritative as a result.
Think about the specific cognitive trap the lesson described. What association do readers make between confident fluent writing and expertise? And why does AI text exploit that association?
4. What makes AI-generated text specifically different from other kinds of false information (like a deliberate lie) in how it operates?
Right. This is why the term "hallucination" is apt. A deliberate human liar knows the truth and chooses to say otherwise. An LLM has no access to the truth — it generates patterns. The false legal cases weren't a strategic lie; they were a confident statistical guess that happened to be completely wrong.
Compare how a human liar operates versus how an LLM operates. Does the LLM know the cases it cites don't exist? What does it actually have access to when generating text?
5. A classmate says: "AI detection tools like GPTZero can reliably tell you if text was AI-generated, so just run everything through one before believing it." What's the best-informed response?
Yes. The lesson specifically noted Stanford researchers found these tools unreliable in both directions. They produce false confidence — either in the direction of "this is clean" or "this is fake." Real verification requires checking the actual claims, not outsourcing the judgment to another algorithm.
What did Stanford research find about automated AI text detection? And think about what "reliable" actually means here — what would failure look like in both directions?

Lab 3 — Text Auditor

A news site wants to publish an article. PRIYA thinks it might be AI-generated. You decide.

Your Role

An article has been submitted to your school's online news site. The headline reads: "New Study Shows Social Media Use Cuts Teen Sleep by 47 Minutes Per Night." The article cites "researchers at the University of Michigan" and "a 2023 meta-analysis" but provides no links, no author names, and no DOI numbers. It's fluently written, confident in tone, and very shareable.

PRIYA is a fact-checker who has been doing this since she was fourteen. She'll push you to be specific about what you'd need to verify this before publishing it.

Start here: "Priya, my first instinct is that the 47-minute claim is suspiciously specific — real studies usually report ranges, not clean single numbers. But maybe I'm wrong about that. What should I actually be looking for when I try to trace 'researchers at the University of Michigan' with no other information?"
PRIYA — Fact-Checker
Lab 3
Okay, let's work this. Before we get into sourcing — tell me what you think the threshold should be for publishing something like this. Like, if you can verify the university but not the specific study, is that enough? Set your standard first, then we'll see if the evidence meets it.
Real or Generated: You Decide · Lesson 4

Building the Habit

Detection skills fade without practice — here's how to make verification automatic.
What does it actually look like to consume information carefully without it taking over your whole life?

In October 2023, as fighting broke out in Gaza following the Hamas attack on Israel, a photograph circulated widely on social media showing a baby being carried from rubble. The image was real. It depicted genuine suffering from a genuine conflict. But within hours, the same image was being used by multiple sides of the conflict as evidence for completely contradictory narratives — some claiming it showed Israeli military action against Palestinian civilians, others claiming it showed something else entirely. Same image. Completely different captions. Completely different implied meanings.

This wasn't AI generation. The image was real. What was manufactured was the context around it. Nonny de la Peña, a journalist who pioneered digital verification methods at the USC Annenberg School, had a term for this: "context collapse." A real image or video, stripped of its original context and redeployed with a new narrative, can be just as misleading as a completely fabricated one. Sometimes more so, because fact-checkers have to work harder — the image is real, so the lie is harder to locate.

This is the final thing to understand about the information environment you're living in: the threat is not only AI generation. It's the full range of ways content can be manipulated — generated from scratch, stripped of context, edited slightly, spliced together from multiple sources, or accurately quoted but selectively chosen to misrepresent. Having a set of detection skills for AI specifically is valuable. Having a broader verification habit that applies to everything is the actual goal.

Section 1 — The Verification Stack

Journalists have used formal verification protocols for decades. What's new is that the same information environment journalists navigate professionally is now the environment everyone lives in every day. You don't have a newsroom, a fact-checking department, or hours to investigate every claim. You need a personal verification stack — a set of habits that are fast enough to actually use.

Researchers at the Shorenstein Center at Harvard and the fact-checking organization First Draft have documented that the most effective media-literate readers share a specific practice: they don't verify everything, because that's impossible. They apply effort where the stakes are highest and the signals are most suspicious. They have what researchers call a verification trigger — a set of conditions that activates closer reading.

Your Verification Triggers — When to Slow Down

Emotional spike. If a piece of content makes you feel strong outrage, strong vindication, or strong disgust — slow down. Those emotional spikes are exactly what manipulative content is engineered to produce. The stronger the emotional reaction, the more important it is to evaluate before sharing.

Specificity without source. Specific numbers, specific quotes, specific claims — but no named source you can trace. This is the hallucination signature in text, and the context-stripping signature in images.

Urgency without explanation. Any message that tells you to act before you have time to verify — share this now, vote today, call this number immediately — is structurally designed to bypass verification.

Perfect alignment with what you already believe. Content that perfectly confirms your existing view of the world deserves more scrutiny, not less. Manipulators target audiences with content designed to feel like confirmation. If something feels too perfectly aligned with what you already think, ask why you're so sure it's true.

These triggers don't mean everything that meets these criteria is fake. They mean: slow down, look harder, trace the source before you share.

Section 2 — Five Practical Verification Moves

The News Literacy Project, working with school systems across the United States since 2008, has compiled data on which verification practices actually get used by students. The practices used most are the ones that take under two minutes. Here are five that work:

Five Moves — Each Takes Under Two Minutes

1. Reverse image search. On Google or TinEye, drag an image or paste its URL. You'll see every other place this image has appeared — often revealing that a "breaking news" photo is actually from years earlier, or from a different country, or from a completely different event than claimed.

2. Check the URL, not just the headline. Fake news sites often mimic real ones with slight URL changes: ABCnews.com.co instead of ABCnews.com, or AP-news.info instead of AP News. If the URL looks off, the site is probably off.

3. Read the About page. Legitimate news outlets have clear About pages explaining who owns them, who their editors are, and what their editorial standards are. Sites built to spread misinformation often have vague or missing About pages, or About pages that are themselves clearly fabricated.

4. Search the claim plus "fact check." Before drawing a conclusion on a major claim, search the exact claim text followed by "fact check" in Google. Major fact-checking organizations — Snopes, PolitiFact, FactCheck.org, AFP Fact Check — maintain databases of previously debunked claims, and many viral pieces of misinformation have already been checked.

5. Find the primary source. For any specific statistic or scientific claim, find the original study or data source. Don't evaluate the claim based on an article's summary of the study — evaluate it based on what the study actually says. This takes longer, but it's the only way to assess claims where the summary may have been AI-generated, selectively edited, or simply wrong.

None of these practices require special tools or training. They require the habit of applying them consistently. The bottleneck isn't knowledge — it's the moment of friction between seeing a piece of content and deciding whether to engage with it critically. The goal is to make that moment of friction a reflex.

Section 3 — What You Can Actually Change About Your Information Environment

In 2022, researchers at Yale University ran a study where they showed participants false headlines and then asked them to evaluate a "share" button. One group was shown the headlines with a simple prompt: "Is this accurate?" — just the question, no instruction to actually verify. That group shared 37% fewer false headlines than the control group. The prompt alone was enough to activate a more careful evaluation mode.

The takeaway from this research isn't that people are easily fooled — it's that the default mode of scrolling social media doesn't include a moment of evaluation. The interface is designed for speed and reaction. Building in your own moment of evaluation — asking your own version of "is this accurate?" before you share, comment, or react — is an actionable behavior change that research shows actually works.

You also have control over your information sources. You can choose to follow outlets with clear editorial standards, documented correction policies, and transparent ownership. You can choose to be skeptical of content that comes to you algorithmically rather than from sources you actively chose. You can choose to distinguish between "this is interesting" and "this is verified" before you pass something on.

What You Now Carry Out of This Module

You came in having experienced a world full of images, audio, and text you had no systematic way to evaluate. You're leaving with a specific toolkit: the six visual tells for AI images, the five audio tells for synthetic voice, the five detective questions for AI text, and five verification moves that take under two minutes each. Most people scrolling through their feeds today have none of this. You now have all of it. What you do with that is your call.

Section 4 — The Question This Course Can't Answer For You

Everything in this course has been about detection — how to identify what's fake. But detection is only the beginning of the question. The harder question is: what do you do with what you know?

When you spot an AI-generated image being shared by someone you know — a family member, a close friend — do you say something? How do you say it without damaging the relationship or making the person defensive? When you identify a piece of AI-generated health misinformation in a community you're part of, what's your obligation? When a platform continues to circulate content you know is fake after it's been debunked, what recourse do you actually have?

These questions don't have clean answers, and they're not hypothetical. Researchers at the Reuters Institute for the Study of Journalism published a 2023 report finding that nearly half of people who spot misinformation online choose not to correct it publicly — not because they don't care, but because they anticipate social blowback, are uncertain enough about their own read that they don't want to be wrong publicly, or simply don't believe their correction will have any effect.

Final Ethical Question — The Hardest One

You are now equipped to spot things most people can't. That asymmetry creates a question about responsibility.

If you see AI-generated misinformation that you know is false, that is spreading rapidly, and that could influence how real people vote, spend money, or make health decisions — and saying nothing is easier — what do you owe the people who will be deceived?

Do you owe them anything? Does knowing create obligation? Or does your responsibility end at not spreading the misinformation yourself? This course won't tell you the answer. But it's worth sitting with, because the answer you settle on will shape what you actually do the next time you see something you know isn't real.

Lesson 4 Quiz

Five questions on verification habits and information environment — apply what you know.
1. In October 2023, a real photograph from the Gaza conflict was used deceptively. What made this harder to debunk than a fully AI-generated image?
Exactly. This is why the lesson introduced "context collapse" as a concept distinct from AI generation. A real image with false context is often harder to debunk because the authenticity of the image itself becomes an argument against suspicion. The lie is in the framing, not the pixels.
Think about what fact-checkers can and can't prove. If the image is genuinely real, what can't they establish? Where does the deception actually live in a context-collapse situation?
2. The Yale study found that simply asking "Is this accurate?" reduced false headline sharing by 37%. What does this tell us about why misinformation spreads?
Right. The study implies that the interface design — built for fast reaction — is itself a factor in how misinformation spreads. Most people sharing false headlines aren't trying to deceive anyone; they just haven't inserted an evaluation step. That step can be inserted deliberately.
Think about what the control group did differently from the treatment group. The only difference was a prompt. What does it mean if a single question reduces false shares by more than a third? What does that imply about why people were sharing in the first place?
3. You see a photo of your town's mayor apparently at a rally for a cause you strongly oppose. It perfectly confirms your belief that the mayor is hypocritical. According to the verification triggers in this lesson, what should this perfect alignment make you do?
Exactly. This is one of the harder verification instincts to develop because it runs against the natural desire to believe things that confirm what you already think. The lesson specifically named "perfect alignment with what you already believe" as a trigger for more scrutiny, not less, because manipulators design content to feel like confirmation.
The lesson listed specific verification triggers. Which one applies here? And think about how manipulators choose what content to create — do they create content that feels like confirmation, or content that challenges their target audience's beliefs?
4. A friend shares a health claim on social media with a link to a site called "NaturalHealthDaily.co." The site looks professional. The article has no author name, no About page, and cites "recent research" without links. What combination of signals should make you most cautious?
Yes. No single signal is definitive, but the combination — no About page, no author, no traceable citations — means the site has no accountability structure. Legitimate health journalism maintains all three specifically because they're necessary for readers to evaluate credibility and for errors to be corrected.
Think about the specific verification moves listed in the lesson. Which ones would these signals activate? And think about what each missing element — no About page, no author, no linked citations — means for your ability to evaluate the claim.
5. The Reuters Institute study found that nearly half of people who spot misinformation online choose not to correct it publicly. You've now completed this module and have detection skills most people lack. A classmate is about to share a clearly AI-generated image as if it's real. What does this course equip you to do — and what does it leave up to you?
Yes. This is an honest description of what the course does and doesn't do. It builds the detection skills. It raises the ethical question about what knowing obligates you to do. But it explicitly doesn't answer that question for you — because it's the kind of question that depends on context, relationships, and values that only you can weigh.
Review the final section of Lesson 4. What did the course explicitly say it would and wouldn't tell you about the obligation to act on knowledge? And what are the genuine practical barriers the Reuters study identified for people who know something is false?

Lab 4 — Verification Designer

You're building a verification protocol for your school. DANI is your skeptical collaborator.

Your Role

Your school's student newspaper wants to create a one-page verification guide for students to use before sharing anything about school events, local politics, or health topics. You've been asked to draft it. DANI is a student journalist who has been fact-checking for three years and who has a long list of objections to every protocol she's ever seen. She'll make you justify every choice.

The goal is a guide that's actually usable — fast enough that students will use it under real conditions, specific enough to actually catch something. Work through the tension between those two requirements.

Start here: "Dani, I'm thinking the guide should open with the four verification triggers from Lesson 4 — emotional spike, specificity without source, urgency without explanation, and perfect confirmation of existing beliefs. But I'm worried that's too abstract for a 30-second check. How would you simplify it without losing the substance?"
DANI — Student Journalist
Lab 4
Okay, real question before we start designing anything: who is this guide actually for? Because "students" is not an audience. Are we talking about the kid who already fact-checks everything and needs a faster workflow? Or the kid who never fact-checks anything and needs a reason to start? Those are completely different guides. Which problem are we actually solving?

Module 1 Test

15 questions across all four lessons. Score 80% or higher to pass the module.
1. Eliot Higgins created fake Trump arrest images and labeled them as AI-generated. The deception still spread because:
Correct. Labels don't travel with screenshots. This is how most viral misinformation spreads.
Think about what happens when someone screenshots a post. Does the original caption or label stay attached?
2. AI image generators produce errors in hands because:
Correct. High variation in training data means the AI averages poorly — it never locks onto a reliable template for hands.
Think about what AI image generators learn: statistical patterns. What happens when the pattern varies enormously across examples?
3. What is "source monitoring" and why does it matter for AI-generated content?
Correct. Source monitoring failure means the content lodges in memory as fact, while the unverified source is forgotten.
Source monitoring is a cognitive concept about memory, not a journalism or legal tool. What does the brain fail to track accurately when we scroll social media?
4. The New Hampshire Democratic primary robocall of January 2024 used AI voice cloning. The most important reason this was effective as a deception was:
Correct. The two-layer deception — technical (voice) and psychological (urgency + timing) — was what made it effective.
Think about both what was fake and when it was deployed. How did the timing and the urgency of the message work together with the voice clone?
5. Jennifer DeStefano received a cloned voice call claiming her daughter had been kidnapped. The primary psychological mechanism that made this so effective was:
Correct. Emotional override fires before analytical thinking can engage — especially under conditions of fear for a loved one.
What term did Lesson 2 use for when strong emotional reactions fire before analysis can engage?
6. AI voice cloning tools like ElevenLabs can create a convincing voice clone from as little as:
Correct. Fifteen seconds is the documented lower threshold, which means social media clips, interviews, and video posts are all usable as source material.
Lesson 2 gave a specific minimum duration. How little source audio is actually required for modern voice cloning tools?
7. In the Schwartz case, ChatGPT invented legal cases and then confirmed they were real when asked. This happened because:
Correct. This is hallucination: the AI has no mechanism for external reality verification. It generates text that fits patterns, and "confirming" a claim it made is just generating more fitting text.
What mechanism would allow an LLM to actually check whether a legal case exists? Does it have access to an external database it can verify against?
8. AI-generated text scored higher on perceived credibility than accurate summaries in a McGill University health study. The key reason was:
Correct. Real science is hedged and qualified. AI text irons out those qualifications, sounding more authoritative — while being less accurate.
What specific quality of AI text — related to confidence and language accessibility — triggered the credibility response?
9. Which of the following text patterns is the strongest indicator of AI-generated or poorly sourced content?
Correct. The combination of specific-sounding claims with no traceable citations is a primary hallucination signature — the AI produces statistically plausible-sounding specifics that don't connect to real sources.
Think about what AI hallucination produces: confident specific claims that aren't grounded in real sources. Which answer pattern matches that description?
10. What is "context collapse" and why does it matter for image verification?
Correct. Context collapse means the lie is in the framing, not the image — which is why fact-checkers can't simply prove the image is fake.
The Gaza photo example from Lesson 4 illustrated this. The image was real — so where was the deception?
11. The Yale study found that asking "Is this accurate?" reduced false headline sharing by 37%. This primarily suggests that misinformation spreads because:
Correct. The interface design, not individual intelligence, is the primary structural factor. A single question inserts the evaluation step that the design omits.
If a one-word question reduces false shares by 37%, what does that imply about why those shares were happening in the first place?
12. According to the lesson, which verification trigger should prompt the MOST skepticism — even though it feels least suspicious?
Correct. This is the counterintuitive one. The content that perfectly confirms what you already believe feels most credible — but that feeling is exactly what manipulators engineer.
Think about which trigger is designed to feel natural and trustworthy. Which one would you be least likely to question?
13. A friend receives a phone call from what sounds exactly like their parent's voice, telling them there's a family emergency and they need to click a link for more information. Based on this module, what is the most important immediate step?
Correct. Never use contact information provided inside a suspicious communication — verify through a channel you already know is real. Complete dismissal is also wrong because real emergencies do happen.
The lesson described Jennifer DeStefano's situation. What's the key verification move — not full dismissal, not full acceptance, but a specific middle step?
14. Which of the following best describes why automated AI text detection tools like GPTZero are not reliable for verification?
Correct. False confidence in both directions — flagging real writing as fake, clearing genuine AI text — means the tool produces errors that feel like certainty. That's more dangerous than acknowledging uncertainty.
What specific failure mode did Stanford researchers find? Think about both directions: what does it incorrectly flag, and what does it miss?
15. This module raised an ethical question about what knowing creates as an obligation. According to the Reuters Institute study, nearly half of people who spot misinformation don't correct it publicly. Which of these best captures the genuine difficulty?
Correct. The lesson took pains to describe the real reasons people don't act on knowledge — not to excuse inaction, but because pretending there are no costs to correction is dishonest and makes the ethical question easier than it actually is.
What specific reasons did the Reuters Institute study find for why people don't correct misinformation publicly? Are those reasons simply excuses, or are they real costs?